infoproc (Steve Hsu)

Downloaded via on 2019-4-25

1963 results

2019-04-19 on infoproc

I have a feeling AI will be much more exciting over the next 30 years, but of course I could be wrong. There are still deep mysteries in physics, like the foundations of QM, quantum gravity, and quantum information/computation. Only in the 3rd case are experiments likely to push things forward materially in the near term… It is only in golden eras that deep and beautiful theoretical ideas get to make contact with experiment. The 1970s to early 1980s were such an era with quantum field theory and the standard model. I don’t know when the next one will come…

2019-04-19 on infoproc

LHC could still discover new physics, but it might be much less exciting than what had been predicted 30 years ago: SUSY at the TeV scale, mini-black holes from TeV quantum gravity, etc. It’s kind of discouraging that for 30 years all experimental anomalies indicating new physics beyond the standard model (with the exception of neutrino oscillations) have gone away after more data / deeper analysis.

2019-04-18 on infoproc

Yes, absolutely. If there is someone you know that would be a good guest, perhaps send them and me an email introduction? Then I’ll try to schedule.

2019-04-09 on infoproc

Need to understand how much additional predictive power PGS yields when combined with test scores and HS GPA. Unless the edge is large it may not be worth the additional cost and controversy of genotyping ISA candidates.

2019-04-05 on infoproc

Ha Ha, funny that you caught that. I noticed it too when I listened to the podcast.

I have no idea why. Perhaps I was just trying to make our guest feel comfortable?

2019-04-02 on infoproc

It’s just a rough way to gauge how unusual the *typical* selected individual would be in the original population. For SD = 10 or 40, the answer is: impossibly rare!

2019-04-02 on infoproc

Thanks! Surprisingly large SD for AMC1957 — 1/6 of bodyweight? I was guessing more like 10%.

2019-03-31 on infoproc

Beautiful stuff — thanks!

2019-03-30 on infoproc

Believe it or not, research grants in the US are still mostly awarded through meritocratic review by other scientists, and not especially influenced by politics (with some big projects proving the exception to the rule). I am of the view that medical research will be transformed significantly in the next decade by new technologies such as genomics, AI/ML, inexpensive sensors (of many types), etc. So there is an opportunity for new entrants to advance against the old guard. Only time will tell…

2019-03-24 on infoproc

The Mueller investigation was cited many times (e.g., by DOJ) as a reason not to declassify specific documents requested by congressional committees. This pretext no longer applies.

2019-03-18 on infoproc

For almost every startup the challenges that need to be overcome are as much on the business side (customer acquisition, execution) as on the technical side. This is even harder when the customer is government!

2019-03-15 on infoproc

It’s not very good yet, but only a matter of time: need photos paired with large numbers of genotypes to analyze.

2019-03-15 on infoproc

As much as I like the science stuff, I will get even more satisfaction from Othram solving cold cases and giving families of victims closure and a sense of justice done.

Othram: “For the main wall of the City was of great height and marvellous thickness, built ere the power and craft of Númenor waned in exile; and its outward face was like to the Tower of Orthanc, hard and dark and smooth, unconquerable by steel or fire, unbreakable except by some convulsion that would rend the very earth on which it stood.” — The Return of the King (The Lord of the Rings Volume 3)

2019-03-08 on infoproc

ex-Left we have 1. GOP establishment, 2. Trump supporters whose main issue is immigration, and much less so competition with PRC, 3. Trump supporters who are closely tracking the China trade war and potential cold war.

I think group 2 is bigger than group 3. There is also a hawkish vs PRC subset of group 1, but counterbalancing it are lots and lots of free trade big business types. Taking all this into account I think there is a lot of pressure on Trump to dial down the trade war — get some meaningful concessions from China but then get back to business as usual.

2019-03-08 on infoproc

Sorry if that was unclear. The guys I spoke with were not claiming that PRC is at complete parity in AI. I believe (as I think you may) that 1. big gains are available via implementation of ideas we already have at hand (deep learning + …), and PRC will be at least as good as US in realizing those, but 2. new ideas (probably originating in academia) are also necessary to go the next step beyond, and it’s not clear that PRC will be as good as the West in making these breakthroughs.

However, what is being discussed is whether PRC is caught up enough across a broad range of tech (jet engines, materials, semiconductors, general hardware / software, AI, biotech, power trains, solar panels, etc.) that a cold war with the US and its allies would not cripple further economic development and tech advancement. Not that a decoupling wouldn’t hurt, but that it’s too late for it to be decisive.

2019-02-16 on infoproc

Roughly speaking, humans cluster into continental populations by ancestry. Within the Euro population a predictor trained in one subpopulation (e.g., UK) seems to work about as well in even fairly distant, but still Euro, populations (e.g., Finns). We know how the prediction accuracy falls off for specific traits like height as we move to more distant validation sets (e.g., train on Euros, test on S. Asians, E. Asians, Africans). The fall off is correlated to a genetic distance one would get by, e.g., just counting average number of base pair differences between groups. We are doing more sophisticated analysis so stay tuned.

To answer your last question: Yes, as a first approximation that would actually work ok, although Middle Easterners and S. Asians wouldn’t fit so well into your categories, and there is a lot of genetic diversity in Africa.

2019-02-15 on infoproc

Out of sample. There are now a variety of successful validations of genomic predictors using populations on different continents, born in different decades, etc. The results are very robust.

However, the predictor quality degrades significantly when the test population is very genetically distant from the training population. This is believed to be due to different SNP correlation patterns: a tag SNP which is highly correlated to a causal genetic variant in the training population may not be a good tag in a distant population.

2019-02-03 on infoproc

Thanks — this is something I’ve worried about and I’m probably due for a bigger car now that the kids are older.

What is the best source for car safety ratings?

2019-01-27 on infoproc

I was trying to avoid doing that… not sure whether it’s good or bad!

2019-01-27 on infoproc

Initial plan is the two of us together for each interview, with one of us (typically the one who knows the guest or their area the best) as lead interviewer. Later I can imagine doing some shows with only Corey or myself present. We also plan to do some shows with just the two of us in conversation, like episodes 0 and 1.

The first guest is a neuroscientist who does brain mapping. We have that video done and will post this coming Thursday 🙂

2019-01-26 on infoproc

Yes, mom is the extrovert 🙂

I can think of some good topics for scientific documentaries. Ask her to email me if she is interested.

2019-01-15 on infoproc

This guy doesn’t make actual trade recommendations (which is in a sense much harder), but he does have a keen eye for macro trends in the digital space.

2019-01-06 on infoproc

Close but not quite — Shell Beach (just north of Pismo) and Morro Bay 🙂

2019-01-05 on infoproc

Please have this discussion somewhere else. It has nothing to do with the post above on disease risk prediction.

2018-12-28 on infoproc

I don’t think we did that specifically but we found that the sibling rank order results are consistent with nearly all of the predictive power coming from direct genetic causality as opposed to genetic nurture (the latter being attenuated for sibling comparisons). I suppose we should publish that result all by itself… but we have too much other stuff to do!

2018-12-28 on infoproc

In our within-family tests of the height predictor we get same-sex sibling height order correct about 90% of the time. This suggests that “genetic nurture” effects are relatively small.

Case-Control data for disease conditions is harder to come by, though.

2018-12-06 on infoproc

Just Google SU(2) SO(3) double cover. If you use the complex 2-spinor and real 3-vector representations of the two groups you can write out the correspondence between elements of the group explicitly.

See also…

This demonstration has been attributed to both Feynman and Dirac, but who knows who came up with it first.

2018-12-06 on infoproc

Oops, thx for catching that! I was thinking of qubits when I wrote that rather than electrons… Fixed now.

2018-12-05 on infoproc

His analysis is, as you say, within the narrow lens of orthodox economics. But most of his facts are correct, AFAICT.

Re: broader considerations, see Munger’s comments: https://infoproc.blogspot.c…

2018-11-29 on infoproc

OMG IQ enhancement via gene editing!!!

2018-11-28 on infoproc

As a first demonstration something which directly eliminates disease would be more defensible. (Even then there is the question of why genetic screening would not have worked in place of editing.)

There is a case to be made (see my comments on the father and HIV) for CCR5 but it is less palatable to most ethicists.

Keep in mind that even people with PhDs or MDs or who write for leading journals or who hold fancy titles typically react emotionally, not rationally, to new developments. It’s sad, but not surprising.

Similar reactions to the first IVF baby can be found in the record. Societal views on ethics change over time, and I expect the same in this case.

2018-11-28 on infoproc

Thx, I created a new post. The talk is interesting and reveals some important details.

2018-11-27 on infoproc

My office built a gene-editing (CRISPR) core facility for MSU, which does animal models.

He Jiankui did his postdoc at Stanford under molecular bio wizard Steve Quake. Quake was trained 100% in physics: BS, PhD (in theory, IIRC). Feng Zhang also has an undergrad degree in physics.

Yes, Physicists can do stuff! 😉

NYTimes: “Dr. He got his Ph.D., from Rice University, in physics and his postdoctoral training, at Stanford, was with Stephen Quake, a professor of bioengineering and applied physics who works on sequencing DNA, not editing it.”…

2018-11-27 on infoproc

I don’t think it was a good first demonstration project, and it has been widely criticized. See, e.g., the comments on the YouTube video above.

2018-11-26 on infoproc

Didn’t FISC report abuses happen well before Halper/Downer HUMINT?

We don’t know specifically whether the people who were surveilled (by “contractors” — including Fusion GPS oppo types?) included Trump orbiters (AFAIK). But shouldn’t we?

Of course you can justify it all by saying that Obama CIA/FBI/DOJ were doing their duty in spying on the opposition candidate. (Because, RUSSIA!!!) But massive surveillance powers were surely in play, and this was denied for a long time.

Rogers at the NSA did not feel comfortable with it. My understanding is that right now UK and Australian (5 eyes) partners are desperately trying to keep Trump from revealing their participation in this. GCHQ and Hannigan? Just a coincidence?…

PS I believe Downer has recanted the whole story now and does not claim that Papa-D mentioned Hillary’s emails in their conversation. It’s only people invested in the Narrative that continue to use him as a key element in all of this. He denies it now himself.…

“In his Australian interview, Mr. Downer said Mr. Papadopolous didn’t give specifics. “He didn’t say dirt, he said material that could be damaging to her,” said Mr. Downer. “He didn’t say what it was.” Also: “Nothing he said in that conversation indicated Trump himself had been conspiring with the Russians to collect information on Hillary Clinton.”

For months we’ve been told the FBI acted because it was alarmed that Mr. Papadopoulos knew about those hacked Democratic emails in May, before they became public in June. But according to the tipster himself, Mr. Papadopoulos said nothing about emails. The FBI instead received a report that a far-removed campaign adviser, over drinks, said the Russians had something that might be “damaging” to Hillary. Did this vague statement justify a counterintelligence probe into a presidential campaign, featuring a spy and secret surveillance warrants?”

2018-11-04 on infoproc

The 5% hooked applicants account for 1/3 of all admits: if whites and asians are treated equally in this group that would dilute the effect of the asian-ethnicity variable in, e.g., a regression run on all the data. If there were a slight favoring of asians in this group (not sure why, but I think that was what was found; might not be statistically significant), that would not only dilute the contribution coming from the unhooked part of the data, but even cancel it out. See quotes from the judge (previous post), who seems to appreciate this.

2018-10-30 on infoproc

Thx for pointing that out — very strange. I think I reopened the comments now.

Not sure how I could have done that by accident, and I hope Disqus didn’t do it all by itself.

2018-10-30 on infoproc

? Comments are not closed AFAIK

2018-10-26 on infoproc

I’m aware of Ron’s claims regarding Jewish overrepresentation at H. But the situation is quite different in terms of the quality of data. Thanks to the lawsuit and ongoing trial, we have good data on whites (as a whole), Asian Americans, African Americans, etc. Arcidiacono can make a detailed analysis (see SFFA brief) and Harvard / Card can dispute his conclusions (which, in the case of unhooked applicants, they have NOT).

But Jews are not broken out from the larger white group. So the trial doesn’t shed much additional light on the question you are interested in.

One thing to keep in mind here is that Asian Americans are very interested in attending H, as are Jews. Plenty of talented white gentiles do not apply to H. So national comparisons can be quite misleading.

I don’t think it’s H’s fault that some groups are not as likely to apply for admission. After all, H has among the best financial aid packages, and sends recruitment letters out to folks in “sparse country” (their terminology) with fairly low cutoffs. This came out at trial but was not widely reported:

PSAT score to receive Harvard recruitment letter in sparse country

black or hispanic 1100
white 1310
Asian male 1380

Does this look like racial discrimination? It sure does to me……

2018-10-26 on infoproc

It just looks better to my aging eyes 😎

2018-10-23 on infoproc

Please see the binomial theorem or central limit theorem. If you flip 10k coins the standard deviation in number of heads is only 50 — i.e., you would expect to get 5k plus or minus ~50 heads most of the time.

So to raise IQ by, e.g., 6 SD, you might only need to flip 300 alleles, even if intelligence is controlled by ~10k variants.

The standard deviation scales like sqrt(N) not N.

Even editing 300 variants is not easy, but we don’t need to edit thousands.

Please alert Matt to this correction. I am an admirer of his writing in most instances.

2018-10-16 on infoproc

If you took the non-physicists in Roe’s group of 64 top scientists, and then removed those that did not obtain high math scores, I wonder how many at age 40-50 would still retain some grasp of special relativity. I suspect the fraction would be fairly high.…

2018-10-16 on infoproc

Thanks for these insightful remarks. I have a friend who is a theoretical chemist, and his path seems to have paralleled yours: he is very interested in QM, relativity, and field theory and studies them in his “spare time”!

2018-10-13 on infoproc

I can’t dispute that many students graduate from HS without any exposure to physics. I suppose that when I wrote the first sentence I was thinking decent HS = good students are exposed to physics. Since I’m talking about fairly high thresholds in ability and curiosity, the reference to HS is unnecessary as you say: people of this type would at some point get curious and look it up on the internet…

Educated adult = someone who, either in HS or college, once took a physics class?

2018-10-11 on infoproc

Yes, there are phenotypes like blood lipids and brain images(!) that are still to be delivered to researchers by UKBB 🙂

You have to do some (not very hard) calculations get odds ratios from the predictor quality, and it depends also on disease prevalence.

2018-10-10 on infoproc

I think I’m quoting Putin in the title.

Khabib and I are not comrades, other than that we both know our way around a single leg (my favorite finish is single to a trip) and hadaka jime.

2018-10-09 on infoproc

The technology for testing embryos and applying genomic predictors is already here. One does not have to know the actual causal variants in order to make statistical predictions which, on average, produce much better outcomes.

This has been going on for a long time in, e.g., dairy cattle / chicken / corn / wheat breeding. They don’t know the causal variants either but they have succeeded in increasing productivity by large factors. Dairy cattle breeding is done using SNP predictors which are very analogous to those recently developed for humans.

Note, however, that for gene *editing* one really does need to know which variants are causal…

2018-10-09 on infoproc

Those numbers don’t seem implausible to me.

You could argue this in various ways. Most of these foreign-born STEM PhDs stay in the US, which from the human capital perspective looks like a huge loss for the sending country and a huge gain for the US. However you might worry about dual loyalties, loss of sci-tech leadership due to leakage of expertise back to foreign countries, etc. I don’t think there’s a simple answer.

Almost everyone would agree that it would be nice if the US could produce more domestic students capable of and interested in pursuing STEM PhDs. But are incentives in the US system encouraging talented kids to do so?

2018-09-26 on infoproc

The scholarship has gone to students from eastern Europe as well as Asia. The goal was to bring talented students to Caltech who might otherwise not have access to world class sci/tech education. Globalization has advanced so quickly that perhaps this isn’t all that necessary now.

I’m not as pessimistic as GM about the US. My father was a very patriotic Republican — he loved Ronald Reagan and knew a lot more about communism, the cultural revolution, etc. than US “intellectuals” (academic fools). His family back in China had to live through it.

My brother and I both seriously considered attending the military academies when we were in high school (probably not a good fit for me, in retrospect). His son is a Marine ROTC officer candidate, soon to be commissioned.

The dream of citizenism is not a dead end. Society can evolve. Also, keep in mind the genomic technologies that are right around the corner.

2018-09-08 on infoproc

Are the washing machines sentient with complex memories, like the virtual AI beings postulated above?

2018-09-08 on infoproc

The assumption is that eventually there will exist huge numbers of virtual AI beings who are self-aware and experience a rich virtual world. They will not know that they are not in base reality.

Among the set of all sentient beings to exist in the universe, at any time, perhaps almost all will be virtual. What are the odds, then, that you are not?

2018-08-26 on infoproc

Here’s something about jamming technologies. There’s definitely military grade stuff out there:…

I think placing a bomb beforehand is much harder than flying a drone onto someone’s stopped car or plane.

BTW, the terrorism use case is a bit different from the assassination one… you’d have to jam the airwaves around every taxiway, not just for Air Force One.

2018-08-24 on infoproc

Not sure this is deployed yet…

Do you think that jamming equipment covers the entire route driven by the motorcade? Or the tarmac on which Air Force One sits before takeoff?

I don’t even think it’s easy to do this at a big Trump rally without messing up everyone’s electronics — cellphones, media equipment, etc.

The Mavic drone above uses (IIUC) wifi frequency range (2.4/5.8 GHz w/frequency hopping) for control. Do you think they wipe out the building wifi everywhere Trump speaks?

PS The drone is better than small arms because the killer can be far away, undetected, and just drive away after the job is done.

2018-08-23 on infoproc

I can’t really comment on your comment in public. But let me just say you are missing a couple of important factors.

2018-08-09 on infoproc

Yes, mean 142, SD 12, so area under normal dist above +1.5SD. Hope I didn’t make a mistake 😉

2018-08-09 on infoproc

I think I fixed the link in the original post. If not see…

The notes don’t address the specific thing you are asking about — they just analyze regression relative to parental midpoint in presence of assortative mating, etc.

The specific result you asked about results from a calculation like the following:

Suppose the regression factor is ~0.7, so parents with midpoint IQ of +4SD = 160 would produce kids whose average is +4 x (0.7) = +2.8SD, or IQ =100 + 0.7 x 60 = 142

The SD among the kids would be about 12-13 points. For a kid to reach 160 they would have to be +18 points above average among all the kids, which is about +1.5SD in that group. The odds of this are ~6 or 7 percent. (Higher heritability would give a less pessimistic result; lower h2 would give an even more pessimistic one!)

Note this assumes that E and G are additive and independent, etc. These are only approximately true at best. It would be interesting to see how well these approximations hold up in a big data set with families.

This comment thread does not belong below a post about cinema and Richard Linklater! Do me a favor and paste your Q and my A into the comment thread under the original regression post.

2018-07-27 on infoproc

Oops, we might have been talking about different things. The *total* inter-sib SD is indeed ~13 points but the SD of (additive) *genetic component* difference between sibs (this is the part the predictor is sensitive to) is about 7.5 points.

2018-07-26 on infoproc

What is written above is incorrect, I believe.

It should be var is cut *to* 0.75 (not *by* 0.75), and the IQ SD = sqrt(0.75)x15 = 13 (i.e., not much smaller than in general population)

var = 1 – 0.5 h2

When comparing sibs half of the variants are the same, so only the remaining half contribute to between-sib var. Total variance is reduced by roughly half of (narrow) h2.

As I mentioned we’ve checked empirically that sib SD in height is almost as large as in general population. I’m pretty sure this is also the case for IQ (according to various papers).

2018-07-26 on infoproc

Degree of assortative mating is an input into the calculation, but of course this can be measured by looking at genomes of married couples.

Leaving aside assortative mating, sibs will be concordant at ~50% of the sites influencing the trait of interest (height, IQ, whatever).

Then the variance is reduced to

var = 1 – 0.5 (h2) [if they were identical twins we’d get residual var = 1 – h2, etc.]

where h2 is the heritability (to be precise something like the narrow sense h2 for sibs but full h2 for twins; in former case nonlinearities are not faithfully reproduced), which might be 0.5 for example.

Then the sib variance is 0.75 compared to 1 for the general population. Assortative mating reduces this a bit more.

We’ve checked this for height and it’s not far off. I’ve seen similar results for IQ.

2018-07-26 on infoproc

It’s a result of women focusing more on careers and getting married later. Fertility decline happens in the 30s — late 30s if you are lucky, earlier if you are unlucky. So it’s not surprising that among the elites IVF is very common. The IVF industry is booming, with very high growth rate worldwide.

2018-07-25 on infoproc

Yes, 3/4 of the variance found in the general population for sibs is about right. This can be deduced from pop gen theory or just by looking at data (i.e., directly calculate SD among sibs for some population sample like UKBB).

2018-07-25 on infoproc

SD / variance among siblings is only slightly lower than in the general population. So there is still (easily) few percent chance of having an outlier among your embryos, and many would think it worthwhile to know.

2018-07-25 on infoproc

It’s very analogous to testing for Downs or rare Mendelian disease. ~1 or few percent chance of something really important and actionable. But still worth it to do the test.

In this case, what if you find that embryo #4 is bottom few percent on EA PGS score? You’d probably just use one of the others. In analogy with Downs/Mendelian risk, it’s worth the expense to do the test.

2018-07-25 on infoproc

The issue you raise is relevant to IVF as a whole, not specifically to genomic prediction, because it is typical for IVF to produce more embryos than are implanted. The “embryo selection problem” is generic.

It is now routine for embryos to be frozen in liquid nitrogen without, as far as we know, ill effects, and thawed later for use, potentially many years later. The unused embryo in the hypothetical might not be discarded — it might remain frozen, ultimate fate unknown.

Worldwide IVF cycles are ~1M per annum, and in some developed countries several percent of *all* births result from IVF.

2018-07-23 on infoproc

We’ve been developing our own version of L1 for summary statistics and have tried it on EA3 as well as other data. It won’t be as good as having all the genotypes in one place, but stay tuned 😉

2018-07-13 on infoproc

“GxE interactions” usually refers to situations in which the genetic effect depends on the environment background. Separate G and E contributions to the phenotype (in this case, life success or SES) are typically not described as gene-environment interactions. The paper does clearly suggest separate G and E contributions to life outcomes (see top figure), but is not primarily about GxE *interactions* as usually defined. I removed the reference to GxE interactions in the earlier comment because I thought it would be confusing (apparently it was).

The underlying predictor used by these authors does indeed predict with r~0.35 when used on an out-of-sample population (i.e., not used in training). This isn’t discussed in the PNAS paper because the original large GWAS paper from which the predictor is derived has not yet appeared (but will soon; this is a slightly odd situation where the follow up work got published before the more fundamental work).

If one tries to correct for parental education one is potentially mixing up both G and E effects (since both are passed from parents to kids) and it’s hard to interpret the meaning of that (reduced) correlation.

The graphs I displayed in Figure 1 clearly show that family SES affects life outcomes, and that genetic influences do as well. I don’t think I said anything about whether intervention is pointless. There is almost nothing in behavior genetics that is entirely G or entirely E.

My guess is that common SNP-based predictors for cognitive ability might reach r ~ 0.7 (i.e., account for about half the variance), given enough training data. This is not a genetic determinist position, in the sense that there is still plenty of room for environmental effects. However, many people are (wrongly, in my opinion) skeptical that we can reach this level of prediction.

I’m sitting in an airport in Amsterdam in the middle of the night (US time) — I hope what I wrote above is comprehensible 🙂

2018-07-12 on infoproc

Two people are looking at the same chess board. One can’t tell which side is ahead. The other says “Checkmate in 8” 😉…

2018-07-11 on infoproc

The paper is an out-of-sample validation of a genomic predictor for educational success that is (in the sense I explain, 2nd figure) about as powerful as existing standardized tests. Individuals who are well-below or well-above average in polygenic score have unusually high probability of having trouble at school or excelling at school. The PNAS paper shows that prediction works even between siblings (within family) and also that it has socioeconomic consequences.

It’s “game-over” because someone who claims that (even stronger) genomic prediction of cognitive ability will not be possible in the relatively near future just does not understand what is already in published (or soon to be published) results.

2018-07-10 on infoproc

Plots would look much noisier if not for averaging many individuals into single point. Keep in mind that socioeconomic success depends on a lot more than just cognitive ability, or even cognitive ability + conscientiousness.

But, underlying predictor correlates ~0.35 with actual educational attainment, IIRC. That is, the polygenic score predicts EA about as well as standardized tests predict success in schooling:


This means you can at least use it to identify outliers: just as a very high/low test score (SAT, ACT, GRE) does not *guarantee* success/failure in school, nevertheless the signal is useful for selection = admissions.

2018-07-06 on infoproc

There are ~900 points and it’s hard for the eye to see the actual distribution in the denser regions. The outer regions away from the fit line are easy to make out because individual points are distinguishable. The denser regions may follow the fit line better than the outer ones.. otherwise I agree, the fit line looks problematic.

2018-06-26 on infoproc

Google for the controversy at Duke over his analysis. IIRC there was a faculty letter denouncing him and Brodhead did not defend his academic freedom very strongly.

2018-06-26 on infoproc

Brodhead, the recently retired Duke president (2004-2017), was formerly Dean at Yale. Duke has been moving in the cookie cutter direction 8-|

There was a big controversy over Arcidiacono’s analysis of the Duke admissions / grade data you mentioned, and IIRC Brodhead failed to defend him against angry faculty that didn’t like the conclusions.…

2018-06-25 on infoproc

Presumably the objective function used to fit the logistic regression parameters is roughly the correct calling of Admit, Waitlist, Reject status, summed over the entire set of individuals.

This is not guaranteed to get the ethnic ratios exactly correct because the actual admissions process is not perfectly modeled by logistic probability functions, etc.

2018-06-21 on infoproc

Harvard prevails in court, media ignores the detailed evidence, business as usual.

Fitz quietly retires.

2018-06-21 on infoproc

That’s one plausible outcome. It depends a lot on how the media spins this. As we know from other (very current) examples, outrageous things that would have been scandal of the decade in the past, are now hidden in plain sight.

2018-06-21 on infoproc

There are some disease risk traits as well as cosmetic traits (skin color, eye color, baldness, etc.) that may have sparsity in the hundreds. That is, only of order hundreds of variants control most of the heritability. Is that also consistent with “~100k”?

Are the authors familiar with the plant and animal literature in genomic prediction? Or is it just humans that are supposed to be omnigenic?

2018-06-21 on infoproc


“Every genetic variant affects every trait!”

“Uh, how much? If the effect size is infinitesimal for all but a small fraction of variants, which account in aggregate for almost all of the heritability, then the genetic architecture is actually sparse.”

“I don’t understand what you just said. BTW, I’m the real biologist here!”

2018-06-21 on infoproc

There are already counterexamples from phenotypes like height.

2018-06-20 on infoproc

1 and 2 probably denote different things like “Alumni interviewer rating of academic strength” vs “Alumni interviewer rating of personality” … or even a composite of several different scores for each of 1 and 2.

2018-06-18 on infoproc

To everyone in this thread: click the links above in the post and you can find a lot of the data you are curious about. It may take some digging because the statistical reports are fairly long, but you can find both the Harvard and SFFA analyses…

2018-06-16 on infoproc

Wow! I was unaware of this. Thx.

2018-06-11 on infoproc

Thx! I saw that and should have linked to it in my post. Made me want to visit Autun even more 🙂

“THIS BLUE, INDOLENT TOWN. Its cats. Its pale sky. The empty sky of morning, drained and pure. Its deep, cloven streets. Its narrow courts, the faint, rotten odor within, orange peels lying in the corners. The uneven curbstones, their edges worn away.”

2018-05-28 on infoproc

I don’t really crave social isolation — I’m pretty extraverted and don’t find it hard to isolate myself in my office or study to get work done. I don’t need to be physically far from other people.

The isolated locations of the homes featured above are pleasing mainly for aesthetic reasons. The locations are beautiful and would be spoiled by having too many other homes around. If I lived there I would want a very fast internet connection and (ideally) a transporter wormhole that can take me to the city when I want 🙂

2018-05-28 on infoproc

The ship is a cool idea. I wonder how hard they are to keep afloat?

There is a lot interest in underground containers. If you go down far enough the temperature is stable year round.…

2018-05-28 on infoproc

>> my postdoc will be in genetics not molbi. Exciting times… << Congrats, Dr. Stovner 🙂

2018-05-28 on infoproc

> human-like intelligence and experiences require the future AI to have a human-like body and immersion < It seems possible that the most efficient way to train a human-like intelligence is to immerse AIs in artificial worlds and give them human like experiences and problems to surmount. OpenAI and other groups are already working in this direction, IIUC. Of course (as we will see) there are many formidable intelligences which are not "human-like" 😉 My problem with Searle is that it is rather obvious that the larger machine can process information correctly (i.e., translate from Chinese to English) without any specific sub-component having to "understand" what is going on. This was made obvious by Turk working examples, and most recently by actual neural nets which do well at machine translation with (obviously) no specific "synapse" or node/connection "understanding" anything but a small step in the process. His insertion of a small philosopher into the scenario (with a mechanical job like sorting papers) is just confusing to the naive intuition, and (in my opinion) irrelevant to the overall question.

2018-05-28 on infoproc

I really know very little about this subject. But I am old enough to remember when CDs first came out and there was a lot of discussion about the digital sound being distinguishable from the “warmer” vinyl sound we were used to. Perhaps we just liked the analog sound better because we were used to it, even though the fidelity is, in fact, worse than digital.

Nowadays I listen mostly to MP3s and usually in environments (car, while working out using cheap earbuds) in which fine distinctions are not possible. But it does seem to me that MP3 quality can be kind of crappy.

On a slight tangent, I also notice that call quality on cellphones or skype really sucks compared to the landline experience I grew up with. I sometimes wonder if there isn’t a business opportunity for provision of high quality voice communication for people willing to pay for it…

2018-05-18 on infoproc

For the type of digital cash currently used in PRC this is true. Cryptocurrency, however, would not have this property. This makes it plausible that PRC will not permit cryptocurrency to gain a foothold there…

2018-05-18 on infoproc

Moto G5 Plus — fantastic phone for about $200!

Mac Air.

2018-05-17 on infoproc

Yes, they are performing very complex optimizations and this is the amount of work done = total number of floating point operations performed.

2018-05-17 on infoproc

petaflop/s-days = day of sustained use at petaflop/s (= 1E15 ops per second)

~ 1E20 floating point operations.

2018-05-17 on infoproc

The Onion can never top that… 😎

2018-05-11 on infoproc

Ha ha — yes, you have a point.

2018-05-11 on infoproc

They probably are somewhat cherry picked, but given that Sundar says this feature of Google Assistant will be released later this year they must be confident that they can get it working reasonably well soon. They are already taking heat in the media over not telling people on the other end of the call that they are talking to a robot, and have conceded that they will issue a warning to that effect at the beginning of each GA call.

2018-04-30 on infoproc

But think of all the kids from his background that might have thrived under better conditions. Lee is obviously a special guy — not everyone can overcome.

2018-04-30 on infoproc

Jiujitsu is chess with the human body 🙂

MMA is jiujitsu (and more), with someone trying to punch you in the face…

2018-04-20 on infoproc

IIUC, accepted in Nature Genetics, should appear very soon.

2018-04-20 on infoproc

Latest study by ssgac.

2018-04-19 on infoproc

Only a small number of people, even now, understand that the hand writing is on the wall. In 2013 almost everyone who was not an expert in the field was taking a wait and see attitude. In 2018 (with EA3 about to appear in Nature Genetics) it is much easier to get that something is happening…

2018-04-10 on infoproc

Not quite 0.9 as the SD is a bit over 3cm IIRC.

When you include the gender information the correlation between predicted and actual height (as shown in the scatterplots) is higher than 0.6-0.7, which is what the predictor reaches on z-scored heights within each gender.

2018-04-03 on infoproc

I originally posted a link to a full version of the 90min documentary, but YouTube has already blocked it. Reload the blog page and it links to a version that you have to pay for…

2018-04-02 on infoproc

Value to consumer atm is not expected +Delta in g, but reduction of risk of very negative tail event.

Note as young women start to freeze eggs early (covered by some health care plans), the typical number of embryos to choose from will go way up.

Finally, eggs from stem cells might not be too far in the future.

2018-03-26 on infoproc

Not a minus sign. $\sim$ or tilde means approximately.

I just changed it to $\approx$ so it will be more readable.

2018-03-21 on infoproc

Your pal entered the PhD program at Yale when I was a prof there. I knew him.

2018-03-15 on infoproc

Also important and in a sense even more fundamental since it suggests we really don’t understand QM itself, let alone QM + GR.

Personally I would say that Bell’s questions were all answered in Many Worlds (Everett), but MW itself has a measure problem (origin of probability or Born’s Rule) that I myself have worked on.

Even fewer people recognize this problem at the root of QM (although Steve Weinberg is one of them) than worry about BH information!…

BTW, Hawking was 100% a MWer (realist). He once said (IIRC) that MW follows trivially if one just requires that QM applies to every degree of freedom in the universe. This is of course correct, but then we are left with the measure problem I mentioned above.

2018-03-15 on infoproc

As far as physics, I would say that Hawking’s contributions are much deeper and across a broader range of topics.

Many theorists (including yours truly) believe that the black hole information problem is the deepest paradox of fundamental physics: it points to a conflict between general relativity (existence of black hole horizons, modified causal structure of spacetime due to gravitational collapse) and quantum mechanics (unitary evolution of quantum states, preservation of quantum information, time-reversal invariance of qm). Hawking noticed this in 1975 and it took others a very long time to appreciate the problem.

2018-03-07 on infoproc

Yours is the advanced version of the lesson 🙂

2018-03-04 on infoproc

Of course a lot depends on how long the good times will continue for the top tier Big Tech firms…

2018-03-04 on infoproc

It is easy to find videos of congressional hearings which show FBI, DOJ, NSA, CIA, etc. directors and senior execs misleading their oversight committees, answering questions in a deliberately deceptive manner. Anyone who has read the material in my post and then listens to these hearings can tell that the agencies are not being up front with the questioners from the legislative branch.

People forget that this mass surveillance without oversight is a very recent development. The necessary technology has only been around 10-20 years, 9/11 and terrorism helped push it forward, and now it has been used to spy on political opposition. You can argue that Obama DOJ/FBI had good reason to spy on the Trump team, but you can’t deny it was spying — using incredibly powerful tools that are hidden from public oversight.

Much more here:…

2018-03-04 on infoproc

Read the FISC report! No controlling authority – IG admits can’t track queries to upstream surveillance data. Only *hope* to get coverage of “most commonly used tools”! You know what that means — they don’t have an actual system in place to track access. Smart Snowden types can go in and do what they want.

2018-03-04 on infoproc

I think it’s clear they wanted the surveillance information distributed as widely as possible so that it couldn’t be hidden by the Trump administration after they took power. But the long term consequences are potentially quite negative. More people querying the data! What could go wrong?

2018-03-02 on infoproc

I should have just linked to this old post:…

2018-03-01 on infoproc

The point is that most people (including many scientists) don’t have a good intuitive understanding of variance and correlation. 0 and 1 they understand, but they don’t know how to think about, e.g., 25 percent of variance. Still sounds small, but the prediction may correlate 0.5 with actual value, which is pretty good, esp. if you are mainly trying to identify outliers.

Squaring a number less than 1 makes it even smaller… gee, what does that mean…

2018-03-01 on infoproc

This comes from the silly variance vs correlation confusion.

Variance is the sum of *squared* deviations so it doesn’t even have the same units as the underlying quantity. It has nice additive properties (you can add variances of independent variables) which is why people like to use it, but it is a confusing measure of predictive power.

Take the square root of variance 0.10 to get ~0.3 which is the *correlation* that current best genomic predictors have with actual cognitive ability (or educational attainment). This is similar to college GPA vs SAT score correlation — not great, but not that bad either. (Note, if you correct for range restriction, difficulty of courses taken, etc. the GPA-SAT correlation is significantly higher.)

I have explained this to journalists and genomics people dozens of times, but usually it does not get through…

2018-02-27 on infoproc

Still not “clinically useful” but it’s only a matter of sample size…

2018-02-26 on infoproc…

Brainiac’s most consistent power (endemic to all versions) is his “twelfth-level intellect”, allowing calculation abilities, enhanced memory and advanced understanding of mechanical engineering, bio-engineering, physics, and other theoretical and applied sciences, as well as extensive knowledge of various alien technologies. For comparison, the population of 20th century Earth as a whole constitutes a Sixth-Level intelligence and the population of 31st century Earth as a whole is a Ninth-Level intelligence. …

Among organic beings, Brainiac views only his frequent partner Lex Luthor as a peer intellect. Brainiac is usually depicted with an incredibly-high degree of super strength and durability; the exact level varies, but usually hovers at around Superman’s strength. Brainiac’s exact abilities vary drastically throughout his various incarnations.

2018-02-23 on infoproc

SVMs and other nonlinear methods have been used, but as far as we can tell most of the variance is additive. The extra variance captured by nonlinear predictors is typically not large, but of course this will be examined more carefully now that we are making strong progress on the linear part.

2018-02-22 on infoproc

I’m not that picky for the purposes of ML research. As long as someone has experience with a modern language (even, say, Python) they can pick up other languages. But I am looking for significant experience, not just a course or two.

Re: lattice, I would find some good review articles on arxiv and look at the software packages used. If you are serious about this you could email one of our 3 lattice people at MSU (one of the largest groups in the US!) for more information.

2018-02-16 on infoproc

YouTube will generate one automatically, I think. Come back later and click on the “…” on the right under the video.

2018-02-11 on infoproc

You called it! 🙂

2018-02-07 on infoproc

Predictors that allow parents to select *against* large negative fluctuations in IQ could be on the market within a few years at most. If sufficient data for large cognitive ability GWAS continue to accumulate then more precise point prediction of IQ will be possible in (say) 5 years or so. But this could take longer if the data are not acquired — most large GWAS today do not measure cognitive phenotypes.

The wet lab technologies (DNA amplification, inexpensive genotyping, etc.) already exist, but are currently only used in embryo genetic testing against chromosomal abnormality (Downs), specific Mendelian conditions (e.g., Cystic Fibrosis), etc.

Any good IVF clinic in the world can send samples for the genetic testing described in the paragraph above. It is a simple matter to extend the testing to quantitative traits once the predictors (derived from very large sample GWAS data) are available.

2018-02-03 on infoproc

This doesn’t directly address what you wrote, but might be of interest:

2018-02-01 on infoproc

If nothing else the particle interacts via gravity (gravitons).

2018-01-28 on infoproc

Could be this:…

… and other related stuff, like physicists reformulating feature selection in terms of (what we call) renormalization group flow or coarse graining.

2018-01-27 on infoproc

1) The application rate doesn’t figure directly in any analysis, although for those familiar with recent history it provides some circumstantial evidence. 20 years ago a much smaller fraction of the applicant pool was A-A and typical strength of application was lower. Yet the fraction of A-A admitted students has not moved much.

2) Sure, non-elite whites might have a case but you have to look at actual admit rate controlling for strength of application. If you don’t apply, you can’t be admitted — some groups are more likely to apply to H than others, so national population fraction is not the only factor.

3) I don’t think genomic predictors will ever be as good as having actual measurements (test scores, grades, evaluation letters, contest performance, extracurriculars) and *using them properly* as predictors.

2018-01-25 on infoproc

I think gcta ceiling may be too low because often sample used does not have high quality phenotype — only crude g score. Could do a bit better if actual large sample has better g measurement. But we will see..

2018-01-25 on infoproc

Selection in the 2020s and editing in the next generation (Nativity 2050)? I don’t think that’s aggressive at all.

2018-01-24 on infoproc

IIRC he transferred to Harvard and graduated in Physics 🙂 He’s doing tech startup things (perhaps related to machine learning) in SV, I think.

2018-01-24 on infoproc

I think you have asked this question before.

One of the stronger arguments demonstrating discrimination is to note that the overall applicant pool of Asian-Americans is much stronger than, say, 10-20 years ago (in fact it is both more accomplished and also much larger — see Blum’s comments about this in the video). This is across the board, not just in STEM. Yet representation at Harvard (or other elite private college that practices non-race blind admission) has not increased very much. One could compare Harvard to, say, Berkeley instead of Caltech. Once Berkeley was forced into race-blind admissions (by CA Prop 209) the A-A percentage shot up. Why would this not happen over the years at Harvard or other places, unless there was negative selection (a de facto quota) on A-A admits?

Blum says about half of all Harvard applicants each year are Asian, and that 55% of >2300 SAT applicants are Asian. So the admit rate for Asians, controlled for strength of application (and even for intended major, legacy status, etc.), might be much lower than for other groups. Only a careful analysis of the data can tell us for sure. See about 25-30min into the video.

2018-01-24 on infoproc

It’s not easy to do this counting. Ron Unz tried but I think there was a lot of disagreement about how it comes out.

2018-01-24 on infoproc

I think you can sort of “select all — copy” in the transcript window and paste the text into another window if you want to read it fast.

One could also listen to the video at 1.5x speed…

2018-01-24 on infoproc

All the talks are up on YouTube now and Google auto-generates transcripts:…

2018-01-17 on infoproc

0,1,2 are actually “dosages” of the SNP variant. In our diploid genomes we either have zero, one, or two copies (i.e., from both mom & dad) of a particular variant. Having said that there is nothing to guarantee that some clever transformation of the data won’t improve the results…

2018-01-15 on infoproc

That’s not a trait that has been quantified very well, AFAIK. You need to first operationalize what you mean by hand motor control, develop tests/measurements, validate them (are scores relatively stable? heritable?), and then collect data on lots of people who will also be genotyped. UKBB has grip strength as a phenotype (measured using some device), which is heritable. We have even built a crude predictor for it.

2018-01-14 on infoproc

Are they using Thought Vectors (or something similar)?

2018-01-14 on infoproc

Taking the most charitable interpretation, I would say Woit has been misled as to the nature of this research, and had he actually attended the talk he would have found it interesting and unobjectionable.

2018-01-14 on infoproc

Effect sizes are given in units of population standard deviation. For males that would be about 7cm or 3in — i.e., effect size of beta = 1 is about 3in.

2018-01-11 on infoproc

Michael below is correct. I should add that

1. One can’t get within 3cm error from parents’ heights alone. I have growing kids and have looked into all the height predictors which use parent and kid growth chart trajectories, etc. Getting within an inch is not easy.

2. The real advantage of genomic prediction is being able to tell which person (or embryo) is likely to be an outlier — e.g., at esp. high risk for a disease, or well below average for some phenotype.

2018-01-07 on infoproc

Just north of Pismo – Shell Beach. Similar landscape (Avila) to the north, though. Happy New Year!

2018-01-02 on infoproc

Not sure I understand — are u saying Buffet knew already the Paulson bailout would pass, or that Paulson would protect GS?

2018-01-02 on infoproc

Thanks! 🙂

2018-01-01 on infoproc

Happy New Year to you! I hope your investing success continues in 2018 🙂

Perhaps the magic of compounding will make a philanthropist of you someday…

2017-12-26 on infoproc

Thanks – likewise to you and your family! We’re having a ball near SLO – 70 and sunny today. Went for a long hike 🙂

2017-12-25 on infoproc

In this case we could be talking about a single interbreeding population, such as white Americans or British.

2017-12-23 on infoproc

This is not a dumb question at all.

The implication of heritability ~0.8 and small GxE is that a big chunk (the majority?) of the *observed* IQ difference between rich and poor (for simplicity just restrict to white people) is due to genetics. Equivalently, most of it is *not* due to SES or other environmental effects. A further inference would be that society is at least somewhat meritocratic and that genetic endowments come into play: rich people are (in part) richer because they have (on average) higher genetic potential for cognitive ability, and this is passed on to their kids through DNA.

Of course these are very complicated issues still under active research and it’s difficult to be sure. But the rough picture that emerges from modern Behavior Genetics is what I described above.

2017-12-16 on infoproc

Thanks for this additional information!

I bet if you email the authors you might get them to explore further within the top few percent to see if there is any evidence of diminishing returns to IQ. As you note the figure I posted suggests no such effect, but it would be nice to see sqrt(N) error bars as you get into bins with small total population of individuals/inventors.

Judging from the first figure I posted above, there appears to be a saturation effect wrt top family incomes – once you are in the top few percent in income the probability of invention seems to stop increasing. Comparing income vs IQ in this way might be very interesting. Email the authors and ask them about it! 🙂

2017-12-16 on infoproc

Actually, Sam understates the case. In Beijing no one would hesitate, in private conversation, to question the communist party or some action by Xi Jinping. There are basically no restrictions on what you can discuss in China — people are “reality based”! You only get in trouble if you *widely broadcast* anti-government views through social media or other platform.

The US is a different story entirely. You can easily get an angry emotional response from someone who is supposedly highly educated and rational by asking simple questions like “How reliable are IPCC climate projections? Can we really model such complex phenomena?” or “Do students admitted through preference with SAT scores hundreds of points below the rest of their class have a good shot at succeeding in challenging majors?” 🙁

2017-12-09 on infoproc

MSU had enrollment increases over the last decade, and recognizing this increase as quasi-permanent created additional recurring budget to increase the size of the faculty.

2017-12-07 on infoproc

Working in a big organization sucks, even if you are near the top.

However, I view it like donning a big exoskeleton or robot suit: uncomfortable and unwieldy, but it allows you to move huge objects that previously seemed immovable.

I take a lot of pride in having reshaped the research program at a major university, helping to create some new things (IQ, CMSE, FRIB, more), hiring lots of talented faculty who are poised to make discoveries.

As one small example, I see the potential to link lattice QCD (rigorous computational results from QCD) to experimental results in nuclear physics (FRIB). MSU now has one of the largest lattice QCD groups in the US, and nuclear theorists who understand effective field theory, nuclear physics, and QCD. I believe this will result in fundamental advances over the next decade. Another example: I believe I have tripled (depends on how you count) the number of faculty who work on ML and AI, including in autonomous vehicles, genomics, deep learning, etc. etc.

Suppose I sat in a conference room with a bunch of young physicists, and asked: If you were in charge of this university, what would you do? What would you create or change? It would only take a few minutes of thought before all kinds of ideas came bubbling up. Executing on those ideas is difficult, but can be satisfying.

2017-11-30 on infoproc

Yes, very nice.

2017-11-30 on infoproc

I’m afraid I don’t know the answer to this question. Perhaps other commenters do.

It’s tough to compare online masters programs because they can change very rapidly.

2017-11-28 on infoproc…

I don’t detect fear of imminent collapse from well-placed people I’ve talked to in China (academics, investors, business people). They could all be mistaken but I’d weight their opinions over most US prognosticators.

2017-11-23 on infoproc

This experiment has been done and the results are clear.……

2017-11-18 on infoproc

I have done so many times but he returns under different usernames .

2017-11-10 on infoproc

The rare isotope accelerator DOE is building (for nuclear physics) on the MSU campus will cost about $750M; over a billion when operating costs are added in.

So, yes, genomics can be relatively cheap if you do it right. OTOH, critics of big science in biology have claimed that returns from genomics, GWAS, etc. are limited so far.

2017-10-29 on infoproc

That’s an amazing story from Minsky!

I like this, from his Wiki page:

He was an adviser[26] on Stanley Kubrick’s movie 2001: A Space Odyssey; one of the movie’s characters, Victor Kaminski, was named in Minsky’s honor[27] and Minsky himself is mentioned in the movie and in Arthur C. Clarke’s derivative novel of the same name:

Probably no one would ever know this; it did not matter. In the 1980s, Minsky and Good had shown how neural networks could be generated automatically—self replicated—in accordance with any arbitrary learning program. Artificial brains could be grown by a process strikingly analogous to the development of a human brain. In any given case, the precise details would never be known, and even if they were, they would be millions of times too complex for human understanding.

2017-10-26 on infoproc

Thanks! I like the berries frozen because they cause the milk to thicken and become creamy. I hope to try some of your recommendations!

2017-10-26 on infoproc

I’ve made this point in the past: you need much more compute to discover/train the AI than to run it after the dust settles. Evolution can be thought of as a very expensive computation that resulted in our highly energy efficient meat brains.

2017-10-26 on infoproc

Now consider the massive advantage in connectivity that our brains have over any extant (Silicon) NN… As compute power and algorithms continue to improve, we can expect shocking leaps in NN performance!

2017-10-25 on infoproc

Not strictly convex, but tractable:…

2017-10-20 on infoproc

You might expect that the spinach gives it a particulate leafy flavor, but it actually blends to a smooth creamy consistency which is sweet because of the blueberries and the (e.g., vanilla) flavor of the protein powder.

Try it and report back! 🙂

You can buy pre-washed bags of spinach, and frozen blueberries last a long time, so the logistics for these smoothies is not difficult. I usually have 1-2 per day and my wife and kids like them as well.

2017-10-10 on infoproc

Although I don’t think Feynman would have scored as high as Schwinger on a test like the old SAT-V, his overall brilliance was apparent to anyone who talked to him. So I doubt you really had to rely on the limited evidence you mention about your girlfriends… 😎

I am sure Feynman would have argued that tests like the SAT-V are “dopey” because they concentrate on obscure words of little practical value. He would probably have asked why “eigenvalue” or “ergodic” (which describe novel, nontrivial concepts) never appear on the test, versus words that are merely arcane stand-ins for common words describing common concepts.

2017-10-02 on infoproc

You are tempting me into looking more closely at the later books 🙂

I like your idea about the fixed point in time plot structure. As you know there are hints that Herbert had in mind some higher order power guiding the Bene Gesserit without their knowledge…

2017-10-02 on infoproc

Just the primitives, not all possible complex thoughts. But it is interesting that a deep net can learn to parse syntax very well, mapping sentences into an internal feature/concept space. Human designed syntax parsers don’t do nearly as well, AFAIU.

2017-09-29 on infoproc

Dweck and other enthusiasts have probably overstated potential gains from (what they call) growth mindset, but that doesn’t mean it’s a bad idea to “believe talents can be developed through hard work, good strategies, and input from others”!

One just has to be realistic about what is possible.

2017-09-29 on infoproc

I think the story is that Watson had/has a modest IQ, not Crick. Watson had trouble with math, but entered the University of Chicago at 15 and obtained his PhD (IIRC) at 21. So I don’t know what to make of the rumor; it may be the mirror image of the Feynman story.

2017-09-29 on infoproc

“And you in your pride thought you could produce the Kwisatz Haderach!”

Someday this science will be very highly developed and these crude early steps will be remembered only dimly……

2017-09-16 on infoproc

I was too early as a dollar bear. I was worried about the dollar already 10 years ago, but it turns out the Euro had even more problems.

Now the Yuan is a candidate reserve currency. Well, maybe, but Capital probably still has more confidence in the dollar and will for some time (how long? no one knows). Note that the Yuan-denominated oil futures contracts are somehow also convertible to gold. While these instruments allow oil trades to circumvent the dollar, the gold convertibility (not sure exactly how this will work) is probably necessary right now due to lack of full confidence in the Yuan.

2017-09-15 on infoproc

Our predictor, constructed using the (linear) L1 methods described above, captures variance comparable to the estimated SNP heritability from GCTA (GREML).

I don’t want to comment in more detail as the paper is under review and some journals are not supportive of the arxiv / biorxiv culture of preprints that I prefer.

2017-09-11 on infoproc

I am sorry to disappoint. Perhaps you can watch the 1:45min video at top and tell me *specifically* what you disagree with Bannon about. Then we can have a discussion as opposed to just virtue signaling conformity.

The longer 27min video is also fair game; I agree with *most* of what Bannon says there as well. Bill Clinton, for example (i.e., any 1990s centrist Democrat), might as well.

2017-09-11 on infoproc

Bray has to explain repeatedly to Remnick that the goals of Antifa are not the goals of the Democratic Party.

Remnick is smart enough to understand that there are radical and violent leftists in the world, but he wants to believe the narrative that the only bad guys in C’ville were on the Right.

“They have no allegiance to liberal democracy”…

2017-08-21 on infoproc…

Probably not omnigenetic in the following sense: projections from existing results suggest there are a finite number of variants that account for most of the variance in the population. That is, a good predictor for the trait need only include thousands or tens of thousands (but not more) variants.

There are many millions of relatively common genetic variants in the population. Most have negligible effect on any specific trait, at least according to the paper above and calculations I have done.

2017-08-21 on infoproc

Thanks for the videos 🙂

We’re working on the draft of the CS applied to 500k UKBB paper. But if we get it into a top journal it will be embargoed from, e.g., bioarxiv, so you won’t see it until it appears in the journal.

2017-08-19 on infoproc

More Than This is one of my favorite songs! I like to think that it’s about chance and determinism, but you can interpret it many ways.…

Just to clarify — we’ve analyzed the 500k UKBB dataset release (July 2017) for height and everything I predicted is correct. We just need to get some clean cognitive score data on a large number of people FTW! (Or gradually get there via proxies like EduYears…)

2017-08-19 on infoproc

Not yet, but 20-30 years from now people will begin to wonder…

2017-08-17 on infoproc

It’s not widely known, but AlphaGo v1 (that beat Lee Sedol) had weaknesses to certain lines of play. Deep Mind were lucky that Lee did not discover them during the match.

If OpenAI can’t fix these problems in their bot then they have no hope of getting to 5v5…

2017-08-17 on infoproc

Emless and Philipp are correct.

1) “candidate gene” means a candidate selected using some kind of human intuition (about biochemistry, biological structure, etc.) and followed up with relatively small sample tests (p < 0.05 significance threshold). These studies were in vogue from early 2000s until a few years ago. But almost none of their results replicate. 2) GWAS = weak priors and pure statistical power. SNP hits found using the p < 5E-08 "genome wide significance" threshold replicate routinely in follow up studies with different populations. #2 works and the results show that #1 does not. There are many SNP hits from #2 that are under intense investigation now (see Emless links below), but they are seldom the "candidate genes" originally suggested by #1. Researchers are getting around to writing papers pointing out what I just wrote above in specific contexts like schizo (e.g., the paper you linked to in your first comment). For smart people this has been obvious for almost 5 years now. There has been a lot of resistance from people who claim to have the magical "biological intuition" used in #1 that couldn't possibly be surpassed by math geeks using raw compute power and statistics (rolling eyes). Sorry I can't explain this any better...

2017-08-17 on infoproc

Good question, although the answer is a bit complicated. CS is better and produces better prediction at the same sample size (amount of data). Our phase transition prediction was ~30s (s= sparsity) for h2=0.5, so that would be 600k for s=20k SNPs. At that point the activated SNPs would capture almost all the variance but the assigned effect sizes wouldn’t be quite right, and would need still more data to converge to more accurate values (move deeper into the “good phase” on the diagram).

These JHU guys also look at whether one should stick to the p<5E-08 threshold for activated SNPs or relax the threshold to get somewhat better prediction. For 5E-08 threshold they would predict you need perhaps 2 million individuals to get everything. Note in the middle graph above the vertical axis is percentage of *genomic* variance (not total variance) captured; extrapolating their curves, one has to go out beyond 1 million to get 100%. I don't think we can predict *exactly* how much data is required to "solve" these phenotypes, but ~million is clearly the right ballpark for h2~0.5 and ~10k causal variants. I made this prediction several years ago and at the time everyone just laughed at me, but in the end I will be proven correct. Our current results with height already show that I was correct -- stay tuned 😎 I just wish we had more clean data with cognitive scores 🙁

2017-08-17 on infoproc

In my experience most former wrestlers love BJJ, once they realize how effective it is — i.e., have been tapped out a few times under realistic conditions 🙂

I prefer no-gi myself.

2017-08-16 on infoproc

Wow, I am impressed 🙂 I keep hoping my son or daughter will get interested so I’ll have some impetus to start training again…

Be sure you are taking BJJ as other variants of JJ are not as good, imo.

2017-08-16 on infoproc

Thanks for the link. The “25 historical candidate gene polymorphisms” that they study are ones that were proposed long ago based on intuition of biologists/medical researchers as likely to be linked to schizo. But modern GWAS show that these suspected polymorphisms do not have significant effects on schizo risk, unlike the variants discovered via GWAS. The article is not negative about GWAS; it is *using* GWAS results to show that human researcher intuition about schizo mechanisms and biological pathways (used to identify “candidate genes” in the past) has little value.

“The history of schizophrenia research should serve as a cautionary tale to candidate gene investigators examining other phenotypes: our findings indicate that the most investigated candidate gene hypotheses of schizophrenia are not well supported by GWAS, and it is likely that this will be the case for other complex traits as well.”

However, modern GWAS results routinely replicate in validation tests and follow up studies. I’ve discussed this history in other posts:…

In the title of this post I mention Candidate Gene Studies. Forget, for the moment, about goofy Social Psychology experiments conducted on undergraduates. Much more money was wasted in the early 21st century on under-powered genomics studies that looked for gene-trait associations using small samples. Researchers, overconfident in their vaunted biological or biochemical intuition, performed studies using p < 0.05 thresholds that produced (ultimately false) associations between candidate genes and a variety of traits. According to Ioannidis, almost none of these results replicate (more). When I first became aware of GWAS almost a decade ago, the field was in disarray, with some journals still publishing results at the p < 0.05 threshold, whereas others having adopted the corrected p < 5E-08 = 0.05 x 1E-06 "genome wide significance" threshold (based on multiple testing correction for 1E06 SNPs)! The latter results routinely replicate, as expected. Clearly, many researchers fundamentally misunderstood basic statistics, or at least were grossly overconfident in their priors for no good reason. But as of today, genomics has corrected its practices and although no one wants to dwell on the 5+ years worth of non-replicable published results, science is at least moving forward. I hope Social Psychology and other problematic areas (such as in biomedical research) can self-correct their practices as genomics has.

2017-08-12 on infoproc

> I’m happy to concede that there is a pro-female bias… at the time of hiring. But after that, it is not like women enjoy a 2:1 advantage on other issues once they are hired. < I doubt anyone at Google HR would concede what you just did. They'd fire you for what you wrote! The post-hiring career and pre-hiring pipeline situations are the subject of hot debate! Most people discussing the issue, and, sadly, many of the researchers, have biases. It would be fun to revisit this issue after you've had +20 years as a professor and perhaps done some administrative work. I think you'll see that the preference survey results I linked to showing significant asymmetries are realistic and that the asymmetries matter for how individual careers progress. (I'll probably be senile, but by then you'll know the score 😉

2017-08-12 on infoproc

Yes, some of us are more focused on Damore and whether he got a fair shake at Google, and whether this case might go to court and provide a national spotlight on these issues.

The larger questions like “What would female representation in tech / STEM be if there were no bias?” are complicated and most people just bring their priors (i.e., are highly biased in one direction or another).

For example, have you looked at the STEM hiring paper I linked to? There is considerable evidence that female candidates are highly favored in STEM (and tech) hiring. As someone whose job (in part) is to *incentivize* departments to hire more female and underrepresented minority faculty I am intimately familiar with this issue! So forgive me if I think the Vox people are just talking their priors without looking at the available evidence. I know how HR actually works at companies and in the academy.

At universities it is (sometimes) *explicitly acknowledged* that we might hire / admit weaker candidates for diversity purposes. We might say that it’s in the long run interest of the school, or the department, to have more female or minority role models, advance social justice, etc. But the tilting of the playing field is *acknowledged* and can (sometimes) be *debated* among professors and administrators. This is what Damore was asking for at Google — it got him fired!

TL;DR What is the *sign* of the obvious bias you refer to?

2017-08-12 on infoproc

Damore acknowledged all of your points, but was still fired. The conversations reported between top G execs show that they don’t understand the Jeremy Lin point. See the Recode article on the firing decision. In particular, you could accept that line of reasoning without it suggesting that female engineers already at Google are somehow less able than their male counterparts.

BTW, did you read the memo?

“Differences in distributions of traits between men and women **may in part** explain why we don’t have 50% representation of women in tech and leadership. …

I hope it’s clear that I’m not saying that diversity is bad, that Google or society is 100% fair, that we shouldn’t try to correct for existing biases, or that minorities have the same experience of those in the majority. My larger point is that we have an intolerance for ideas and evidence that don’t fit a certain ideology. I’m also not saying that we should restrict people to certain gender roles; I’m advocating for quite the opposite: treat people as individuals, not as just another member of their group (tribalism).”

> female colleague in STEM, what is the relative magnitude of these effects?

This is a highly studied subject with lots of published papers…

Cited in this earlier post:…

Note the preference surveys don’t assume biological causes. They do, however, strongly suggest that a smaller fraction of women want to put up with what it takes to be startup founders, tech leaders. It may not be as simple as basketball, but the basic statistical point stands.

The point of the blog post above was to show that media (and even “serious people”) claims that Damore relied on “outdated bro pseudoscience” are contradicted by the actual scientific literature.

2017-08-11 on infoproc

If it goes to trial it could be quite a sensation. The press coverage is already intense.

Imagine a national referendum on both the science in the memo and the social issues of diversity, hiring preferences, political correctness, etc. Steve Pinker and Ceci and Williams called as expert witnesses …

2017-08-02 on infoproc

I think the idea of prioritizing H1Bs by the compensation offered is not a bad idea, as long as one considers industry and academia separately, etc.

2017-08-02 on infoproc

Thanks for that link. It doesn’t seem to contradict the specific result concerning robots and job loss, but rather notes that effects of globalization might be much larger:

3. Globalization is a much bigger deal than automation for work and wages. The economists Daren Acemoglu and Pascual Restrepo recently published a paper that found some evidence that automation has already nibbled away at work and wages. They concluded that, for every new robot per 1,000 workers, six people lose their jobs and employment fell by about one percent.

But Mishel and Biven call hogwash. They point out that, in Acemoglu and Restrepo’s own paper, “job displacements caused by trade with China in the 2000s were four times as large as their estimate of job loss due to robots.” This echoes other economists, like David Autor, who have argued that globalization, and especially trade with China, has had much more to do with stagnating wage growth than anything related to technology.

2017-07-23 on infoproc

At the beginning, best to focus on excluding really negative conditions like Downs or outlier risk for disease conditions. As you note it won’t be possible to be confident about detailed tradeoffs until we know much more.

2017-07-16 on infoproc

“phenotypic N-S gradient discussed here”…

2017-07-08 on infoproc

No sane person could disagree with “we must secure our borders, we must enforce our laws” (SJWs of the current year are not actually sane).

So pretty much every serious politician has to say those things. But only Trump is carrying them out!

2017-07-08 on infoproc

The stated policies — “we must secure our borders, we must enforce our laws” are not different. The difference is whether the rhetoric is backed up by action.

2017-07-03 on infoproc

Prosopagnosia is heritable, IIUC. So variations in this particular brain function are almost certainly influenced by genetics.

2017-07-03 on infoproc

>> your opinion of Chinese professors mostly hiring Chinese graduate students, Indians hiring other Indians, etc. << This could be the result of ethnic nepotism. But it could also mean that the professor is making use of a competitive advantage: he/she has contacts in their country of origin and can recruit talented students from there. If I came from Caltech and had a pipeline to recruit superior grad students from there I would take advantage of it. Substitute China, India, or Russia for Caltech and it doesn't sound so sinister or negative... In the case of this research my suspicion is that one or both authors have better math chops than the typical lab neuroscientist because they were able to design their experiment to focus on features in the space of facial variation. This combination of skills might be easier to find in students/postdocs originally trained in China where the math education is stronger. Most neuroscientists in the US know very little math and would not understand at all how face recognition algorithms work.

2017-06-29 on infoproc

Some clarification:

1. US carriers very vulnerable to *conventional* Russian and PRC missile (cruise, ASBM) weapons.

2. Within ~10y (i.e., well within projected service life of US carriers) I expect missile systems of the type currently only possessed by Russia and PRC to be available to lesser powers. I expect that a road-mobile ASBM weapon with good sensor/ML capability, range ~1500km, will be available for ~$10M. Given a rough (~10km accuracy) fix on a carrier, this missile will be able to arrive in that area and then use ML/sensors for final targeting. There is no easy defense against such weapons. Cruise missiles which pose a similar threat will also be exported. This will force the US to be much more conservative in the use of its carriers, not just against Russia and PRC, but against smaller countries as well.

Given 1. and 2. my recommendation is to decrease the number of US carriers and divert the funds into smaller missile ships, subs, drones, etc. Technological trends simply do not favor carriers as a weapon platform.

Basic missile technology is old, well-understood, and already inexpensive (compared, e.g., to the cost of fighter jets). ML/sensor capability is evolving rapidly and will be enormously better in 10y. Imagine a Mach 10 robot kamikaze with no problem locating a carrier from 10km distance (on a clear day there are no countermeasures against visual targeting using the equivalent of a cheap iphone camera — i.e., robot pilot looks down at the ocean to find carrier), and capable of maneuver. Despite BS claims over the years (and over $100B spent by the US), anti-missile technology is not effective, particularly against fast-moving ballistic missiles.

One only has to localize the carrier to within few x 10km for initial launch, letting the smart final targeting do the rest. The initial targeting location can be obtained through many methods, including aircraft/drone probes, targeting overflight by another kind of missile, LEO micro-satellites, or even (surreptitious) cooperation from Russia/PRC (or a commercial vendor!) via their satellite network.

2017-06-28 on infoproc

If it’s not true already it will certainly be true soon (10 years max?) that putting $billions into a large, easily visible target that has (due to basic asymmetries of physics) very little chance of defending itself against cheap mach 10 ballistic missiles (or even slower cruise missiles) is really, really stupid. Such missile weapons won’t be wielded only by major powers like Russia and China. They are cheap and will be exported to the same 3rd rate powers the carriers are supposed to intimidate (e.g., Iran). Missiles are cheap, ML is cheap, sensors are cheap, hard-to-detect drones that can acquire a rough targeting fix are cheap, LEO micro-satellites are cheap, etc. etc.

But, say the pro-carrier crowd, we’ll just tell the bad guys they are *not allowed* to take shots at our aircraft carriers, even as they are being bombed from those same carriers. (You wouldn’t dare hit our vulnerable carrier, because it cost us so much money and we have so many men on board! We’ll be really mad! Meanwhile, let us bomb your cities…)

If you can dictate like that your enemies, just tell them to report to the disintegration chamber — no need for any local dust up.

2017-06-27 on infoproc

Yes, but the US military maintains that ASBM can’t hit carriers! Seems crazy to me, but I think USN procurement priorities (and groupthink) prevent analysts from giving realistic analysis.

2017-06-27 on infoproc

Once the carrier is located the ballistic missile is launched toward that coordinate. Early guidance could be via GPS or simply inertial (~km inaccuracy does not matter!). Once the missile gets close (within few x 10km) it needs some way to target the carrier which will have moved as far as ~10km during the 10 min flight time. That’s where the ML comes in, utilizing optical + IR + radar imaging data. On a clear day ML + optical should make locating the carrier during the final ~10s of flight trivial (think of a human pilot locating the carrier visually on approach). The final targeting system is only on briefly so I don’t think battery power is a big issue.

2017-06-27 on infoproc

This works both ways. If the US attacks a PRC target from a carrier they have to be ready to risk the end of mankind. This means A2AD fait accompli.

It’s not a viable game theory proposition to take an aggressive action using a vulnerable platform (carrier) whose main protection is the *threat* of full nuclear escalation. What it really means is that that platform cannot be used (against PRC, Russia) for any purpose short of all out war (= end of mankind).

US should be diverting budget from carriers to subs and smaller surface ships.

2017-06-26 on infoproc

I think China only has, for now at least, regional projection capabilities. They are mainly planning for a conflict in Asia. But in a (conventional) conflict in Asia I don’t think either the UK or France will be a major player.

2017-06-26 on infoproc

Maybe, but that’s a big escalation, creating space debris and all kinds of nasty consequences that will affect nations not part of the conflict.

Strangely, I don’t see reference to the “destroy all satellites” step in scenario planning for this conflict. Most US planning (at least in the open literature) doesn’t admit that ASBM will be effective. If they thought it was effective they would shift money to subs and away from carriers.

Even without satellites, other systems such as drones, subs, surveillance planes, etc. can be used to roughly localize a carrier.

PS Pretty hard to stop the enemy from re-deploying LEO micro-satellites…

2017-06-20 on infoproc

Hard to imagine that the conflict would stay conventional if it reaches the level of intensity imagined in the video.

2017-06-19 on infoproc

Yes, there is variability in the maturation process itself that reduces early correlation.

Also, measurements of intelligence are noisier for kids than for adults.

2017-06-19 on infoproc

Oops, should have put this reply here instead of above.

I recently looked into the situation with NATO militaries. I had sort of tuned out on that topic after the cold war ended. It’s shocking that essentially *none* of the the (non-US) NATO countries, not even UK or France, has any serious power-projection capability left (other than nuclear weapons).

In a US-PRC confrontation in Asia, the other major player would be Japan, followed possibly by Russia, Australia, India. I doubt NATO countries would play a role.

2017-06-19 on infoproc

I recently looked into the situation with NATO militaries. I had sort of tuned out on that topic after the cold war ended. It’s shocking that essentially *none* of the the (non-US) NATO countries, not even UK or France, has any serious power-projection capability left (other than nuclear weapons).

2017-06-18 on infoproc

I am not an expert on this topic, so take what I say with low confidence!

IIRC, Ravens has had a huge Flynn effect. So it’s possible that Ravens scores from two different populations lead to misleading comparisons, esp. if those populations have experienced different levels of exposure to the kind of abstract, algorithmic thinking found in modernity.

2017-06-16 on infoproc

You might be amused that I attended a small dinner in Pasadena recently at which Munger was not the only billionaire and I was not the only theoretical physicist 🙂

2017-06-16 on infoproc

If you look up bios of prominent CCP leaders who are, say, 40-50 years old, do they typically have very superior credentials, like Tsinghua engineering degrees?

2017-06-16 on infoproc

I think careful meta-analysis shows that shared environment is not exactly zero. It’s just smaller than a lot of environmentalists had hoped for.

See Tim Bates article — Turkheimer cites it in one of his rebuttals in this recent debate.…

2017-06-16 on infoproc

When you are a child your parents can partially control your environment, and the environment somewhat controls you. An adult has much more freedom to follow their impulses. A bookish kid will eventually end up at the library or bookstore once *they* control their own activities — even if the parents keep few books at home. This is to me at least a plausible explanation for why heritability of cognitive ability increases with age, hitting an asymptote in adulthood.

2017-06-16 on infoproc

Can you link to some references describing this process in more detail?

2017-06-16 on infoproc

Yes, the key is per capita GDP. This is not understood by immigration advocates. The quality of human capital and its ability to assimilate to the existing national culture are the most important factors — not just counting heads.

Barring a disaster, it is hard to imagine China’s GDP per capita will not continue to increase. If they ever get close to Taiwan’s GDP per capita and overall level of development they will eclipse any other country or political entity in comprehensive national power.

2017-06-16 on infoproc

It depends on what happens to the EU, whether western powers other than the US want to play the role of global hegemon, etc.

The situation today is that the US is focused on preserving its primacy, wants to deny Russia and China any local sphere of influence, etc., whereas Europe has little appetite for any of it. They can barely allocate enough resources for their own defense.

Europe and the US have their own demographic problems to deal with in the coming decades. An aging population may turn out to be less challenging than the consequences of mass immigration (note population trends in Africa, so close to Europe).

If China behaved as an aggressive hegemon like the US or former USSR, it would probably elicit a collective back reaction from the West. But I think its first step is simply to consolidate influence over Asia.

2017-06-16 on infoproc

In my estimate I reduced effective total GDP by 25% to account for the relative demographic burden — i.e., reducing a PRC/US 4x population ratio to 3x. That’s a pretty large correction given that the actual difference in elder population between the US and China in, say, 2050, might be 10% of total population (e.g., 30% vs 20% over 65). It’s like allocating an extra 15% of GDP to deal with elderly care.

As far as cultural differences I think elderly care will be less of a problem there than here. They are more comfortable with collectivization (communal solutions like group living; rationing health care) and most of their elderly will not have high expectations for resource consumption (“golden years”) after retirement. I think this is completely misunderstood by westerners.

2017-06-14 on infoproc

IPCC 1.5-4.5 range has not changed in ~30 years. This “best estimate” presumably includes all information available to the community. Read the APS 570+ page transcript of discussion between physicists and top climate modelers for more color.

2017-06-14 on infoproc

US is looking at huge estimated expenses to ensure safety/reliability of our stockpile. Everyone has these problems.

Because of test ban treaty the aging of nuclear weapons can only be studied indirectly through simulations, complex materials modeling, etc. Dangerous!

2017-06-13 on infoproc

I’m OK with UBI. But we have to watch out for unintended consequences. Work has positive value beyond just the income it generates. People need a purpose and some structure in their lives…

2017-06-07 on infoproc

They may not have seen VC dimension in a course, but the idea should not be totally alien to them. When I wrote “whether the available data, given its own uncertainties, can constrain this high dimensional parameter space sufficiently” I almost used “VC dimension” but I thought many readers would not recognize the term!

In reading through the transcript I was a bit disappointed that Koonin didn’t push this point. But he was only gently feeling his way around their world to learn how they think, and perhaps felt reluctant to confront them with hard stuff.

2017-06-06 on infoproc

Yes. Several actually have a background in physics, and would know a Vapnik when they see it.

2017-06-06 on infoproc


Your characterization of the problem is spot on. I read through most of the 570+ pages of the APS workshop transcript. I wonder what response your comment would elicit from the 6 climate scientists who presented at the meeting…

2017-06-05 on infoproc

I specifically noted that the page I linked to has a certain perspective. But I think it does capture the tone of the workshop.

AFAIK Muller only looked at the temperature record, but not dynamics such as the climate models or CO2 sensitivity.

Someone like Koonin can easily get up to speed on the relevant science and lend rigor to the discussion, just as Muller did.

2017-06-05 on infoproc

The link below goes to a description of the 2014 APS workshop, chaired by Koonin, which featured 6 climate scientists, on both ends of the spectrum. The article uses terms like “alarmist” so it presumably takes a certain perspective. But at least it gives the flavor of the 1 day event and links to other resources like the 500+ page transcript.

The workshop seems like a valuable meeting. Why not have something similar that runs for 6 months and produces formal Red and Blue reports with rebuttals, etc.? Koonin obviously believes that his workshop isn’t the last word since he is now calling for a larger exercise.

This is a trillion dollar issue so I don’t see why it should be left to specialists if there is any chance that groupthink or concern for research funding or politics are affecting the conclusions that reach the public. All communities are vulnerable to these forces.…

2017-06-05 on infoproc

IIUC the increase in solar heating causes CO2 release from the oceans. So there are (at least) two things going on during the interglacial warming. Do you have a reference for how the feedback is estimated?

I know even less about this topic than about climate modeling but a cursory look at the links that appeared up-thread suggests that there are still mysteries to be resolved in climate history on geological timescales ………

2017-06-04 on infoproc

Thanks also for your comment and the links. Can I ask you the same questions about a Red Team exercise that I asked Matt?

Do you think Koonin’s Red Team idea is a good one? How would you organize it?

PS Your post at the link below, in quoting from those published papers, illustrates very well how the coarse graining of a specific microsystem can have an *under-constrained* parameter space description with large families of satisfactory values (at the micro level) of the parameters, but with different impacts on the global output of interest (e.g., temperature). Anyone experienced with modeling knows that bodes ill for reliable predictions!…

2017-06-04 on infoproc

Thanks for your comments!

Do you think Koonin’s Red Team idea is a good one? How would you organize it?

It sounds like you think the consensus prediction of (say) 2-3 degrees C per doubling is well-supported both by modeling and geological data. Does that mean the main benefit of a Red Team exercise would be to convince skeptics? Are there weak aspects of the scientific consensus that need to be examined more critically?

2017-06-04 on infoproc

“bit of genetic tweaking”: I’m pretty specific about the assumptions that lead to this conclusion — that the genetic architecture for cognitive ability is polygenic and approximately additive, like the many other traits in plants and animals that have been deliberately shifted many SDs. I’m not so confident in the hypothesis that I’d bet trillions of GDP on it, though. At this point it’s all in the domain of speculation.

In climate models there are many heuristic assumptions (coarse graining some complex physics into a tuning parameter). The uncertainty quantification of these models is very poor. I believe that if it were done properly the confidence intervals would be very broad.

Are you claiming that the geological record tells us what the temperature rise would be due to a doubling of CO2? If so, please elaborate.

2017-05-28 on infoproc

Thx! I’ve used this in the past but it seems to have improved a lot. I’ve installed it … let’s see if it tells me more than StatCounter + Blogger logs.

2017-05-27 on infoproc

Most of our readers (and certainly the most discerning) in the forward light cone will undoubtedly be AIs 🙂

2017-05-27 on infoproc

Blogger has its own tools. I also use StatCounter for cookies. You are right that I might be counting bots among my readers :^)

2017-05-25 on infoproc

I got Marina vN’s book out of the library but didn’t find it very interesting.

Believe it or not there is a monograph by vN’s brother which I managed to get via our library. It also wasn’t so informative. I’ve found the comments of his peers more insightful than those of his family, who were not scientific experts.…

2017-05-25 on infoproc

Great to hear from you! We will see much more work along these lines in the future……

2017-05-23 on infoproc

SNPs only tag causal variants. Correlation between SNP and specific variant can vary between ethnicities. Predictor trained on one group could work poorly on another.

ADDED TO ANSWER QUESTIONS BELOW: A specific SNP hit might be causal, or it might just be in LD (correlated) to some other variation that is the actual molecular cause. LD patterns vary by ethnicity. So the predictor SNPs might not work so well in more distantly related ethnic groups. This is all true even when the genetic architecture is strictly linear. In the linear case the effect size of the *actual causal variant* would be the same across groups. But there is no guarantee that the algorithm that produced the predictor in the original training set found only actual causal predictor SNPs. The predictor might be using correlations that are specific to the original training group.

I cannot elaborate more without charging you money 😎

2017-05-23 on infoproc

Not sure what Taleb points you are referring to. Current GWAS practices are ANTI p-hacking:

Threshold for genome-wide significance (definition of a “SNP hit”) corrects for multiple testing: requires p < 5E-08 NOT p < 0.05 ! SNP hits routinely replicate in independent samples. Standard ML methods using training set + validation on independent data set show real predictive power from SNP hits.

2017-05-23 on infoproc

Give Carl a break. His paper is run by the ideologues I mentioned…

2017-05-18 on infoproc

Sure, it’s possible that Comey was giving a very narrow answer. But it doesn’t seem to me that he’s parsing the question that carefully (watch the video). He seems to be saying that the FBI has not, in his experience, been forced to stop an investigation for political reasons.

Imagine him saying “that would be a very big deal. It’s not happened in my experience” all the while thinking that he had recently made notes about Trump trying it on him. Seem plausible to you?

2017-05-18 on infoproc

I think my headline is a faithful description of Comey’s testimony.

2017-05-17 on infoproc

Those sentences are a bit misleading. I don’t think McCulloch would include vN as part of their core group that worked on “cybernetics” and neuroscience, although they were in communication. There was some vN-Wiener rivalry and then later a split between L-M-P and Wiener himself. I think M meant that in the L-M-P group Pitts was the genius.

2017-05-14 on infoproc

Thanks! The wordplay I referenced is 白痴 (idiot).

2017-05-09 on infoproc

OK, so perhaps I should have written ~20 years. But we are getting close 🙂

2017-05-09 on infoproc

I’m a big fan of Sesardic and his muse, Professor Carmen de Macedo 😉

2017-05-04 on infoproc

Thx. James Lee is a collaborator of mine…

2017-04-28 on infoproc

The most off-scale people I know tend to be in math and theoretical physics.

2017-04-28 on infoproc

If someone in your household is a Prime member, then so are you. I get it through my wife.

2017-04-27 on infoproc

Believe it or not, my YouTube feed!

AI at work … Lord knows I’ve spent enough time training it 😎

2017-04-27 on infoproc

It varies. But a fair number who have hit the ultra-high net worth level (e.g., $30+ M) try to be “well hedged” against various tail risk scenarios. US dollar collapse is one of the biggies. Stockpiling bullets and seeds is a bit more rare.

2017-04-26 on infoproc

That’s a great question. There are definitely still geniuses around but science has become so vast and specialized now that I can’t think of anyone whose influence is so broad as vN’s was.

2017-04-20 on infoproc

Thanks for that reference. I wouldn’t say the hit percentage is all that high for recommendations, but since I see ~20 at a time on the page often there are a couple that are very high quality (i.e., very relevant to what I am interested in). I also notice some spooky image recognition capability — like if I watch a music video featuring large breasted women or corsets it then recommends videos with similar content to me for a while. This presumably comes from their automated image labeling work. It’s a bit eerie.

2017-04-17 on infoproc

I doubt that the effect is that large. The DeCode study suggests something like an IQ point per decade or generation.…

2017-04-16 on infoproc

I suppose that earlier post has to be read in the context of someone (like me, or apparently LeCun 20+ years ago) who had the prior that local minima would be a big problem… If you don’t suffer from that prior it might appear incoherent 😎

2017-04-16 on infoproc

I’m not an expert on any of this, but my last (physics) postdoc just made this transition. I can put you in touch — he probably has a good list of the resources he found useful. Send me an email.

I think anything Nielsen writes is very clear, so that would be a place to start.

2017-04-16 on infoproc

The story LeCun tells is that people abandoned backprop at first because they thought it wasn’t working *specifically* due to *local minima* (not because of saddle points). That turns out not to be true.

In the 80s and 90s people in physics and neural nets were discouraged about optimization due to fear of local minima (see simulated annealing, etc.). But in certain systems (like large NNs) it turns out not to be a big problem.

I’m not sure what you find incoherent about the story.

2017-04-16 on infoproc

Sry — posting from phone.

U get caught in a local minimum but can still escape a saddle point (gradient vanishes at a point but discrete numerics… slowdown but not trapped). When number of eigenvals gets large unlikely to have all positive.

I think Bengio or Lecun even wrote a paper on this.

######## Now off phone

Here it is:….

1. exponentially more saddle points than local minima in high dimensions
2. saddle points cause slowdowns
3. various methods that can readily escape from saddle point region but would not work at a local minimum

2017-04-14 on infoproc

According to NLSY / Pew Trust results I linked to, someone born in the bottom income quintile has an ~80% chance of moving to a higher quintile if they have at least average IQ. This is true for both blacks and whites and the probability seems to be mainly driven by cognitive ability, not race.…

2017-04-14 on infoproc

I thought he did an admirable job of capturing all sides of the issue in the podcast. Esp. the last 10 min or so.

2017-04-11 on infoproc

The problem is with the decision makers, who are unlikely to be statistically sophisticated. They will also be *combining* predictive results across multiple traits in order to make their decision. I think that a *good* rank order predictor (for which the score order correlates well with actual rank order) is something an average person can process, even if they need to combine prediction results across multiple traits. But something that only accounts for a small amount of total variance is hard for ordinary people to think about, esp. across multiple traits.

It’s not a question about whether there is positive expected return. The question is whether the decision makers have a good intuitive feel for the tool. Perhaps this can be improved through good exposition, but I have my doubts (even for the clinic docs, let alone the parents).

Warning against elevated risk (in this case, from large negative fluctuations) has a long history in medicine, so that should be the least controversial. A noisy statistical predictor that only accounts for a small fraction of total variance would (my guess) generate controversy and negative reaction.

2017-04-11 on infoproc

Almost 1M couples use IVF each year, worldwide, typically for fertility reasons. Most of them have to make an “embryo choice” decision (which one to implant). Are you against giving them the best information to help with their decision? Are you against testing for Down Syndrome? What about other, equally serious, health conditions?

2017-04-11 on infoproc

What I mean by rank ordering capability is (roughly; this can be made more precise): if you have, say, 10 embryos the one with the highest predictor score is very likely to be near the top in actual phenotype, and the one with lowest predictor score is very likely to be near the bottom. Current predictors for Eduyears are not good enough for this, but height predictors soon will be. Essentially, the rank ordering in predictor scores correlates reasonably well with actual rank order. Before this is true I would not be comfortable with using the predictors for more than warning about negative outlier risk of certain embryos.

2017-04-10 on infoproc

Rank ordering embryos for cognitive ability is some ways away but possible soon for height.

Eduyears is probably a combination of things like conscientiousness and cognitive ability. Within a couple of years at most I would guess that a combination of more data and better analysis might get this to ~0.25 of total variance, which means the predictor would correlate ~0.5 with the real phenotype. This is not good enough for rank ordering but probably good enough to warn against large negative fluctuations. (Sort of like detecting Down Syndrome today, but in this case depressed capability due to many negative variants.)

What is really needed is a big cohort (e.g., ~500k or 1M individuals) with decent cognitive measurements. This would cost tens of millions of dollars (cf UK Biobank), but it would be well worth it just in terms of what we would learn about neuroscience, the brain, etc., not to mention the IVF applications.

2017-04-09 on infoproc

Thanks! I think I’ve fixed the link.

2017-04-07 on infoproc

You read that the wrong way. The point is younger lefties might not be that into Chomsky, compared to older ones.

2017-04-07 on infoproc

You’re probably the wrong demographic. Talk to an older lefty — perhaps someone on the faculty 🙂

2017-04-07 on infoproc

Yes, see the earlier posts linked at the bottom. But I like to quote Chomsky because it causes cognitive dissonance in certain brains 😉

2017-04-05 on infoproc

If on average M and F brains work about equally well*, and M brains are 10% bigger (all around), then absolute volume may not be the right variable. Instead, one could look at typical *fractional* variability in volume.

* Not everyone agrees with this.

2017-04-05 on infoproc

The effect is definitely much smaller if you normalize that way. I did a couple of example cases and still found slight effects, though. (I did this in my head so I might have made mistakes.)

It’s not clear biologically which is more important — the absolute fluctuation in volume or the percentage (dimensionless) fluctuation.

2017-04-05 on infoproc

Recent expansion of access might be OK if enough (e.g., database system-level) safeguards are in place.

However, in the excerpt I added from your Observer piece, you can see that agencies are subject to pressure from, e.g., the WH to bend or break the rules. It’s inherently dangerous: imagine some Snowden-level guy who has broad access and is being pressured by his high level (political appointee) superior to pull up some desired info.

The LOVEINT examples (+ Snowden himself) suggest that system-level protections even at NSA do not prevent lower level individuals from conducting invasive searches on raw data. Instead of tracking/stalking a love interest, imagine that politically motivated requests are coming down from high up…

2017-04-05 on infoproc

The problem: based on their public comments, I doubt whether the media and even intel committee members understand how bulk collection, mass storage of intercepts, cellphone malware, etc. work. So basic competency may be lacking.

2017-04-05 on infoproc

The excerpt is from David’s linked article, written by:
John Schindler is a security expert and former National Security Agency analyst and counterintelligence officer. A specialist in espionage and terrorism, he’s also been a Navy officer and a War College professor. He’s published four books and is on Twitter at @20committee.

2017-04-04 on infoproc

Excerpt: (note last 2 sentences below!)

Not to put too fine a point on it, but Susan Rice is a deeply unpopular figure with our Intelligence Community. Her abrasive personality and overall incompetence grated on the IC. Her habitually coarse language was inflicted on senior intelligence officials more than once, while nobody outside Obama’s inner circle considered Rice even marginally competent at her job. Simply put, she was the worst National Security Adviser in American history—at least until Mike Flynn’s dismally failed three-week tenure.

In addition, Rice didn’t like to play by the rules, including the top-secret ones. On multiple occasions, she asked the NSA to do things they regarded as unethical and perhaps illegal. When she was turned down—the NSA fears breaking laws for any White House, since they know they will be left holding the bag in the end—Rice kept pushing.

As a longtime NSA official who experienced Rice’s wrath more than once told me, “We tried to tell her to pound sand on some things, but it wasn’t allowed—we were always overruled.” On multiple occasions, Rice got top Agency leadership to approve things which NSA personnel on the front end of the spy business refused. This means there may be something Congress and the FBI need to investigate here.

2017-03-30 on infoproc

[Wigner] Until 1925, most great physicists, including Einstein and Planck, had doubted that man could truly grasp the deepest implications of quantum theory. They really felt that man might be too stupid to properly describe quantum phenomena. … the men at the weekly colloquium in Berlin wondered “Is the human mind gifted enough to extend physics into the microscopic domain …?” Many of those great men doubted that it could.…

Beyond Human Science: [This Ted Chiang short story envisions a future in which science has become the province of genetically enhanced “metahumans” — leaving non-enhanced humans to gape from the sidelines.]

2017-03-30 on infoproc

The truth is even more brutal than you can imagine. Are there no concepts that are challenging to the top 1%? What if they had to be invented by someone in the top 0.01%?…

Feynman notes that only a few dozen students (out of a Caltech cohort of perhaps 200) were able to fully appreciate the material in his lectures as delivered.

“the powers of instruction are of very little efficacy except in those happy circumstances in which they are practically superfluous.”

2017-03-29 on infoproc

I think Joe is a smart guy and I really enjoy his podcast. But I don’t think mathematical concepts are his strength.

2017-03-29 on infoproc

> A random person off the street can understand the concept gauge symmetry given enough time to absorb it and its pre-requisites. < I'm just not sure this is true:….

Suppose we define “understand” in some rigorous way. A person only “understands” the material if they can pass a test involving equations, calculations etc. Would you say it’s true for someone with an IQ of 100? 90? 80?

Note what I wrote above in the comment thread:
> the concept of gauge symmetry is quite difficult for most people to understand. It’s actually not understood well by *most physicists* ! <

2017-03-29 on infoproc

It’s not that the clone would pick it up all in one shot through casual conversation (although he would probably pick up more than Joe). It’s that if the clone spent, say, an hour or two with the real LK in front of a blackboard, he could make progress in understanding that Joe might not be able to make at all, even given unlimited time.

2017-03-28 on infoproc

The clone would crush Joe.

2017-03-28 on infoproc

In case you are interested, there have been many posts on this blog concerning exactly this question — how much raw democracy vs technocratic leadership is optimal?……

Whether or not the masses can overcome elite policy preferences even in the US system is debated among political scientists:…

2017-03-28 on infoproc

I think it’s possible to do a better job than Krauss did but I still think the concept of gauge symmetry is quite difficult for most people to understand. It’s actually not understood well by *most physicists* !

2017-03-28 on infoproc

I certainly share your view that popular books on theoretical physics are doomed to fail. That’s why I’ve not written one. What is more depressing is that the general population might also be doomed to fail in making democratic policy decisions on complex topics like genomics (or computer hacking to influence elections).…

Yes, it is an exaggeration to suggest that component concepts leading to gauge symmetry are as far beyond average human minds as ordinary human concepts might be to a dog. But it gets the point across.

2017-03-28 on infoproc

In the cartoon the dog is not just having trouble with the overall idea, but with each of the smaller components from which it is constructed.

2017-03-24 on infoproc

I was joking — I didn’t really challenge him 🙂

2017-03-24 on infoproc

It seems that lately Wikileaks is the only viable option for whistleblowing / leaking.

The stuff they post gets at least some attention and they have yet to have any challenges to its authenticity.

2017-03-19 on infoproc

As I keep explaining to you, the “robot” can live in a virtual environment. It doesn’t have to be a mechanical robot. See, for example, the paper that Seth (STS) linked to in this comment thread.

You can verify the amazing performance of deep learning by using voice recognition on your Android phone, or Google Translate, etc. All of these have improved drastically in the last years.

2017-03-02 on infoproc

“… no journal is interested in failed research. No research grant is interested in dis[-]proving others findings.”

In fundamental physics replication is treated with respect. If one group claims a result of fundamental importance, then other groups will rush to try to replicate, and funding agencies and journal editors recognize the importance of this activity.

Other fields need to understand, collectively, the central importance of replication to the functioning of their field as an actual science. A collective endeavor which does not produce reliable, enduring truth is not successfully practicing the scientific method.

Part of the problem is a lack of statistical (and epistemological) understanding. Negative results are themselves inputs into a Bayesian confidence level in the original claims. Since our goal is to obtain high confidence conclusions (i.e., to know the difference between claims in which we have high vs low confidence), even negative results are valuable if the replication attempt is performed well.

2017-03-01 on infoproc

I’d like to add that the basic questions addressed by this work are quite interesting:

1) Is there a compressed measure (“c”) of group effectiveness analogous to g for an individual?

2) How strongly dependent is c on the average or max g of individuals in the group?

3) Are there best practices (turn taking? Robert’s Rules of Order? ensuring diversity of background?) that enhance c at fixed g distribution within the group?

These are obviously tricky questions to get at and the answers may depend on the types of tasks assigned to the groups. It seems to me that the Bates replication tried to follow the original work closely, but of course faithful replication is quite difficult.

I can attest that the original claims of the 2010 work have already made their way into real world situations — I’ve been told during “leadership training” exercises involving top-level university administrators that group effectiveness is enhanced by diversity and specifically by the presence of women in the group. This may (or may not) be true, but certainly the people making the claims were not carefully examining the underlying science at all!

2017-02-28 on infoproc

Note to commenters: we’re lucky to have authors of both papers under here in the thread. I’m not going to allow hostile or ad hominem attacks on either of them. If you have actual questions about the research post them here (read the papers first!), but be civil or I won’t approve your comment.

I just removed 2 comments that didn’t add much to the conversation, but were hostile to Chris.

2017-02-24 on infoproc

Sadly, I agree with your prediction.

2017-02-22 on infoproc

Translation was the first application, and the first demonstration of automated mapping of thought vectors, but the next (proximate) application is to reading comprehension (document processing). I expect that within a few years, perhaps five at most, we’ll be able to query Google in natural language and receive responses as if it had read and (at least partially) understood large bodies of text like Wikipedia, standard textbooks, etc. Note that Google is sitting on a huge corpus of scanned books from university libraries. They can’t give us access to those books due to copyright, but they can use the data for applications like this. Note also that (reportedly) TPUs running deep learning are 20% of ALL of Google’s compute. Their basic search function switched to using deep learning already some years ago. The competition (except maybe Baidu?) has been left in the dust.…

Building Large Machine Reading-Comprehension Datasets using Paragraph Vectors

Radu Soricut, Nan Ding….
(Submitted on 13 Dec 2016)
We present a dual contribution to the task of machine reading-comprehension: a technique for creating large-sized machine-comprehension (MC) datasets using paragraph-vector models; and a novel, hybrid neural-network architecture that combines the representation power of recurrent neural networks with the discriminative power of fully-connected multi-layered networks. We use the MC-dataset generation technique to build a dataset of around 2 million examples, for which we empirically determine the high-ceiling of human performance (around 91% accuracy), as well as the performance of a variety of computer models. Among all the models we have experimented with, our hybrid neural-network architecture achieves the highest performance (83.2% accuracy). The remaining gap to the human-performance ceiling provides enough room for future model improvements.

2017-02-20 on infoproc

It’s happening!!! 😉

2017-02-20 on infoproc

“Neural nets used in language translation have mapped out an abstract ~1000 dimensional space which coincides with the space of “primitive concepts” used in human thought and language. It appears that rapid advances in the ability to read human generated text (e.g., Wikipedia) with comprehension will follow in the coming decade. It seems possible that AGI — Artificial General Intelligence (analogous to a human intelligence, with a theory of the world, general knowledge about objects in the universe, etc.) — will emerge within our lifetimes.”…

2017-02-20 on infoproc

> untapped potential is huge < That's my sense. If the neural encoders/decoders are good enough to learn reasonably good mappings between English sentences and thought vectors (perfection is not required), which seems to be the case, there are lots of things which can be built now which were not imaginable even a few years ago ...

2017-02-12 on infoproc

I have experienced this first hand… 🙁

2017-02-12 on infoproc

Fixed bases are vulnerable for both sides, although the new SCS bases are probably being constructed from the beginning in a hardened way with precision strike capability in mind (see link). This is not true of existing US bases.

Escalation to anti-satellite ops is pretty serious as the US military depends heavily on them as well. Also lots of collateral damage to commercial satellites due to debris, etc. To hit fixed bases using just inertial and terminal guidance is probably doable (satellites not necessary). Tracking a carrier group probably relies heavily on satellites.

2017-02-12 on infoproc

“I voted for Trump because I thought that a war monger like Clinton would give the Deep State a free reign, unlike Obama who tried to keep its various schemes in check. Unfortunately I was mistaken, and Trump’s administration is likely to expand our wars in Africa, the Middle East, Central Asia and Eastern Europe.”

Let’s hope you are wrong about this. Where do you think Bannon is on the war-monger axis?

2017-02-12 on infoproc

“…carriers [won’t] play an important role in hypothetical wars with Russia and China.”

$10+ billion for a new carrier is a lot for something just to impress Solomon Islanders and Iran. (Why do we need 11 if they aren’t useful against PRC or Russia?)

The few US strategists who publicly align with your quote above are considered heretics. I recently questioned a former commander of the US Pacific Fleet on this topic and he (even privately) would not agree to your quote. DANGER! DANGER!

In Australia, which has to think these issues through very carefully, there is a lively debate over whether they should spend their money on subs or surface ships…

2017-02-12 on infoproc

I think that even with control over SCS it would be difficult for PRC to keep oil flowing.

But the SCS presence raises the cost of interdiction (e.g., in the Malacca Strait) for the US.

Somewhere I even recall seeing claims that the Chinese are building deep sonar sensors in the region to hamper US submarine operations. The ability to maintain these sensors depends on having some level of regional dominance, etc.

The other advantage of SCS dominance that I didn’t mention is the ability to protect nuclear subs coming in and out of Hainan island. Obviously the region is important for more than one reason.

2017-02-12 on infoproc

No one really knows what the escalation criteria are on either side.

In any case, it’s dangerous for the US to field carriers as primary forward platforms if they are indeed highly vulnerable and only protected by threat of (nuclear) escalation. Since the US and PRC are now in a MAD nuclear balance (both sides can annihilate the other, and neither plausibly has first strike capability), it also isn’t clear that PRC would believe that taking out a carrier would lead to a nuclear response. If a US carrier group was attacking one of their artificial islands, they might feel justified in knocking it out with some DF21s.

Xi Jinping (i.e., in building the islands) seems to behave as if he knows he has little to fear from the US fleet. But I don’t know whether most US Navy planners understand this.

The danger is compounded by the lack of understanding of “experts” (both military and foreign policy) about the basic realities.

2017-02-09 on infoproc

Barring disaster, peak cognition is in our future due to genetic and AI developments. But there is a very strong “undertow” we are fighting against…

2017-02-06 on infoproc

I don’t anticipate neural nets replacing humans wholesale in financial markets in the near future. (It’s not a low-skill job 😉 For some narrow applications or specialized funds, perhaps.

2017-02-04 on infoproc

Steve is fine 🙂

I like Lydon and regularly listen to his podcast. Nader and I were on the same ticket for Harvard Overseer last year, but alas none of us were elected!

2017-02-04 on infoproc

The mobile robot can now be created within a rich virtual world that obeys the usual laws of physics but runs at much faster speed.

2017-02-01 on infoproc

I regret the possibility that my blog post caused them to be removed. They had been up for some time! 🙁

2017-01-30 on infoproc

That was only in the tweet 🙂

2017-01-22 on infoproc

I didn’t mean that a kid would discover the precise formulation of correlation or variance or higher-dimension spheres or geometrical projections. But it would be easy for a far tail kid to discover/invent the intuitive idea for these things on their own. Then, when they encounter the formal definition they understand it immediately and can use it, whereas many educated adults don’t even have the intuitive concept. I think all the examples I gave above are things I had thought of on my own (again, in an intuitive rather than precise way) long before learning about them in a classroom or from a book.

2017-01-22 on infoproc

I’ve been too busy to follow this discussion the past few days, but I’d like to clarify something that was probably confusing in the original post and my subsequent comments.

One of the things I noticed in the SAT reading comprehension sections my kids were looking at is that one is NOT being asked to making subtle distinctions between nearby concepts/words. One is merely being asked to know that X (esoteric word) is a synonym for Y (common word), without having to know the subtle difference between X and Y.

So, if my kid didn’t know that “Brobdingnagian” is a synonym for “big” they might not be able to answer a multiple-choice question about a paragraph containing the sentence: “But of course the error was of Brobdingnagian proportions.” To answer the question doesn’t require knowledge of Gulliver’s Travels — I could un-befuddle my kid (allowing him or her to easily answer the question) just by saying “Brobdingnagian means big”!

So, at least this psychometric exam (the SAT) isn’t even testing fine distinctions — it just tests whether you know that X1, X2, … , XN are synonyms of a very primitive concept like BIG. What is the value of taking N larger and larger (in this sense; not the fine distinction sense)? Surely there are diminishing returns…

2017-01-19 on infoproc

I’m thinking about far tail individuals for whom many of the math concepts I listed are intuitive and will have been conceptualized already without formal training. Lubinski et al. found that SMPY qualifiers could often solve the high school level problems without having been exposed to the material via coursework or instruction. Certain brains are exploring this space already the way that other brains have grasped the distinction between opaque and enigmatic. It’s not purely an aptitude vs attainment dichotomy.

Ordinary vocabulary covers a finite conceptual space and there are diminishing returns to covering this space more and more finely with additional terminology. On the other hand (as Oppenheimer notes) there are vast conceptual spaces beyond the vocabulary covered in most psychometric tests.

My kids are going to take the SAT soon as part of the talented and gifted testing in our school system. It is quite interesting to see how they approach the math problems, reading comprehension, etc.

2017-01-18 on infoproc

The distinctions in meaning between those words, while real, are small (which is why they are grouped as synonyms). Perversely, psychometric verbal scores reward knowledge of these small distinctions more than mastery of deep new concepts.

2017-01-11 on infoproc

Much of his best work is single author stuff done without collaborators. He was clearly a genius all by himself and everyone knew it.

2017-01-10 on infoproc

It’s something you would think is widely discussed among the grandmasters who have been playing AlphaGo. At least from the comments, it appears they think that AG has deep insights, not just superior tactics. But as you say, time will tell.

2017-01-09 on infoproc

Or autonomous vehicles … Or rich virtual worlds … 😉

2016-12-31 on infoproc

Likewise to you and your family! 🙂

2016-12-29 on infoproc

Recent results are shifting my priors!

2016-12-28 on infoproc

He was not retired, and still enjoyed interacting with undergraduates. The main course he was teaching was an advanced graduate course on the physics of computation (essentially, laying the groundwork for what is now quantum computing). But he also hosted something each week (Thursday evening?) called Physics X, open only to undergrads. He would answer *any* science question asked to him by the students. As with the original Feynman Lectures the people likely to have gotten the most out of this kind of interaction would be more mature — grad students, postdocs, even professors. But even for us it was quite a treat. I suppose just seeing him think on his feet was a special experience.……

2016-12-27 on infoproc

I had totally forgotten about the existence of this photo, and many of the others in the post, until discovering them while digging around in boxes and old albums.

2016-12-18 on infoproc

Another good example — again about 2000 primitives are enough for basic communication.

What’s interesting here is that the deep translation algorithm is able to extract the primitives automatically from translated text across any pair of languages.

2016-12-14 on infoproc

The average person has a vocabulary of perhaps 10-20k words. But if you eliminate redundancy (synonyms + see below) you are probably only left with a few thousand words. With these words one could express most concepts (e.g., those required for newspaper articles). Some ideas might require concatenations of multiple words: “cougar” = “big mountain cat” , etc.

But the ~1k figure gives you some idea of how many distinct “primitives” (= “big”, “mountain”, “cat”) are found in human thinking. It’s not the number of distinct concepts, but rather the rough number of primitives out of which we build everything else.

Of course, truly deep areas of science discover / invent new concepts which are almost new primitives (fundamental, but didn’t exist before!), such as “entropy”, “quantum field”, “gauge boson”, “black hole”, “natural selection”, “convex optimization”, “spontaneous symmetry breaking”, “phase transition” etc.

2016-12-05 on infoproc

1) It’s a matter of years before we have predictors with correlation ~0.3 to 0.5 with actual IQ or some kind of adult achievement measure. This is (numerically) roughly the correlation between SAT score and college GPA or HS GPA and college GPA. That would be good enough for some IVF applications. See…

2) It’s 3 points per SD of variation of the polygenic score. If you are screening against outliers this is almost enough to be actionable.

3) This area of research deserves much wider attention, better funding, etc. It could result in a discontinuity in human history and in human evolution. Compare to AI research, which has similar potential. The resources available for AI/ML are incomparably greater than for this work.……

2016-12-01 on infoproc

Any predictions on whether this project will succeed?

2016-11-29 on infoproc

There are well-known examples of relationships between morphology (appearance) and cognition. You can spot someone who has Down’s Syndrome from appearance alone, although probably the worst aspect of the condition is depressed IQ. In the Belyaev fox breeding experiments the domesticated foxes acquired certain traits (floppy ears, spotted coats, head shape) not found in the ancestral line. This is all just pleiotropy at work, presumably.…

2016-11-28 on infoproc

It’s confusingly written — there were certainly humans in the loop, which is why there is a 15min delay.

2016-11-28 on infoproc

My guess is that Trump support was stronger among first-generation Chinese Americans.

2016-11-26 on infoproc

I think your experiences in an urban center where Westerners are likely to be found is far from typical for the rest of the country, so probably best not to extrapolate.

Even if the crime rate is the same as in the US most identified criminals (including those that appear in wanted posters) are outliers in the general population.

2016-11-26 on infoproc

AFAIK crime rates in China are very low…

2016-11-26 on infoproc

It’s certainly possible that odd looks cause a biased reaction from the criminal justice system. But I suspect the criminals in the data set and criminals in general are, on average, only slight outliers in appearance, whereas being a criminal makes you a very strong statistical outlier in behavior. So your model is not very plausible to me. Another possibility is that the criminal group has an overrepresentation from some ethnic minority who are not predisposed to criminality but are discriminated against. I’m not sure how likely this would be in PRC, which is pretty homogeneous.

In any case I think this direction of research requires much more work.

2016-11-25 on infoproc

The big question is whether they are detecting something real or they have fooled themselves. The purely descriptive observation that there is more variance in criminal faces is probably correct but again it could simply be an artifact. People have hypothesized for a long time about physical correlates of criminality or low IQ. It seems that a concerted effort could answer this question now, but I’m not sure it will happen.

See my twitter feed for people already objecting to this work.

2016-11-24 on infoproc

re: Iowa lady, I don’t completely endorse her point of view, but I understand where she is coming from. She seems like a good person, certainly not a racist, voted for Obama twice. She’s part of the NOT deplorable Trump base.

You can model *her* reaction to Dem identity politics positions on immigration (amnesty? open borders? social benefits for illegals?), BLM (“systemic racism”? ALM = hate speech?), etc. How are they going to win her back while keeping the non-white part of their coalition engaged? I can certainly see it if Trump really screws up. But as Bannon said, if their team performs reasonably while Dems stick to their current positions on these issues, Trumpers (= populist, anti-globalist, anti-PC) can rule for many election cycles…

I think Bill Clinton understands all of this very well. But HRC and her team, as well as the media, the Left, coastal elites, etc. are just now waking up to the problem. In the end demographics may eventually deliver the prize to them, no matter what Trump does now about immigration. But that future America, I fear, will be deeply polarized along racial lines and highly dysfunctional.

2016-11-22 on infoproc

What do you mean by “is the Problem”? Do you mean a thing which has strong emotional appeal to wcw (working class white) voters and played a big role in their defection from D to Trump? The big issues are economic difficulties of wcws + immigration + identity politics, and they are all intertwined.

The reason I posted the Times interview of the lady from Iowa was to illustrate this.

I think it will be hard for D’s to please the non-white part of their coalition while reassuring wcws on these issues.

Alternatively, you might mean “is the Problem” to be something about justice, morality, etc. But that is much less relevant to electoral politics.

2016-11-22 on infoproc

Thanks for the comment and links. I’m familiar with MPC, TRS, etc. and even the pleasureman / Spencer purity feud. Spencer has become the public face and is pushing his movement into the real world. In some parallel universe future (hopefully not our future) Spencer is the 21st century American Hitler. The other Alt-Right factions are even more hardcore than Spencer, and even further from Trump and Bannon.

The purpose of my blog post was not to fully explore the Alt-Right underworld, but to check the lying press on whether there is really any link to Trump / Bannon.

2016-11-21 on infoproc

IIRC this guy was a humanities grad student at Duke not so long ago. They are a movement of ideas, not violence.

2016-11-21 on infoproc

Both parties are going to change, I think.

But you could take a minimalist perspective by just saying that working class whites, who went with Bill Clinton and Obama in the past, now are with Trump. Perhaps Dems can win them back, but there is a tension between identity politics pandering to black and hispanic voters, and reclaiming “silent majority” working and middle class whites.

2016-11-21 on infoproc

The term alt-right is up for grabs right now and people use the term in different ways.

Bannon in some of his interviews clearly thinks it means (literally) an *alternative* to GOP conservatism, and not specifically WN. I think his “platform” quote is using the term in this way.

Spencer sometimes refers to the non-WN version as alt-LIGHT (as in, watered down), IIRC. I think Spencer is being honest in the first clip. From his perspective Trump and Bannon are not sufficiently “woke” 🙂

Probably all of this media attention will cause Spencer’s narrow (WN) definition of alt-right to take hold.

I think the recent version of the GOP is nearly dead and there will be a non-WN, new version of the American right led by people like Trump and Bannon, but it won’t be called the alt-right.

2016-11-21 on infoproc

They’ll have to agree to disagree 😉

2016-11-20 on infoproc

It was great fun! We’ll have to catch up some other time…

2016-11-19 on infoproc

> High SES whites get the “white quota” admissions slots, and slots for any low SES applicants go to racial minorities. < "... slots for *any* low SES applicants ..." is a bit of an exaggeration, but I believe the available data supports the direction of your claim. It's not in anyone's interest to have the Dems become the party of non-whites + certain elites while the R's become the party of the white working class + some rich people.

2016-11-18 on infoproc

The focus of this meeting was mainly on automation, robots, information technology, etc. The claim is that these have had and will have larger effects on American jobs than foreign competition. My feeling is that both are important.

We had a long discussion of solutions which might ease the burden. I was able to slip in the idea that limiting low-skill immigration might be a good idea, given their assumption that automation would make many low-skill people unemployable!

2016-11-18 on infoproc

It’s a great idea, and what Keynes thought the world would be like by now, but there is the competitiveness issue. Developing countries won’t cut worker hours as fast as developed countries, leading to job migration.

A more fundamental issue is that the most capable or ambitious people actually *want* to work a lot, either because of the joy of creating or accomplishing things, or because of rivalrous status competition.

2016-11-12 on infoproc

1. I didn’t suggest that there is a simple or universal solution to the utility weighting problem (in fact I said the opposite). I’m not telling other people what to do — that’s why we have elections. I don’t see the point in belaboring any of this further; it’s obvious stuff for many people who have thought about moral philosophy.

2. I noticed your old Disqus comment because my admin panel shows your earlier posts when I type my reply. I wasn’t trying to ridicule you. I just wanted to point out that you might have a particular cognitive bias that makes it hard for you to comprehend the real world. Life is all about learning and *updating* our beliefs when we get new information.

2016-11-12 on infoproc

Do you also have trouble with someone prioritizing the interests of their children or family over those of others?

Once you allow for differential weighting of others’ utilities, the problem of how to assign those weights arises and there isn’t an obvious unique function that one has to arrive at. (I doubt the constant or uniform function is workable because most other people won’t adopt it. It goes against our human nature, produced by natural selection and competition for resources.)

If one shares common interests, values, tendencies toward reciprocity (the Golden Rule) with co-citizens, shouldn’t one favor their utility over that of others?

Within a nation state it seems reasonable that as a part of the *civil contract* citizens should agree to solidarity with each other over non-citizens. Our government and military should place the interests of *our* nation and citizens above those of outsiders.

I confess to finding it very strange to have to explain these ideas so explicitly.

I noticed this old Disqus comment of yours. Do you feel like your understanding of the world has improved in the last 2 years?

Andy Sebastian UVa President Recommends Series of Changes

1. Whatever the exact details of the assault were, various parties have testified visible trauma Jackie suffered following the alleged incident. Are you skeptics seriously to suggest she meticulously feigned emotional distress for 2 years in the hopes of a Rolling Stones author finding her and offering her an op-ed?

2. Does the veracity of Jackie’s individual account somehow affect whether or not action ought be undertaken to address rapes on campuses? That, if you think this one report was frivolous, you will happily cherry pick it as an excuse to refuse action on the larger issue of real rapes in universities?


2 years ago

2016-11-12 on infoproc

Strange that I didn’t notice that. I just cut and pasted the text from the Morning Joe site.

2016-11-10 on infoproc

Redistribution is necessary for social stability and I think there is a moral argument for providing a basic level of support for those at the bottom. However, one factor that has not been considered seriously enough is the longer term effects of welfare dependency vs the positive non-monetary benefits of work.

2016-11-10 on infoproc

I’m not at all confident about the future of an advanced nation that opts for open borders. I don’t see how the argument for open borders proceeds without strong assumptions about the consequences.

I do value the utility of my children, family, and nation above the utility of others, and I don’t feel that this is an unethical position to take.

2016-11-09 on infoproc

You don’t need to believe in HBD to conclude that low-skill immigration might be undesirable. Just calculate transfer cash flows (“makers vs takers”) to see that it isn’t necessarily a win to allow low-skill immigration. You can be 100% anti-hereditarian and still come to this conclusion. (Although you can’t be smart and also be 100% anti-hereditarian.) This is basically what Borjas would say privately, i.e., if he weren’t self-censoring.

“Don’t grok Trump” refers to elites being out of touch with the *electorate* and making incorrect predictions about how the election might go, etc. Case closed on that one.

2016-11-05 on infoproc

Thank you for the two you added. Fantastic! 🙂

2016-11-03 on infoproc

2-3% avg GDP growth rate seems too low for China anytime in the next decade or two because there is still a lot of catch up innovation for them to implement. I’d guess 3-5%, which still leads to increase in GDP per capita relative to Western economies.

If China is really becoming a more inwardly (as opposed to export) focused continental scale economy, they will be able to grow for some time just through internal demand and innovation.

The current debt build up is really due to post-2008 stimulus that they are afraid to turn off. If they were willing to take some short term pain they could kick the habit, but who knows what will happen.

2016-11-01 on infoproc

They are still trying to stimulate the economy. The debt levels are becoming a problem but they are afraid to stop.

2016-10-28 on infoproc

The problem is not so simple that it will be solved cleanly once and forever. The models will require constant updating as more data become available, for different groups and under different environmental conditions. Also, the technical methods will evolve, from array genotyping to sequencing; from embryo selection, perhaps to stem-cell based production of embryos, to editing.

However, if you are actually interested in medicine *and* have the ability to use computers and statistics, you might consider a medical degree. There will be many opportunities for people with both kinds of skills in our future of precision / personalized medicine.

2016-10-27 on infoproc

First, stick with the subject you find most interesting if you can.

Second, areas like genomics and genetics need people with strong computational skills. So you can stay in software engineering but try to learn about biology/IVF/statistics/genetics on the side. It would be easy for you to move into bioinformatics some day if you are still interested.

2016-10-26 on infoproc

Yeah — eventually we’ll need some international treaties to prevent too much of this, and perhaps some kind of cleanup missions …

2016-10-24 on infoproc

I love all the UFC and fighting stuff. But this interview wasn’t especially good on those topics. What’s unusual are Bilzerian’s experiences and his life insights.…

2016-10-21 on infoproc

Yes, it seems to me that ground target accuracy is harder than satellite placement.

2016-10-16 on infoproc

The first plot I show has the situation when I actually got interested in genomics, circa 2010. The second has some hand-made George Church charm 🙂

2016-10-13 on infoproc

Heritability estimates come from many sources: mono vs dizygotic twins, sibs, half-sibs, etc. As long as you know the relatedness and can compare the phenotypes. Ideally you would randomize over environments (raised apart) but that is only a small fraction of cases. The easiest thing to explain conceptually is of course identical twins raised apart, but those cases are rare.

2016-10-07 on infoproc

Yes, ideally we would control for distribution of majors, but we don’t have that data. It’s not quite as lopsided as you might think due to the large number of engineering degrees at places like MIT and Caltech. The prizes go to sciences, not engineering…

2016-10-07 on infoproc

“all winners of the Nobel Prize (physics, chemistry, medicine, economics, literature, and peace)” + Fields + Turing

But it is overall STEM biased. On the other hand, Institutes of Technology graduate a lot of engineers who are not really trained to win any of these prizes, with the possible exception of computer scientists (Turing prize). At Caltech when I was there two of the most popular majors were EE and Engineering and Applied Science (E&AS — the most popular major of all), neither of which lead to any of these prizes. I doubt more than a couple of kids graduated each year with an econ major.

2016-10-07 on infoproc

[ Note: includes Literature, Economics, Peace, as well as science prizes. ]

2016-10-05 on infoproc

Let us know how you like it! Post some video on YouTube 🙂

2016-10-04 on infoproc

@DougK and @BobS: it seems plausible the cost trend with solar will continue, and we’re talking about +30y or more long term planning when it comes to big projects. It’s possibe that grid/distribution/storage will itself evolve to take advantage of cheaper energy sources. The home/car battery idea of Musk’s might come to fruition …

2016-09-29 on infoproc

Yes, I wasn’t directly addressing your comment. I just thought you might find the data at the link interesting.

2016-09-29 on infoproc

The 1E06 multiple testing correction is a bit ad hoc. The current situation actually allows imputation of ~10 million SNPs from a typical gene array (chip) with ~1E06 probes. But I think the logic behind the 1E06 is based on the idea that there are of order 1E06 independent degrees of freedom (i.e., effectively uncorrelated single nucleotide genetic variants, or about 1 per thousand bps in the whole genome). I don’t know that this is exactly correct, but the reasoning isn’t meant to be completely precise.

In some sense it’s just terminology (“genome wide significant”) and an arbitrary threshold. But it was eventually adopted as a standard for publication by journals. The real proof in the pudding is whether SNP hits at the corrected level of significance (i.e., p < 5E-08) tend to replicate, and I believe that is still the case even if using large imputed SNP sets. Actually, now that people understand that most complex traits are highly polygenic, emphasis on being absolutely sure that SNP X is associated with the trait has given way to the notion that one wants to build a good predictor for the trait, and it doesn't matter if (say) a dozen out of 500 SNPs used in the predictor are actually not associated with it. Allowing in some false hits (or sometimes estimating an effect size to be nonzero when it is actually zero) to get a better overall prediction is OK. In the early days when some people thought there would be a gene for this and a gene for that (largely Mendelian architectures; who could have possibly believed that I don't know), there was much more emphasis on being very confident in the status of a specific SNP. But times are changing.

2016-09-20 on infoproc

Early predictors such as this one, or analogous predictors for disease risk (e.g., diabetes), are not likely to be used as an optimization target. Rather, they can be used to help parents identify embryos with unusually high risk of a negative condition. For example, unusually high risk for diabetes or cognitive impairment.

2016-09-18 on infoproc

Thanks. I think I covered Selzam et al. in…

Selzam et al. + this Dunedin paper = replication on different population samples.

2016-09-17 on infoproc

People always ask this question and my guess is that the researchers weren’t careful to calibrate the 95% income level — perhaps they used household income but survey respondents reported their individual income, etc . The main take away is that avg salary continues to increase even in the tail.

2016-09-17 on infoproc

The magnitude of correlation is modest but it is statistically significant. This kind of result — ability of polygenic score to predict achievement — has been replicated in more than one study now. The correlation strength will increase as more GWAS data accumulate and effect size estimates for each SNP become more accurate. At the moment the highest correlation is about 0.3 with high school level achievement in the UK. Height prediction using SNP genotypes is up to r ~ 0.5 and will also improve.

2016-09-12 on infoproc

IIRC, number of college age Asian Americans went up much faster than other ethnicities (except possibly Hispanics) during this period — their percentage of the total college age population almost doubled over this period.

Their representation among the most academically accomplished HS kids also went way up:…

2016-09-01 on infoproc

Dominance is a long way off. Even the challenges of ensuring a supply of oil from the ME in the face of submarine ops are formidable.

Hint: what do you think these artificial islands and forward military bases (Djibouti, etc.) are good for?

But the era of US dominance in the western Pacific may be drawing to a close.

2016-08-31 on infoproc

> What probability would you assign to the claim “China could destroy a carrier with their current capabilities”? < Again, this depends on the circumstances. But it seems very likely that there are a range of circumstances in which they could do it with current capabilities. Note we have not even talked about cruise missiles, which some rate as more dangerous than ASBM. http://www.andrewerickson.c……

“… the YJ-12 is the most dangerous anti-ship missile China has produced thus far, posing an even greater risk to the U.S. Navy’s surface forces in the Western Pacific than the much-discussed DF-21D anti-ship ballistic missile. The arrival of the YJ-12 is one more indication of how the U.S. Navy is falling further behind in the missile competition against China, exposing flaws in operating concepts that U.S. and allied commanders and policymakers have relied on for years.”

2016-08-31 on infoproc

Anti-satellite ops are a big escalation.

It seems likely that the missile launch will coincide with an interval during which the satellite still has coverage of the carrier. If updated targeting information can be sent to the missile in flight then the problem you point out is reduced considerably. Even with final targeting done by the missile, this would happen during approach, not necessarily in the final 3 sec. Think about what SAMs are capable of against much smaller, faster targets like fighter jets.…

2016-08-31 on infoproc

Are there actually blind spots in the satellite coverage? It seems one could design the network so that this is not possible. (See the Indian analysis.)

Carriers are not fast relative to missile speeds. There will have to be some terminal sensing on the part of the missile, as opposed to just hitting a fixed GPS coordinate.

2016-08-27 on infoproc

If she’s required not to disclose the answer to the question, why not just say that instead of lying?

I’m pretty sure Paul didn’t “officially” know what was going on in Benghazi at the time — he said this publicly in interviews. In other words, his committee was kept in the dark about these ops. Of course plenty of journalists eventually reported on it so he knew unofficially.

2016-08-25 on infoproc

Thanks — fixed the link to TC’s paper. The video explains the basic ideas pretty well, but the paper is better for experts.

2016-08-23 on infoproc

Not sure I understand your comment. I think she was lying (she knew very well what was going on there) and yes, I think it’s a crime to knowingly lie to Congress even if you are doing it for what you might think are reasons of national security. In this case she was protecting herself, not national security — every intelligence service in the world already knew what was going on (Libyan arms to Syria via Turkey) by the time these hearings were held. The real issue is whether she/Obama would admit to their actions in support of Syrian rebels (surely some of these weapons ended up with ISIS).

2016-08-18 on infoproc

It’s hard to war game a US-China pacific conflict, even a conventional one. How long before the US surface fleet is destroyed by ASBM/ASCM? How long until forward bases are? How long until US has to strike at targets on the mainland? How long do satellites survive? How long before the conflict goes nuclear? I wonder whether anyone knows the answers to these questions with high confidence — even very basic ones, like how well asymmetric threats like ASBM/ASCM will perform under realistic conditions. These systems have never been tested in battle.

The stakes are so high that China can just continue to establish “facts on the ground” (like building new island bases), with some confidence that the US will hesitate to escalate. If, for example, both sides secretly believe (at the highest levels; seems that Xi is behaving as if he knows this) that ASBM/ASCM are very effective, then sailing a carrier group through the South China Sea becomes an act of symbolism with meaning only to those that are not in the know.

2016-08-17 on infoproc

I had to do it because of a single annoying guy. I don’t know what happened with Scott Adams.

2016-08-16 on infoproc

Quantum key distribution …

2016-08-16 on infoproc

I started moderating all comments a while ago 🙁

Sometimes there is a delay if things get caught in the queue.

2016-08-13 on infoproc

Here’s a simple example which I think conveys the basic idea.

Suppose you have 10k variants and that individuals with 5.5k or more + variants are at the limit of cognitive ability yet seen in history (i.e., at the one in a million or billion or whatever level). Now suppose that each of the 10k + variants comes with some deleterious effect on some other trait(s) like general health, mental stability, etc. (This is actually too pessimistic — some will actually come with positive effects!) These deleterious effects are not uniform over the 10k variants — for some fixed number of + variants (i.e., 5.5k) there are many different individuals with different levels of overall health/robustness.

Let the number of distinct genotypes that lead to (nearly) “maximal historical” cognitive ability be n = (number of ways to distribute 5.5k +’s over 10k variants). Now, we know of many actual examples of historical geniuses who were relatively healthy and robust. The probability that these specific individuals achieved the *minimum* level of negative or deleterious effects over all n possibilities is vanishingly small. But that means that there are genotypes with *more* than 5.5k + variants at the same level of general robustness. These correspond to individuals who are healthy/robust but have greater cognitive ability than any historical genius.

You can make this argument fully realistic by dropping the constraint that + effect sizes on cognitive ability are uniform, that effects on different traits are completely additive, etc. The point is that there are so many genotypes that realize (cognitive ability ~ historical max), that the ones produced so far are unlikely to maximize overall health/robustness given that constraint. But that means there are other genotypes (off the surface of constraint) with even higher cognitive ability, yet still healthy and robust.

2016-08-12 on infoproc

Yes, you are correct that doping works and can have a big effect on a specific competition. The point I was trying to make is that going further and further out on the tail can yield even bigger effects than doping, albeit over a long time.

2016-08-04 on infoproc

I’m not trying to make any money from the blog — I don’t run ads, etc.

The number of uniques per month is a bit below 100k but greater than 60k for sure.

2016-08-02 on infoproc

I have great sympathy for the people of Ukraine but I don’t see realistically that it is in US interests to press so aggressively into what Russia regards as its traditional “sphere of influence” or “sphere of interest”. I think it risks a nonlinear response, which could be very dangerous. In the same way I don’t think the US will fight a war against China and DF-21s over some man-made islands in the South China Sea. I am sure that some Malaysians and Filipinos feel very angry about this but nations have to do what is in their own best interests 🙁

I have relatives in Taiwan, but any declaration of independence from Beijing would be an act of “suicidal nationalism,” and the US needs to make sure Taiwan understands that.

2016-08-01 on infoproc

As Cohen says, “Trump speaks elliptically…” = you have to piece things together, and you can’t be sure what he thinks or what he’ll do.

So my statement was probably a bit too strong.

I do think that Hillary is a liberal hawk when it comes to FP, and Obama and Trump are more realist in approach. The US should have learned (the hard way) that utopian positive outcomes from foreign intervention are unlikely. I think Trump understands this instinctively; note Obama’s admiration for Scowcroft.

2016-08-01 on infoproc

I doubt it. How do you think she’ll staff her security and foreign policy team?

2016-08-01 on infoproc

Missiles in Poland/Romania are potentially destabilizing to MAD and hence the benefit should be very high to compensate. As ABM I doubt these systems provide much protection against the Russian arsenal, but do create an additional US first strike capability, hence very dangerous.

Also, bar should be very high for US to intervene in Russia-China disputes.

I hope you can understand the difference between “Putin bad” and “We should risk nuclear war over an unfortunate situation in their sphere of interest”. Obama now does.

2016-08-01 on infoproc

Your link is critical of Cohen’s position on Ukraine. It’s an aspect of our Brawndo moment that he’s subject to this kind of attack for deviating from the Narrative. None of it has anything to do with the comments he made in the CNN clip I posted.

From the Obama interview I linked to above:

“I suppose you could call me a realist in believing we can’t, at any given moment, relieve all the world’s misery,” he said. “We have to choose where we can make a real impact.”

… “The fact is that Ukraine, which is a non-nato country, is going to be vulnerable to military domination by Russia no matter what we do,” he said.

I asked Obama whether his position on Ukraine was realistic or fatalistic.

“It’s realistic,” he said. “But this is an example of where we have to be very clear about what our core interests are and what we are willing to go to war for. And at the end of the day, there’s always going to be some ambiguity.”

After an 8 year learning curve as POTUS, Obama finally arrives at a realist foreign policy perspective that is closer to Trump’s than to Hillary’s.

2016-07-29 on infoproc

I think Obama, with an eye on his legacy, recently tossed Hillary under the bus for much of the FP stupidity (e.g., Atlantic interview, I think I blogged it). But certainly he deserves some of the blame. I’m looking ahead, though, not backward.

2016-07-29 on infoproc

Hillary foreign policy is very stupid and very dangerous.

2016-07-29 on infoproc

Thanks! Please post the link if you can. I’ll add it to the main post.

2016-07-13 on infoproc

They were grading the exams last night 🙂

2016-07-12 on infoproc

Gordon meetings are confidential. But I can comment on already published work. I was especially impressed by the ability of researchers to detect cancer cells in circulating blood at a very early stage via sequencing. Simple machine learning + omics allows to detect cancer or infectious disease very early using signals of moderate dimensionality. Also, ability to look at single cell function using advanced imaging is really impressive. We still have a couple days to go … 🙂

US lead on researchers in Asia is narrowing very fast in many of these fields.

2016-07-11 on infoproc

You could email the math dept. at HKUST. They aren’t really set up to sell them to the public, but might help you out anyway.

2016-06-26 on infoproc

Certainly there are people capable of doing good research who don’t have PhDs.

2016-06-23 on infoproc

I thought it was good. But I had already seen the notes from the startup course he taught at Stanford, so I knew what to expect.

2016-06-20 on infoproc

That was my recollection, too. It appears he is not in the academy now but he has a PhD and published research.

2016-06-18 on infoproc

Psychometric measures are not really that accurate, nor that (noiselessly) predictive.

Intense motivation can, to some degree, overcome cognitive limitations. I wouldn’t try to stop someone who didn’t score well on a test from trying to do something they are very interested in. Hopefully they can figure out for themselves whether they are really able to get up to speed and contribute.

When dealing with large numbers of individuals — for example, admitting students to a freshman class of thousands of kids — the noise tends to cancel out, so that test scores yield a reasonable *statistical* predictor. But one should always allow more room for uncertainty when dealing with a specific person.

2016-06-14 on infoproc

Hi Malcolm,

Any cognitive measure that has a decent correlation with g is useful for identifying cognitive variants. So SAT, ACT, Wonderlic, etc. are all good. If cost were no object we’d opt for a fairly sophisticated battery of cognitive tests, perhaps covering M, V, and Spatial reasoning, personality inventory, etc.

10 points at the societal level would be very noticeable. At the individual level not so much — in fact the intrinsic uncertainty in the accuracy or meaning of the scores is probably in the 10 point range.

There is an elementary primer on population genetics by Gillespie that might be a place to start. My article on arxiv might be useful as well:

There are also many elementary articles / lectures on machine learning or compressed sensing or LASSO that might be useful.

2016-06-14 on infoproc

Thanks! I ran it through their processor but the output still needs some editing.

If anyone is bored and wants to improve the transcript, it’s here:…

If more than one person gets involved, you may want to alert each other to avoid edit collisions.

2016-06-11 on infoproc

Many thanks!

2016-05-27 on infoproc

I hope so too!

2016-05-27 on infoproc

I wouldn’t take the “IQ = 160” line on the graph very seriously.

I do think that one can consider this a reasonable collection of high cognitive ability outliers, for research purposes.

2016-05-27 on infoproc

I would be interested in this as well. Note, though, that US-born S. Asians often have “ethnic” first and last names, unlike E. Asians. The anti-semitism measurement mentioned by Anon IV would also be interesting.

2016-05-26 on infoproc

It’s relative to the response rate to white males.

2016-05-26 on infoproc

I think ablation-reaction pressure is also important.

2016-05-26 on infoproc

I thought the same thing. The researchers unwittingly chose the wrong “white” name!

2016-05-23 on infoproc

There are lots of videos showing the secrets behind famous magic tricks, as I recall from looking a few years ago. There have even been TV shows devoted to this topic. In the case of nuclear weapons, the devil is in the details, so knowing the basic idea is very far from making a real bomb.

2016-05-17 on infoproc

Harvard is not #1 if you look at per capita measures … it’s another school 😉

SAT scores:…

Nobel prizes won by alumni:…

2016-05-15 on infoproc

Correlation between EA and g is positive but less than 0.5 or so (depending on country). So EA is only a noisy estimator of g. That’s why you need more EA data than height data to achieve similar statistical power. But there is evidence that the EA hits are also hits on cognitive ability. Find someone to fund a study with good cognitive phenotyping (e.g., with SAT or IQ scores) of a few hundred thousand individuals and we’ll get many more g hits.

2016-04-28 on infoproc

I’m not responsible for Mensa 😎

I don’t know of evidence that supports g or any specific IQ measure as more predictive of later life success than, e.g., SAT (new or old). The problem is getting sufficient data to test these hypotheses. SMPY had to rely on SAT administered to <13 year olds, and that has become fairly standard practice for gifted programs in the US. Is there something better? Perhaps, but no one really knows. For high stakes testing you need lots of psychometricians (people who know item-response theory and some statistics, as well as having some psychological or educational insight) producing and validating new questions every year, and large samples of test takers. Not very many entities can pull this off. Re: grinding, I'd like to separate cognitive power from conscientiousness or work/effort. They are not the same thing. One of the problems with the college entrance exams in Asia (China, Japan, etc.) is that they jumble all of those things together by mixing cognitively hard problems with memorization of facts. There are really brilliant individuals who don't score at the top of such exams because they can't bring themselves to spend years grinding in preparation. In my terminology a "g man" is someone who ascribes excessive meaning to the g = general factor construct as opposed to taking a pragmatic view of testing. Perhaps the Mensa gatekeepers are g men -- they dropped the SAT/GRE as insufficiently g-loaded.

2016-04-28 on infoproc

I’m not sure what you mean by my position on IQ. I think cognitive testing allows us to make crude estimates of individual differences in cognitive ability. I am not sure that IQ tests have higher validity (predictive power) for college/graduate school success than SAT/ACT/GRE. In general, if we can develop better psychometric tools, then we should. For example, some have argued for adding a spatial ability measurement to standardized tests. I do think we should stop reducing the ceiling of the SAT.

I am worried that sufficiently determined individuals (e.g., some Asian Americans, but also others) can increase their scores (reducing the validity) by grinding. But that would probably be true for any high stakes test that becomes widely adopted.

I often argue against denying the existence of individual differences, or denying that we can even make crude measurements of those differences. But I don’t consider myself a “g man” — i.e., IQ extremist.

2016-04-21 on infoproc

From first link: “In the end, however, I have to agree with old Wilbur Bender, the Harvard admissions dean who fought off idealistic faculty committees in the the 1950s. A Harvard that followed Pinker’s advice would, after a generation or two, be reduced in status, prestige, and endowment size, to a mere Caltech or Ecole Normale Superieure.”

2016-04-21 on infoproc

My detailed thoughts on Harvard admissions:……

I consider myself a pragmatist on these issues.

2016-04-20 on infoproc

Business guys would dominate in Extraversion!

2016-04-20 on infoproc

Thanks! I will fix the post.

2016-04-20 on infoproc

Who needs video games when Julian can do it via a brief conversation? 😎

2016-04-18 on infoproc

Our platform does not have the elimination of affirmative action as a goal. We merely want greater transparency about admissions.

It is interesting that over the last ~25 years, the Asian American percentage of the admitted class is at its highest exactly during the periods when Harvard has been under scrutiny (early 90s and now). Our candidacy and the ongoing lawsuit probably contributed to the 22% number.

More here:

2016-04-17 on infoproc

As it was explained to me, someone working for Harvard on the Overseer ballot thought MI was the state abbreviation for Missouri. That person evidently didn’t read the rest of my entry, or else assumed that I have a very long commute to work at Michigan State University. I’d like to think this sort of thing wouldn’t happen at Caltech 😉

2016-04-11 on infoproc

I added links to references in the post above.

2016-04-09 on infoproc

The identified SNPs will likely allow for phenotype (IQ) prediction, which has obvious utility even if the underlying neurobiology is not yet understood.…

2016-04-09 on infoproc

For example, rather than alter the RNA or protein itself the variant could change the conditions under which the gene is expressed. Given the complexity of living organisms, there have to be lots of controls of this type in DNA. Otherwise, the “programming language” is too simple.

2016-04-07 on infoproc

No idea yet, because we only know the genetic regions of a small number of early hits.

2016-04-07 on infoproc

Read the figure caption — raised identically.

2016-04-07 on infoproc

Consistent with GCTA h2 for somewhat better phenotyped samples.

2016-04-07 on infoproc

I’m sure Mark would appreciate any specific feedback you have — for example, if there are specific definitions or claims that are wrong, please let him know. He chooses to emphasize the term locality, which I would not have done, but that isn’t necessarily wrong. BTW, he’s not exactly a crank — his advisors were Coleman and Wilczek 😎

” … The violation of the Bell inequality is often described as falsifying the combination of “locality” and “realism”. However, we will follow the approach of other authors including Bell [2–4] who emphasize that the EPR results violate a single principle, strong locality.

Strong locality, also known as “local causality”, states that the probability of an event depends only on things in the event’s past light cone. Once those have been taken into account the event’s probability is not affected by additional information about things that happened outside its past light cone. … ”

2016-04-07 on infoproc

Yes, originally he quipped that 10k g variants meant that half of all genes were involved in intelligence. It showed a lack of understanding of modern genomics. I wonder if he’s now edited out this part of his original post.

2016-04-07 on infoproc

Hi Lubos: Mark’s paper is going to appear in AJP, which is a pedagogical journal. He’s trying to explain everything in the simplest way possible. If you look at the paper he spends half the time trying to explain (philosophical) terminology. He makes a distinction between strong and weak locality, etc. I personally would have put more emphasis on reality than locality — his locality might also be called local realism by other people.

2016-04-06 on infoproc

There are many genetic variants that are known to affect cognition, for example:………

Biochemical and neurophysiological effects of the ~70 common SNP hits are under active investigation. Stay tuned.

2016-04-06 on infoproc

Tonon’s leglock game is amazing. He had Palhares in trouble many times.

2016-04-06 on infoproc

I think there will definitely be limits of the type you mention for height. The IQ=1000 scenario is subject to lots of unknowns. But I would be surprised if we couldn’t do better than all historical geniuses. That only requires a small extrapolation of linearity.

2016-04-06 on infoproc

My guess is that horse breeders could do better if they went SNP-based like the cattle breeders. But maybe they’ve really hit a limit.……

2016-04-06 on infoproc

I agree — an SD or two would totally transform society.

The 1000 IQ thing is mainly a fun SciFi idea. It does play an important role in the discussion of AI threat, though.

2016-04-06 on infoproc

Additive models work well in cattle genomics as well:……

2016-04-06 on infoproc

[ Myers seems to be confused about the difference between specific (protein coding) genes, of which there may be only ~20k in the human genome, and the set of all variations in the DNA code, of which there are many, many more. Thousands of variants (or 10k) out of this much larger number is a tiny fraction much less than one. ]

2016-04-06 on infoproc

We can imagine that the tradeoffs (e.g., brains are energetically expensive; big heads cause problems in natural childbirth; etc.) required to increase intelligence had negative evolutionary fitness effects 50k years ago, but that does not mean they would be debilitating today. You could make the same arguments about extra muscle in chickens.…

2016-04-05 on infoproc

IIRC in QM you can give up CFD *or* (some level of) locality. Instead of giving up on certain “elements of reality” (i.e., that determine the spin measurement outcomes), one could assert that an influence passes (outside the lightcone) from one measurement to the other. To have independent elements of reality (i.e., one for each photon) requires imposition of locality. Hope that makes sense 🙂

2016-04-02 on infoproc

Great Diaz story. But he probably burned half or more of the calories he got from the burrito doing BJJ, and what if he didn’t get the burrito at the end?

2016-03-25 on infoproc

Obama’s presidency is almost over. He’s working on his legacy. But it’s valuable to see what he learned in his 8 years. Very few people have seen the world from his position, and you can tell from his comments that he’s thoughtful and intelligent. You can examine his world view now, and then ask yourself which of the current candidates is closest to that.

2016-03-24 on infoproc

I can certainly imagine examples where the algorithmic speedup is greater than the hardware speedup over some recent period. However, the hardware improvements help *every* computational task. So if we had to assign overall credit for information technology advances over the last decades, I’d have to give most of it to applied physicists, semiconductor engineers, etc.

2016-03-20 on infoproc

Both nice, but I guess I prefer DM.

2016-03-10 on infoproc

Yes, it will be video recorded.

2016-03-08 on infoproc

Regression to the mean happens even in a purely additive model, because heritability is less than 1 and the environmental “luck” that outlier parents had will not, on average, be passed to their kid.…

Regression to the mean is just a little bit too g loaded to be understood by the typical reader of any blog that I know of 😉

2016-03-07 on infoproc

3 gun — nice! But we already knew Keanu was good athlete, so no surprise.

Note that Diaz can make 155 and although he’s good he’s not the champ. So this suggests to me that McG should stay at 145 or at least move up with caution. 170 is totally out of the question for the “anorexic leprechaun” (as Rockhold christened him). He’s a little dude who has succeeded by beating even smaller dudes. Can you imagine him against 6’2 Condit?

I think McG is smart and I enjoy watching him train and fight and promote fights. I think he’ll learn from this experience.

“It is not the critic who counts; not the man who points out how the strong man stumbles, or where the doer of deeds could have done them better. The credit belongs to the man who is actually in the arena, whose face is marred by dust and sweat and blood, who strives valiantly; who errs and comes short again and again; because there is not effort without error and shortcomings; but who does actually strive to do the deed; who knows the great enthusiasm, the great devotion, who spends himself in a worthy cause, who at the best knows in the end the triumph of high achievement and who at the worst, if he fails, at least he fails while daring greatly. So that his place shall never be with those cold and timid souls who know neither victory nor defeat.”…

2016-03-06 on infoproc

Conor should stick to fighting little guys @145. He just got smoked by Diaz, a 155er. He actually wanted to fight Lawler @170 which is a joke.

2016-03-04 on infoproc

Wow! Thanks 🙂

2016-03-04 on infoproc

I’ve been a big fan since he joined the UFC. But I’d like to see him face a good grappler after a full camp.

2016-03-04 on infoproc

I think I’m among the 5-10% most pessimistic (is that an outlier?) on the chart: very low probability of HLMI by 2050.

2016-03-04 on infoproc

How do your HLMI predictions compare to the experts in the chart above?

How big a factor is deep learning in your expectations? How about the end of Moore’s Law?

2016-03-02 on infoproc

You’re right that he really didn’t seem to understand the question.

2016-03-01 on infoproc

Some people have been touting AlphaGo as the most impressive feat towards AGI (Artificial GENERAL Intelligence; i.e., “Deep learning changes everything!”, “The geek rapture (singularity) is near!”). That’s why it is interesting that without big GPU clusters AlphaGo is only a bit better than previous Go playing programs.…

2016-02-28 on infoproc

Trump is obviously smarter and more thoughtful than people give him credit for. They don’t understand he is playing a role and has been for many years. A lot of the glitz is just for the rubes (now voters).

He talks specifically about analyzing downside risk (worst case scenarios) in the interview. He would stop the demonization of Putin and Russia. He is moderate to left on social issues. He has thought a lot about the risks of nuclear war (see interview).

The recent Obama/Hillary foreign policy has been a terrible failure in the middle east, Ukraine, and elsewhere. I don’t think Trump would have made those mistakes, which resulted from stupid idealism (same could be said of W and Iraq).

Business leaders can be fools just like politicians, but people who have built businesses have seen the direct consequences of their decisions, and generally tend to have a realistic view of how the world works.

I was an early Obama supporter, and of course he was barely qualified to be Senator, let alone President. But the fact that he was able to run his campaign so effectively, and beat Hillary, suggested he was an able leader. The same could be said for Trump (although of course he’s had much, much more experience in life and business than the 2008 Obama) in terms of his current campaign. But people are reluctant to give him credit. They think he’s the character he plays on TV and on the political stage.

2016-02-26 on infoproc

Trump is not an idiot, despite what his detractors say. His people smarts are probably very high — he’s a dealmaker and negotiator, after all.

2016-02-24 on infoproc

See Schmidhuber at the link below for a scrupulous attempt at credit assignment:…

He was doing video games well before DeepMind:…

2016-02-24 on infoproc

I think almost all the claims of the PNAS paper are disputed, so take that sentence with a grain of salt.

Yes, it would be great to have a better “natural unit” for the measurement of intelligence. Clearly it’s much more complex than height. What cognitive ability score would we give to a chimp, or to AlphaGo? Underlying this are biological mechanisms that make a particular neural net (i.e., human brain) work slightly better or worse.

2016-02-24 on infoproc

Not sure I understand your question. As long as the underlying causal variant does the same thing in both populations, and it is well-tagged by a particular SNP, then you should detect the same effect in both populations.

2016-02-23 on infoproc

Thanks for the link!

2016-02-23 on infoproc

This is from GIANT so I think it’s all (well-nourished) Europeans. A genomic predictor is built using a sample population that experienced a specific environment, and its accuracy will degrade if it is applied to a group that experienced a different environment (e.g., famine, malnutrition).

2016-02-23 on infoproc

First hits will be variants with larger effect size and MAF, and in good LD with SNPs found on standard chips. Therefore, will not be representative of the entire spectrum of causal variants. Although we have ~ thousands of height hits there may be ~10k in total and we can’t say much about the ones that remain to be discovered.

2016-02-15 on infoproc

See the blue columns in the Economist chart. These numbers are from the Census CPS survey, so they are noisy from year to year. Smooth the blue columns and you have a sense of the growth in college age AA population since around 1990 — very roughly, it looks like a doubling.

2016-02-15 on infoproc

Evolution works fastest on existing additive variance but for really significant changes complex (non-additive) “gadgets” have to evolve via mutation. The latter takes much, much longer.

Dominance effects are non-zero, but smaller than additive effects (“to first approximation”).

2016-02-15 on infoproc

I think the most robust observation one can make is that the Asian American college age population is significantly larger than 20 years ago but the Ivy/H component hasn’t changed much. If anything, I suspect the quality *and* quantity of the pool have improved over time. Why that is not reflected in the Ivy/H numbers, but IS reflected in the Caltech numbers is a mystery that can only be resolved through greater transparency. We may already have de facto quotas, as we had in the past.

2016-02-01 on infoproc

Thanks, I think the links are fixed now.

2016-01-28 on infoproc

There is an accreditation process. But re: admissions practices, it’s up to you, me, and SCOTUS to keep them honest …

2016-01-28 on infoproc

Established beyond doubt using Duke data, despite the fact that incoming recipients of racial preferences were more likely to declare STEM interest in their applications. Rich legacy admits (also recipients of preference) underperformed regular admits as well.……

I wonder if determined Overseers would be able to obtain similar data at Harvard?

2016-01-28 on infoproc

As I mentioned in this earlier post, part of having a realistic discussion on this topic is admitting that subjects vary in difficulty. So, of course there are easy paths through Harvard, and one can’t really claim that mismatch applies for students in those majors. But for STEM at least, mismatch is real and actually hurts the people that preferences are trying to help.…

2016-01-27 on infoproc

In any case, it will be very interesting to see if they can edit back using CRISPR. Did you catch the part about how they accelerated the generational timescale using fancy reproductive tricks?

“… the next generation of macaques, which the team created with unprecedented speed. When the monkeys were 27 months old and not yet sexually mature, Sun’s team took testes from the males, matured the tissue artificially by grafting it under the skin on the backs of castrated mice, and used the resulting sperm to fertilize eggs from non-engineered macaques. The offspring showed asocial behaviour at about 11 months.”

2016-01-26 on infoproc

How about “via genetic modification”?

2016-01-23 on infoproc

Obama in the video here is at least an inch or two taller than 6″2
Bill Clinton. Either Obama is significantly taller than 6″1 or he wears

2016-01-16 on infoproc

Thanks! I’m glad I avoided my trivial altercation 🙂

Afterwards I was talking rationally to my friend about how stupid it would have been to fight. But I was on autopilot for those few seconds. I suppose there’s a Kahneman fast vs slow thinking aspect to threat response.

2016-01-15 on infoproc

Contact Ron directly (see email in main post) with your mailing address.

2016-01-15 on infoproc… ??

Or is there another paper on non-family violence?

2016-01-15 on infoproc…

High Trans-ethnic Replicability of GWAS Results Implies Common Causal Variants (PLOS Genetics)

2016-01-15 on thecrimson

Our platform is simple. We are in favor of

1. More transparency in Harvard admissions
2. Increased use of the endowment to make Harvard more accessible

Harvard degree holders (including from the graduate and professional schools, but not current faculty or officers) who would like to support Free Harvard, Fair Harvard can sign our petition. Please contact Ron Unz with your (physical) mailing address:

2016-01-14 on infoproc

Thanks for your support 🙂

Munger actually attended Caltech although he did not graduate. I would love to get in touch with him…

2016-01-14 on infoproc

Every time I visit my mom I consider chucking everything and retiring there 🙂 It must have been great to grow up in that neighborhood. Did you go to Pismo Beach schools? How are they?

2016-01-12 on infoproc

Near SLO, just north of Pismo.

2016-01-01 on infoproc

I took a quick look at it. My guess is that Visscher et al. will respond to the paper. It has not changed my opinion of GCTA. Note I have always thought the standard errors quoted for GCTA are too optimistic, as the method makes strong assumptions (e.g., Gaussian random effect sizes). But directionally it is obvious that total h2 accounted for by all common SNPs is much larger than what you get from only genome wide significant hits. For example, the set of genome wide significant hits for some traits (e.g., height) has been growing steadily along with h2 accounted for using just those hits — eventually approaching the GCTA prediction. That is, even *without* GCTA the steady progress of GWAS shows that common variants account for significant heritability (amount of “missing” heritability steadily declines with GWAS sample size), so the precise reliability of GCTA becomes less important.

Regarding this paper, they make what sound like strong theoretical points in the text, but the simulation results don’t seem to justify the aggressive rhetoric. The only point they really make in figs 4,7 is that the error estimates from GCTA in the case where SNP coverage is inadequate (i.e., using 5k out of 50k SNPs) are way off. But this doesn’t correspond to any real world study that we care about. Real world results show that as you approach ~few x 100k SNPs used the h2 result asymptotes (approaches its limiting value), because you have enough coverage of common variants. The authors of the paper seem confused about this point — see “Saturation of heritability estimates” section.

What they should do is simulate repeatedly with multiple disjoint populations (using good SNP coverage) and see how the heritability results fluctuate. But I think that kind of calculation has been done by other people and does not show large fluctuations in h2.

Well, since you got me to write this much already I suppose I should promote this to an actual blog post at some point … Please keep in mind that I’ve only given the paper a quick read so I might be missing something important. Happy New Year!

2015-12-26 on infoproc

Nope, not yet. It’s just something I made up 🙂

2015-12-24 on infoproc

I’m partial to the King James version 🙂

Merry Christmas!

2015-12-19 on infoproc

“Among known U.S. baccalaureate-origin institutions of 1997–2006 black S&E doctorate recipients, the top 8 and 20 of the top 50 were HBCUs. ”…

“The top 5 baccalaureate-origin institutions of 1997–2006 black S&E doctorate recipients were: Howard University, Spelman College, Hampton University, Florida A&M University, and Morehouse College.”

2015-12-04 on infoproc

Editing and replication are not the same thing. Replication touches (duplicates) *every* base in the DNA, so has many more opportunities for error. Guide rail targeting is the key step for editing — the overwhelming majority of the DNA is untouched in the editing process.

You can conceptually break the two processes (editing vs replication) into steps and assign error rates to the steps. It then becomes clear that the *total error rates* (which you are using to guide your intuition) are not directly related to each other. Evolution could produce a targeting efficiency which is better than the improved Cas9 but a *total error rate* in replication (which involves more steps than just targeting a single region) which is higher.

Why don’t we bet $1k on whether Zhang’s results in the recent Science paper will replicate? You should give me 100 to 1 odds since you are so confident. The answer will be known within a year or two at most.

2015-12-04 on infoproc

IF gene editing continues to improve at this rate, then the main bottleneck will be the sample size of good (cognitive, genotype) data sets necessary to extract the genetic architecture. IF we can get to ~ millions (very plausible in 5-10 years), then we can extract SNPs accounting for most of the genetic variance. Editing would then allow many SDs of improvement. So 10 years is not impossible, and in 30 years the super-geniuses would be adults … But who will have the guts to try it? There could easily be a decade or two lag between when it first becomes possible and when it is actually attempted.

More modest advances are likely via IVF selection (driven by self-interested parents), which then raise confidence levels in our ability to do genomic prediction. Plenty of parents want to make sure their kid is above average, or to select against negative fluctuations. Not so many want a historical super-genius via a riskier path.

2015-12-04 on infoproc

The new Cas9 variants are going to be widely used — every major research university in the world has people working with CRISPR already. So false claims are going to be detected very quickly.

2015-12-02 on infoproc

Right on schedule 🙂

Lubos even has a long post on this topic on his blog:…

Without getting into an all out religious war, a few comments:

1. I think Lubos and I agree on the measure problem in MW, and that it’s not solved.

2. I wrote this paper because there are now giant tomes (e.g., [14]-[17] in the paper, including Wallace’s huge book from Oxford Press) claiming solutions based on subjective probability. But I don’t believe any of these solutions are satisfactory.

3. The current paper overlaps with my earlier pedagogical paper on probability in QM ( ) because I re-use some of the introductory material (section I and part of section II). The first sentences in the new paper: “We begin with a brief introduction to no-collapse or many worlds quantum mechanics. A longer version of this material can be found in an earlier paper by the author [1].”

4. Actually, Lubos is right that the new paper is mainly an elaboration of points I made (much more tersely) in earlier papers. He’s even right in his blog post in saying that anyone sufficiently smart would find these points obvious 😉

5. Lubos and I disagree about the Copenhagen Interpretation. I don’t like it and would prefer to find something better, but it’s possible that there isn’t anything better. There are greater physicists than either of us on both sides of the issue.

2015-11-29 on infoproc

No, they live down the street 🙂

2015-11-24 on infoproc

I liked the new episodes, overall. They stretched the material out more than necessary, though. Some of the “tactical” plotting seemed stupid. I really hated the Marshall character (something out of Stephen King?). But the material has emotional power. Each of the main characters is trapped in a moral quandary, which I found powerful. (The Juliana-Frink-Joe triangle is a bit tiresome, though.)

In the Amazon version is Hitler the Man in the High Castle? Where do the movies come from?

2015-11-18 on infoproc…

Episodes 1 and 2 are available now, I think.

2015-11-18 on infoproc

There were two episodes (I guess you could call them a pilot) made already. I think on Amazon they refer to those as season 1.

2015-11-13 on infoproc

In the realm of entertainment I agree with you re:VR. Do you have any links describing what’s in store in the near term?

The potential I had in mind inspired by FPV drone racing is, e.g., in the area of military applications.

2015-11-09 on infoproc

I wasn’t trying to make a direct connection. I have multiple interests — unlike some others I don’t try to link everything together under a single theme 😉

I’m actually dubious that the size of dysgenic effects resulting from, e.g., the Industrial Revolution, so far are very large. Negative effects are probably swamped by better nutrition, access to education, globalization, etc. There are many more smart people around today than there were in the past.

But Huxley is a fascinating figure and in many ways Brave New World is deeper and more appropriate for the 21st century than 1984.…

2015-11-09 on infoproc

Ronald Reagan used to say that he wasn’t responsible for the opinions of his supporters. After all, they were following him, not vice versa.

2015-10-25 on infoproc

Yes, the big decline on the graph represents what has been happening over the past few years, but for some reason NIH decided to recognize it all at once. The important point is that sequencing costs have continued to decline.

$100/genome would require a big technological breakthrough (imagine a small desktop device or even a USB device like Oxford Nanopore). The bioinformatic cost to deal with the data is more than ~$100 per genome so there would have to be advances on that end as well.

2015-10-25 on infoproc

I said NOT as precipitous as shown on the graph. It’s been happening over time.

I know for a fact HLI has gotten their all-in costs below $1k for high coverage sequencing. But the cost varies depending on who is doing the work. See link for more


2015-10-25 on infoproc

Illumina is the market leader and have been setting prices for the last few years. The cost decline is not as precipitous as shown on the graph — Illumina have been dribbling out the decreases bit by bit, in response to perceived competitive threats, etc.

2015-10-19 on infoproc

It’s amazing how little intellectual integrity even professional “scientists” and academics have. I suppose much of this is subconscious — it takes discipline to regularly note strength of belief in various hypotheses over time. Unlike portfolio managers, academics are seldom made to pay for their mistaken judgements …

2015-10-19 on infoproc

Very strange, considering I am using a standard (Blogger = Google) template. The embed scripts I use for video come directly from YouTube, etc. I do load a fair number of posts on the front page, so if you have RAM limitations this might be an issue.

2015-10-15 on infoproc


Thanks for your kind note. My dad was very dedicated to his students, and a good father as well.

Best wishes,

2015-10-12 on infoproc

There are estimates from classical family/twin designs but the uncertainties are significant. It seems likely that the additive part is much larger than the non-additive part. But even if (say) the non-additive part is half the additive part, the nonlinear methods we’ve developed should remain effective. Another important parameter is the *number* of QTL-QTL interactions (the specific genetic architecture — not just the variance accounted for). As long as this doesn’t scale up too rapidly (worst case: each QTL interacts significantly with every other one) then methods like ours will work. This seems to be the case in the yeast results.

2015-10-07 on infoproc

The throwing part of judo (tachiwaza) is higher impact and more dangerous than the groundwork part (newaza). BJJ is mostly newaza and the takedown game is emphasized less. The takedown techniques in BJJ are often more wrestling style than judo. These can also lead to injury but are lower in impact than the high amplitude throws that are common in judo.

2015-10-05 on infoproc

I can’t speak for the other panelists but I find the hypothesis of (significantly) lower genetic g due to dysgenics to be implausible.

2015-10-02 on infoproc

Effect sizes for these variants are very large — i.e., they probably boost probability of having autism from something like 0.01 to 0.2 or more. So the sample size required to detect is much smaller than for the traits like height. The PNAS worry you have I can’t really address…

2015-09-28 on infoproc

>> if something can be approximated by a series, it does not in any way mean understanding of the underlying function. << I totally agree that to understand the underlying function is much harder than just making predictions using some mathematical approximation. But the latter may have some "engineering" applications 🙂 >> biophysics and molecular biology completely lost their original meaning << Yes, the terminology today is different than in the early days. I just wanted to point out that some physicists already snuck into biology a long time ago and did make some contributions 😉

2015-09-27 on infoproc

I’m afraid you misunderstand the difference between claiming that something could be linear to first approximation and claiming that it is completely linear. If you would just read the animal/plant breeding literature you would find that there is already plenty of evidence that linear models are useful (not exact, but no one ever claimed that). Very few of the people working on human genetics are familiar with predictive genomics in animal breeding. This is simply due to bounded cognition — they can’t pick up a slightly mathematical set of papers / textbooks and read them with comprehension. They can’t even follow Fisher’s fundamental theorem of population genetics although it is directly relevant to their work.

The failure of the field wrt “candidate gene” (early GWAS) studies cannot simply be waved away with the usual “it’s complex, how could we have known!” excuse. The field (a) was overconfident in its “biological intuition” and (b) did not understand how to do a proper correction for multiple testing and so fooled itself for many years before the leading journals (finally) adopted the “genome wide significance” threshold (which merely includes the correction for multiple-testing for ~1E6 SNPs) that has (now) led to routine replication of results. It was a lack of understanding of basic statistics. Most of the people in the earlier generation (and still many today) are just not quantitative. It was bounded cognition no matter how much you want to use “complexity” as an excuse.

Of course you can’t accept me as an ugly physicist pointing these things out. But instead perhaps you could trust Ioannidis who is a bona fide genetics and biomedical researcher.

PS Physicists *did* revolutionize finance (quants; for better or worse!) and *did* revolutionize biology — Delbruck, Crick, etc. Molecular biology once went under the name “biophysics”!

>> Perhaps the only mathematician to name a new biological discipline, in 1938, as Director of the Natural Sciences Division of the Rockefeller Foundation, Warren Weaver coined the term molecular biology to describe the use of techniques from the physical sciences (X-rays, radioisotopes, ultracentrifuges, mathematics, etc. ) to study living matter. <<

2015-09-26 on infoproc

It’s not clear to me what the take home message is from Tetlock’s most recent work, other than that general methods of rationality (Bayesian thinking, rigorous self-examination, etc.) are useful. My guess is that most of his “supers” are highly rational, high g individuals who for some reason are sufficiently motivated to look into minutiae. (Since they were not being paid there was probably a lot of variance in how much effort different forecasters devoted to the competition.)

Note that rationality and g, although somewhat positively correlated, are independent constructs. You need both to be a good predictor: http://www.keithstanovich.c…

The message from Tetlock’s earlier work was simply that returns to “expertise” could be quite low in certain (complex or soft) areas of prediction. That seems obviously true, for example to anyone familiar with financial markets.

Here’s something from BBC:…

>> Hare and Roberts alike agree that an added benefit of the Good Judgment Project was facilitating ways for hyper-intelligent “supers” to find each other and develop ways to collaborate. Hare’s first super-forecaster conference, he says, was something of a revelation. “It’s like that bit at the end of ET,” he says, “when all the other ETs come and get him. He’s not an alien anymore.” << You always seem peeved when I suggest that in certain soft scientific subjects it is common for "expert researchers" to be really confused about their own specialty. But I am quite sure this is true and the work by Ioannidis supports my contention. Careful review (which is seldom done in most fields) of, for example, what "leading experts" thought was true about genomics over the past 20 years would easily validate my perspective. Ioannidis found that ~(5-10) years of "candidate gene" studies from the early 21st century can all be tossed into the garbage -- replication rates are in the low percents. I wonder how much NIH spent on these confused researchers and their studies? On the other hand physical scientists are very familiar with Bayesian/statistical reasoning. They have a good feeling for levels of certainty in their specialty: what is really known, and what has yet to be rigorously established. This is all part of the transition (if it ever happens) from soft to hard discipline.

2015-09-17 on infoproc

22lr is a versatile round, and would be a key unit of exchange in a post-apocalyptic America. It’s actually very effective for defense (depending on situation) due to low recoil, low noise, and less chance of over-penetration. Chechen snipers used 22lr and the Russians now issue a 22lr sniper rifle. Of course, at longer distance or against body armor or in a direct firefight a larger caliber would be better.

I wonder whether preppers are the cause of the multi-year scarcity of 22lr we’ve been experiencing.

2015-09-16 on infoproc

It’s like being able to swim or perform CPR. Most people can get through life without it, but it could come in handy!

Or maybe I read too much Heinlein when I was growing up 🙂

2015-09-16 on infoproc

Never shot anything full auto. Something strange is happening with 22 ammo — hard to find and (relatively) expensive right now. Might have to move to 223 🙂

2015-09-13 on infoproc

Yes, good programs can add value through superior player development, or recruiting undervalued HS athletes (judging talent more effectively). It’s possible that small colleges in the lists above outperform by doing a better job of educating students. Also, some schools may have a better admissions filter: numbers over (clueless) holistic review, or really really good holistic review.

The question is whether the recruiting “star system” has validity: do players with higher ratings coming out of HS have better careers *on average*? Well, a 5 star is >100x more likely than a 2 star to become an NCAA All-American or make it in the pros. It turns out the academic situation is similar but people are reluctant to admit it.

For every Kaminsky there are hundreds of lightly recruited players who never even get to start a game in college.

2015-09-12 on infoproc

Let me know when your page is up and I will link to your paper 🙂

2015-09-12 on infoproc

Four and Five star recruits almost always end up at the powerhouse programs, and they are 100x to 1000x more likely to make it to the pros than lightly recruited athletes who are nevertheless offered college scholarships. You can’t argue by anecdote — you have to add up the numbers and use statistics.………

2015-09-10 on infoproc

Can you post a link to a copy of your paper (ungated)? I’m sure readers would be interested!

2015-09-04 on infoproc

I agree with you about Sanders. The question is whether he can mount a *competitive* campaign that is mostly crowd funded. Trump, at almost 70, and with billions in net worth, could self-fund such an effort. Ideally, we would modify campaign finance laws so that crowd funding can compete with oligarchs and special interest groups 🙁

2015-09-03 on infoproc

Genetic data is generally protected under IRB. You won’t see it made public unless individual participants in studies want to make their data public, and this is rare.

For example, the study this post is about used ~44k individuals, but that data is only available to a select group of researchers. There are issues of scientific competition as well, of course.

Worldwide there are probably ~2 million SNP genotypes stored somewhere, but most of the data is in silos and can’t be merged. This is holding the field back, actually, but can’t be avoided due to privacy regulations.

You could write a Sci Fi story about hackers who steal data in a race to decode the genome 🙂

See earlier comment:…

2015-09-02 on infoproc

I don’t think any of these quantities should be trusted to 10% accuracy. The main point is that there is no reason to doubt that we will accumulate more h2 as statistical power goes up, and no reason to think that classical methods of heritability estimation aren’t in the right ballpark.

2015-09-01 on infoproc……

In general, these data are not public. They are tightly controlled by various collaborations.

2015-08-31 on infoproc

EBI is an improvement over sport BJJ, but I prefer the no time limit format to sudden death overtime. The great thing about the BJJ Kumite videos is that both fighters know they will have to compete until someone taps. You can see them wearing down as the 30 or 60+ minutes elapse.

2015-08-22 on infoproc

Yes, I doubt the 100 Harvards/Berkeleys in China, albeit low-tuition and state-supported, would be much for soul building. In fact they would probably have ruthless test-based admissions policies and emphasize STEM subjects.

2015-08-21 on infoproc

> the unregulated market squeeze the life out of “the pursuit of knowledge for knowledge’s sake” < Not really sure this is happening to the extent that WD claims. Also, I would differentiate between pursuit of (e.g., scientific) knowledge and what he calls "soul building" ... States have been cutting support for higher education for 20+ years, and most of the increased burden on families is a result of this. I agree with WD that this is very bad. But the decline of the pursuit of soul building (or navel gazing 😉 by students is mainly due to societal/cultural forces outside the university (i.e., income inequality makes middle class aspirations much more tenuous) -- it's not really something professors can do much about. Generally I find people who throw the term "neoliberal" around do not have reasonable views about markets and capitalism, but ymmv.

2015-08-21 on infoproc

Yes, “fiscal situation” refers not just to finances but also attitudes of state legislatures, etc.

2015-08-21 on infoproc

It is the traditional role of scholars, professors, universities to fight on behalf of the pursuit of knowledge for knowledge’s sake. In this I endorse WD’s exhortations. But I don’t share WD’s disdain for “neoliberalism” — markets are useful, if imperfect, tools.

2015-08-21 on infoproc

I thought WD held his own on the panel. He and Homi were quite entertaining. I also support his 100 Berkeleys comment, although it’s pretty unrealistic here in the current fiscal situation. (Note the 100 Harvards in China *would* be low-tuition, state supported schools.)

“I think the more interesting story is the ways in which we are impotent to solve that problem.” sums it up pretty well.

Note that Steve Pinker was persona non grata at this event, even though he and Brooks were the most prominent thinkers to directly engage WD’s thesis.

2015-08-14 on infoproc

Not to mention the ability to absorb that education 😉

2015-08-13 on infoproc

There don’t seem to be diminishing returns at the +(4-5) SD (or whatever it is) tail. The linear model predicts a Gaussian distribution but the tails seem to be fatter than they should be, which is the opposite of diminishing returns.

My guess is that people who join high IQ societies are a selected population. Individuals who find a productive use for their brainpower are unlikely to seek out such groups. They probably get enough stimulation from their colleagues at the university or research lab.…

2015-08-12 on infoproc

I don’t think we know enough to answer these questions, but consider that much of the stress associated with being a high g person in today’s society comes from the gap between the exceptional person and the rest of society. It would be stressful to be the only modern living among Neanderthals. A community of exceptional people might operate under different rules and individuals who can’t survive in ordinary society today might thrive in the new community. Many kids are quite relieved to arrive on campus at MIT/Caltech as freshmen 😉

From the crude variables we can measure, increased g is mainly adaptive — to income, wealth, longevity, etc. If there are problems they are in the far tail, but then such individuals are very different from the people around them.

If most parents are using IVF/PGD just to eliminate negative g fluctuations or make small gains (no fitness problem with that), this will still have the effect of raising the average, and the consequent population of the + tail.

2015-08-12 on infoproc

>> pretty sure that caltech loses out in cross-admit battles to those schools.

This tends to vary over time.…

2015-08-12 on infoproc

You could have made the same argument regarding traits which agricultural breeders pushed for. Additive variance in oil content of corn has persisted as the trait shifted 30 SD over hundreds of generations of selection. Agricultural chickens lay roughly an egg per day, whereas in the wild it is more like one per month. Again, many SD of change achieved using *existing variance* (existing variants of genes; no new mutations — just flip (-) to (+)) in the ancestral population. We don’t know for sure what will happen with humans, but it’s a safe bet that +7 SD is not the limit. Just ask Ook the Neanderthal.

Re: head/brain size, yes there is a correlation with cognitive ability, but not all historical geniuses had exceptionally large heads, which suggests that there are other ways to get higher ability than simply a volume increase. Similarly, physical strength is associated with size but you can get stronger creatures without making them larger.

2015-08-11 on infoproc

CHH are talking about variants of large effect. Whether or not CHH alleles are *part* of the story of AJ intelligence, the bulk of any group difference in a quantitative trait is likely to be just frequency differences in variants of small effect, as discussed in the post above. Recent N-S European height results support this picture: N. Euros have higher frequencies of (+) variants across the board (the result of selection pressure from some, as yet unknown, source; this could have been exerted in the N, S or both).

If you do the math there are many tens, possibly 100, SDs up for grabs so even if we can’t reach the full potential in these numbers it seems almost certain to me we can get beyond the +5 or +6 of historical geniuses (the biggest of the 1957 chickens on the left ;-). Some of these historical figures were fairly normal in health, personality, appearance, so it’s almost obvious that by further tweaking one could go beyond them while retaining viability. The genomic space is very high dimensional, so it would be surprising if there were some kind of absolute barrier.

Think of a group of Neanderthals debating whether there could be a humanoid intelligence beyond that of Ook, the cleverest of them all, inventor of the obsidian chipping tool. Ridiculous.

2015-08-11 on infoproc

This is on the list because there is an Anne Roe file in the archive. She was not among the 64 scientists studied in her book!

2015-08-11 on infoproc

Perhaps one of the benefits of watching it during a long Pacific flight was having my brain partially off. I don’t think you can analyze any Disney movie for internal consistency as if it were a “hard sci-fi” novel. I liked the characters, effects, and acting.

2015-08-09 on infoproc

Nevertheless, talk to your Android device and tell me you aren’t impressed by the level of speech recognition and natural language processing. It’s not general purpose AI, but has (at minimum) vast economic/social consequences.

In my view, we can expect big advances in focused machine intelligence applications, but a long wait for human-level general AI.

2015-08-09 on infoproc

No, I expect marvels when a million genotypes with good phenotypes are analyzed using *our methods*. Currently, most samples are in silos due to limitations of subject consent agreements that prevent data sharing between different institutions. Simple regression (standard GWAS) can be performed using summary statistics, which can be shared across silos. But more sophisticated methods require having all the data together at once. It will take some time for such samples to accumulate.

If you read my recent slides to the end you will see a plea to NIH to require that funded studies use consent forms which allow better data aggregation. Even a small step like this is difficult.

On the positive side, HLI intends to get to a million genomes in ~4 years and will be able to perform our analysis. There are other efforts, public and private, which should reach fruition on similar timescales. Note that genotyping costs alone for a million individuals are ~$100M for SNPs and $1B for whole genomes.

2015-08-09 on infoproc…

It looks like about 40% of H students are in Science + Engineering, and another almost 40% in Social Science which is National Academy and possibly Nobel eligible (econ). H has a lot more premeds than Caltech, which produces very few students eligible for the Institute of Medicine. There’s also the literature Nobel — not sure whether that is counted here.

There is definitely a correction factor to be applied for H, although probably less of one for MIT.

The Caltech vs Harvard-MIT stuff is just nerd bragging, but I think the likelihood ratios relative to less selective (but still very good — i.e. flagships) universities are much more significant in their implications. If these schools really produce top scientists and engineers at 100x or even 1000x the rate of ordinary universities, there has to be an explanation… (i.e., psychometrics works! 😉

2015-08-08 on infoproc

You have to read the papers, understand the reasoning, and follow the experimental progress. There’s no shortcut.

If someone does lasso/CS on ~1M (genotype|phenotype pairs) for a trait with h2 ~ 0.5 and can’t get good results (capturing much of the heritability in predictive variance), then there is something wrong. It could be that s >> 10k, or architecture is not approximately additive, etc. If you actually understand the theoretical claims it will be obvious eventually whether they are correct (working) or not.

2015-08-08 on infoproc

In biological science no one cares about theory. A prediction about performance of an algorithm is of little interest, compared to whether one actually obtains working predictive models (whose validity can be easily checked on small test samples and through a process administered by a neutral third party). Of course, in physics or mathematics a lot of credit can be claimed for predicting something like a key data threshold for some method to work. But biologists could care less.

Aside from performance of convex algorithms, I’ve predicted that human cognitive ability has a largely additive genetic structure and sparsity of order 10k. This implies that many, many (~100) SDs of improvement are up for grabs. The consequences of this prediction are much more significant!

2015-07-28 on infoproc

Your summary is correct, AFAIU. Below is a bit more detail about the 4 gamete test, which differentiates between a recombination event (which breaks the haploblock for descendants of that individual; recombination = scrambling due to sexual reproduction) and a simple mutation at that locus. The goal is to impute identical blocks of DNA that are tagged by SNPs on standard chips.

Algorithm to generate haploSNPs

… Given two alleles at the haploSNPs and two at the mismatch SNP, a
maximum of four possible allelic combinations can be observed. If all four combinations
are observed, this indicates that a recombination event is required to explain the
mismatch, and the haploSNP will be terminated. If, however, only three combinations are
observed, the mismatch may be explained by a mutation on the shared haplotype
background. These mismatches are ignored and the haploSNP is extended further. We
note that this approach can produce a very large number of haploSNPs and very long
haploSNPs that could tag signals of cryptic relatedness. …

>> This estimated heritability is much closer to the full-strength twin study estimates, showing that a lot of the ‘missing’ heritability is lurking in the rarer SNPs << This was already suspected by some researchers (including me), but the haploSNP results provide support for the hypothesis. It means that, e.g., with whole genomes we could potentially recover nearly all the predictive power implied by classical h2 estimates ...

2015-07-26 on infoproc

If frequency is too low then in a given sample of individuals you won’t have very many people who have the variant, so your ability to say something about it is limited. Of course if the effect size is large enough (i.e., makes people +1 foot taller) you might detect it from just a few cases. So you can see these things trade off against each other.

2015-07-23 on infoproc

From: Kenneth W Ford. “Building the H Bomb: A Personal History.”

“What then happens when the fission bomb “match” explodes? Its radiation runs out ahead of its expanding material and almost instantaneously vaporizes the polyethylene, creating a very hot plasma. The pressure of this plasma adds to the pressure of the radiation, pushing outward on the outer steel cylinder and at the same time inward on the uranium cylinder, thereby keeping the channel open long enough to let more radiation stream in. At the same time, the outer layers of the inner cylinder “ablate” (boil off), creating even more inward pressure. Before long—within microseconds—an inwardly imploding shock wave has compressed and heated the cylindrical container of deuterium and also, through compression, caused the plutonium sparkplug to go critical. The sparkplug, which is really another fission bomb, helps to ignite and then enhance the deuterium burning—hence the name “sparkplug.” Finally, energetic neutrons emitted in the thermonuclear burning (especially those of 14 MeV from the DT reaction) cause fission in the ordinary uranium that surrounds the deuterium as well as more fission in the sparkplug. Altogether quite a maelstrom. So this “H bomb” is really one part fusion and three parts fission. Its total energy release (its “yield”) is estimated, in fact, to have come about three-quarters from fission and one-quarter from fusion.”

I think subsequent designs get a larger fraction of their energy from fusion than the original Teller-Ulam version.

See also…

2015-07-21 on infoproc

1. Perhaps it depends on the trait?
2. Perhaps it is a typo?

2015-07-21 on infoproc

>> I don’t think that the history of genetic associations that you present to Ioannidis is correct. I believe that there was a period of GWAS with correction for multiple comparisons that still failed to replicate before the current period of replication. << I'm not aware of this history you reference, but I am only a recent entrant into this field. On the other hand Ioannidis is both a long time genomics researcher and someone who does meta-research on science, so he should know. He may have even written a paper on this subject -- I seem to recall he had hard numbers on the rate of replication of candidate gene studies and claimed it was in the low percents. BTW, this result shows that the vaunted intuition of biomedical types about "how things really work" in the human body is worth very little. We are much better off, in my opinion, relying on machine learning methods and brute force statistical power than priors based on, e.g., knowledge of biochemical pathways or cartoon models of cell function. (Even though such things are sometimes deemed sufficient to raise ~$100m in biotech investment!) This situation may change in the future but the record from the first decade of the 21st century is there for any serious scholar of the scientific method to study. Both Ioannidis and I (through separate and independent analyses) feel that modern genomics is a good example of biomedical science that (now) actually works and produces results that replicate with relatively high confidence. It should be a model for other areas ... Re: particle physics, AFAIK the p value threshold used at, e.g., LHC is (1) not based on any rigorous calculation, and (2) not just taking into account multiple tests. There are always hard-to-estimate systematic errors (for example, in monte carlo of background events using theoretical approximations) that cannot be taken precisely into account. So 5 SD is really a rule of thumb. Even a 6 SD result could be wrong due to a subtle mistake by the experimenters. Science is hard!

2015-07-07 on infoproc

Yes, this is evidence for dominance, which is a nonlinear effect. For typical traits we expect most of the genetic variance to be linear, with a smaller nonlinear part.

2015-07-03 on infoproc

They discuss this a little bit in the paper, but you’re right that it’s not completely excluded as a confound.


Furthermore, humans also mate assortatively on the basis of body mass index, for
which we see no effect. A more complex possibility, a form of reverse
causality, could arise when subjects from one trait extreme (for
example, people with high educational attainment) are on average
more geographically mobile, and thus have less homozygous offspring,
with those offspring in turn inheriting the trait extreme concerned15.
We do not think that this mechanism can account for our results, since
it does not readily explain the constancy of our results under different
models, especially the similarity in bFROH for either more or less homozygous
populations. Moreover, we observe similar effects in multiple
single-village cohorts, and the Amish and Hutterites, where there is no
geographic structure and/or no sampling of immigrants, hence such
confounding by differential migration cannot occur. …

2015-06-17 on infoproc

Lots of mysteries here to explore…

2015-06-16 on infoproc

We are focused on the gauge-side explanation for how the thermalization can be so fast. We believe we have identified the mechanism, which connects to work on entanglement and thermalization (coming from quantum information theory and foundations of quantum stat mech). The evolution of a quantum pure state to “typicality” can be surprisingly fast when considered in the full Hilbert space. The usual kinetic theory picture of approach to thermal equilibrium assumes incoherent scattering of individual particles, and requires more time.

The connection to holography is through the Ryu-Takayanagi observation that the entanglement entropy of a boundary state (in this case the HIC state) is related to the area of extremal surfaces in the bulk which terminate on the border separating the boundary state into disjoint regions (one of which is traced over to obtain the entanglement entropy S_A; in our case this border separates the central region A of the collision from the rest of it, B). When a black hole is formed it forces a jump in the area of the extremal surface and through R-T this is directly related to a jump in S_A. Thus our physical observation that there is a rapid jump in S_A due to coherent scattering of the complex nuclei corresponds to rapid horizon formation in gravity.

People who believe strongly in gauge-gravity duality should have asked already what, exactly, is the entanglement entropy relevant to the heavy ion case? The gravity side predicts a jump in this quantity in the collision, but what is it? (In some cases of thermalization or quenches of condensed matter systems or conformal field theories S_A is known explicitly, but not in HIC.) One could just say that we have clarified this question and explained the mechanism for the rapid growth of S_A.

2015-06-11 on infoproc

You just need a sufficient overall correlation between edu years and cognitive ability in order to use it to find cognitive hits. Of course some of the hits could be on something other than cognitive ability, such as Conscientiousness, but you follow up by testing your hits against samples for which you have direct cognitive measurements.

2015-06-05 on forexiv

Jess, I’m not the only one who thinks additive heritability constitutes much or
most of total heritability for many traits. Just ask plant and animal
breeders/geneticists who have been working on this sort of thing much
longer than human geneticists. It’s not ugly physicists here, it’s one
biological field being ignorant of results in another proximate area.……

Of course, physicists might be different from other scientists in that they can often read (with comprehension) papers across multiple disciplines …

2015-05-24 on infoproc

I’m not sure what game theory actually accomplishes in the real world:…

But presumably vN was not of this mind at the time (cf his book with Morganstern).

2015-05-22 on infoproc

> I worry about naive use of purely additive models <

2015-05-09 on infoproc

Most of what I have heard from the current Google HR guy seems implausible to me.

2015-05-09 on infoproc

Personality traits are harder to measure than cognitive ability. If one properly accounts for the additional noisiness in the measurements of personality traits, I believe the resulting heritability is similar to that for cognitive ability.

2015-05-06 on infoproc


I don’t know for sure — let’s see what the Science paper says when it comes out.

2015-04-29 on infoproc

To be honest differences between the top few schools in this ranking are likely noise dominated. Do we really know mid-career median salary to within $5k for MIT vs Caltech? …

2015-04-29 on infoproc

They regress against student test scores and family income and use them in the predictor model. Schools are measured relative to the *predicted level* of, e.g., salary. Note there are ceiling effects, so Caltech kids might be better than can be captured from the SAT. But there are obviously other schools with good students (HYP etc. noticeably absent from top of the rankings) in the study, so they can easily build a predictor model that extends to the tails.

>> Sounds like you say something quite a bit different in the linked post. 🙂 << ?Not sure what confuses you. Is it the normalization by alumni population?

2015-04-26 on infoproc

Unz has replied to these claims in detail. I’m sure you can find his rejoinder easily via Google.

2015-04-15 on infoproc

Thanks! Added to main post.

2015-04-14 on infoproc

I know the authors of this paper and they sent it to me last night. Please post links to the other studies — I’m all for balance.

2015-04-12 on infoproc

If I had time I would look into this more seriously. Thompson’s collaboration has ~21k individuals so if these ML techniques work at all they should be able to build a pretty good predictor.

2015-04-12 on infoproc

Yes, it’s a crude method. But, to rebut far left IQ skeptics it is useful to show that these “socially constructed” test results are correlated to (indeed, can be predicted by) measurements of physical quantities — in this case, brain region volumes. The Thompson lab results indicate that brain structure itself is heritable.

2015-04-09 on infoproc

In cases of real deprivation there are studies claiming a big drop in heritability, although I don’t think the subjects were followed to adulthood. (Early environmental effects tend to fade away.) The PNAS paper I linked to in the post describes a Swedish study where adoption from lowest quintile to highest leads to a gain of ~ 1/3 SD.

2015-04-04 on infoproc

If the variable plotted is family income then one has to apply a further adjustment since some of the income is likely due to the spouse. (Although maybe the husband should get some credit for managing to marry a good earner …)

2015-03-30 on infoproc

The quote about Ballmer you are referring to is not from me. I did not attend HBS.

The quote is from a comment on EY’s LessWrong post.

2015-03-30 on infoproc

Excellent meeting. Would definitely attend again. Attendees are very open to learning and discussion, which makes it different from other meetings of similar exclusivity. Discussions are all moderated, which improves quality significantly.

2015-03-27 on infoproc

Meetings like these are actually very valuable, not just for networking but for learning what is going on in the world. I don’t suppose you know who Peter Thiel is?

2015-03-26 on infoproc

If only 5 or 10% of bio majors become doctors/dentists, the effect on average income is not that significant.

2015-03-23 on infoproc

“These schools simply select for high general ability” — If high general ability is the whole story, then why is Harvard so dominant over some of the other schools (e.g., YPS) that also get very strong students? I doubt any other school comes close to matching the statistic below, which seems to suggest a special institutional component:

>> more than ten percent of all Senators, billionaires, Federal judges, and Fortune 500 CEOs hold Harvard degrees of some kind <<

2015-03-21 on infoproc

PGD = preimplantation genetic diagnosis

2015-03-21 on infoproc

Yes, I think direct editing is not as imminent as the current fuss would suggest. However, IVF has become widespread since the first “test tube baby” in 1978. Genomic selection as part of the IVF process is already possible.

2015-03-17 on infoproc

PS I would rather have the data displayed with SAT score as horizontal axis. I suspect we would see stronger correlations. All engineering programs are selective, regardless of the overall admit rate of the university. Just being able to do the work and graduate is a pretty decent filter.

2015-03-17 on infoproc

If a student who attended an expensive, highly selective college gets the same ROI as someone who attended an inexpensive, not-so-selective school, then the former must have earned more money during the ensuing 20 years. (ROI is computed on the money spent on higher education.) However, whether this has anything to do with the school itself, as opposed to the higher ability level of its students, is not clear from this data.

2015-03-16 on infoproc

On average, costs rise with selectivity. So constant ROI as selectivity approaches its maximum (elite privates) suggests that incomes are also rising with selectivity. I also don’t understand why there is no data for highly selective non-quant majors. “Science” is a mixed bag, ranging from phys and chem (small fraction of total majors), to biology and other life sciences that are not (on average) well remunerated.

2015-03-13 on infoproc

Note use of “typical” 🙂

“A typical human behavioral trait is associated with very many genetic variants, each of which accounts for a very small percentage of the behavioral variability.”

They could also add a qualifier about “normal range of behavior” to exclude (rare) mutations of large effect.

2015-03-07 on infoproc

Like everyone else I find that there are not enough hours in a day to do all the things I want! Unfortunately I’m not one of those people who can get by on little sleep. I perform best on 8-9 hrs per night.

2015-02-27 on infoproc

Thanks, and good luck with your blog!

2015-02-26 on infoproc

IIRC from talking to Schmidhuber some of their techniques for determining NN weights exploit high dimensional sparsity. But it’s probably worth thinking about more.

2015-02-24 on infoproc

Somehow the cuteness of the lightbot attracts my kids. The puzzles and solutions have a certain elegance. They can be challenging, esp. with memory/command limits.

The book you recommended looks good!

2015-02-22 on infoproc

Thanks! Great recommendation. Right now my kids are enjoying driving bots around the screen 🙂

2015-02-18 on infoproc

Isn’t 70 grad students on the small size for most CS departments? Even if CS is now the most popular major, Caltech can’t have more than ~200 undergrads total in the program. The UG student body (and almost every department, in number of faculty relative to competitors) is quite small. (Smaller than most high schools!)

2015-02-12 on infoproc

He’s just being poetic, or narrowly defining history as the story of the original homo sapiens.

2015-02-07 on infoproc

It’s possible. Only time will tell.

2015-02-02 on infoproc

With lending and derivative contracts the effective money supply isn’t the same as the number of coins in circulation.

2015-02-02 on infoproc

The central bank doesn’t have to peg the coin price to the cost of blockchain verification.

2015-01-27 on infoproc

The cosmological constant pulls everything apart. So the attractive force needed to keep a satellite in orbit is increased slightly. The effect does not violate isotropy. The existence of the central mass M violates isotropy.

2015-01-22 on infoproc

“none of the people interviewed had studied astronomy, geography or meteorology”

How about elementary school or middle school science? Do you think they were exposed to that? It seems likely that all had been exposed to the idea that the Earth’s axis is tilted relative to the orbit plane, etc. Whether they could recall this *geometrical* picture and its implications does have something to do with math/spatial ability. They might also have wondered why their elliptical orbit model doesn’t explain variation in day length.

If the film makers had tried something similar down the road at MIT graduation, would they have had a harder time finding clueless but confident grads?

2015-01-20 on infoproc

“However, when the individual student is the unit of analysis, multiple-choice measures are known to yield more reliable scores per hour of testing time than do constructed-response measures.”

Don’t know of specific studies, but this seems very plausible. See, for example, the writing portion of the SAT.

2015-01-14 on infoproc

For many traits additive heritability is > 0.5 and broad sense h2 > 0.7 or so. Therefore you can build good quantitative predictors just treating environment as noise. The problem is to extract the parameters in the models (see Turing quote).……

2015-01-07 on infoproc

Looks like we are in for a long conversation 🙂 I think I know where we disagree, but it will take some time to write it up in comprehensible form — I’ll send you an email. I don’t think you need “out of equilibrium detectors” for the system to evolve away from the stationary state you consider. The state of the universe we describe (QFT modes thermalized; gravitational back reaction changes geometry slightly) has higher entropy than the stationary one (yours), and there are nonzero quantum amplitudes connecting the two. So in an Everett description I think most of the measure (probability) is concentrated on the former. More later …

PS I noticed in the original Gibbons-Hawking paper on dS temperature that Hawking promises to address some of the issues you tackle in your paper in a forthcoming publication (ref. 16; see discussion in left column of p.2751). AFAIK ref. 16 never appeared! So it’s definitely a tricky question …

G. W. Gibbons and S. W. Hawking, Phys. Rev. D 15, 2738 (1977)

2015-01-06 on infoproc

I believe the decline in state support for higher education in the US is mainly driven by budget pressures — specifically, health care and other costs. The availability of student loans has cushioned the problem (from the perspective of universities), but is not the cause.…

“Education used to make up a bigger share of state spending. When the association first began compiling the report in 1987, elementary and secondary education made up the biggest share of state spending, and higher education the second-biggest share. Medicaid surpassed higher education as the second-biggest state program in 1990, and in 2003 it became largest state program for the first time.”

2014-12-14 on infoproc

Not sure — implantation rate might refer to individuals (60 women had an implantation cycle) whereas pregnancy refers to transferred eggs (80 eggs transferred)? I think there are more details in the paper itself.

2014-12-04 on infoproc

von Neumann to Nash: “That’s trivial, you know. That’s just a fixed point theorem.”

2014-12-03 on infoproc

I think strong privates inhibited development of strong publics back east. This is changing now, however. Rutgers and Maryland are now in B1G and academically they are roughly comparable to other schools in the conference.

2014-11-22 on infoproc

See top right figure.

2014-11-09 on infoproc

It was very off-topic.

I’m not in the Q&A business, but if you send me an email containing your real identity I will try to find the time to reply. Note, some of the questions you asked are not really in my area of expertise.

2014-11-07 on infoproc

I think Munger just gave $65M to KITP at UCSB! Caltech is screwing up again if they don’t get a much bigger gift from him, a former student who lives in Pasadena.

2014-10-31 on infoproc

Players can develop a lot between the age of 17 and 22. Some blossom physically, get really good technique coaching; others not so much.

By the time the NFL combine and scouting process take place, the rank order of players in a cohort may have changed considerably.

The same could be said for 13 year old kids growing up to become scientists or technologists.

Talent selection is noisy, but signal is not zero…

2014-10-21 on infoproc

We don’t know much more about growth and metabolism, but nevertheless can produce bigger and faster growing chickens and other agricultural animals. You can do engineering without understanding fundamental principles 😉

2014-10-20 on infoproc

Tom: Thanks for the offer — probably too busy on this trip, but maybe someday! Good luck with fusion; we need it 🙂

2014-10-16 on nautilusmag

I just realized I’ve said all of this already in… (p.16):

… The preceding discussion is not intended to convey an overly simplistic view of genetics or systems biology. Complex nonlinear genetic systems certainly exist and are realized in every organism. However, quantitative differences between individuals within a species may be largely due to independent linear effects of specific genetic variants. As noted, linear effects are the most readily evolvable in response to selection, whereas nonlinear gadgets are more likely to be fragile to small changes. (Evolutionary adaptations requiring significant changes to nonlinear gadgets are improbable and therefore require exponentially more time than simple adjustment of frequencies of alleles of linear effect.) One might say that to first approximation, Biology = linear combinations of nonlinear gadgets, and most of the variation between individuals is in the (linear) way gadgets are combined, rather than in the realization of different gadgets in different individuals.

Linear models work well in practice, allowing, for example, SNP-based prediction of quantitative traits (milk yield, fat and protein content, productive life, etc.) in dairy cattle. …

2014-10-16 on nautilusmag

The question of additivity of genetic effects is discussed in more detail in reference [1] above (sections 3.1 and also 4):…

In plant and animal genetics it is well established that the majority of phenotype variance (in complex traits) which is under genetic control is additive. (Linear models work well in species ranging from corn to cows; cattle breeding is now done using SNP genotypes and linear models to estimate phenotypes.) There are also direct estimates of the additive / non-additive components of variance for human height and IQ, from twin and sibling studies. Again, the conclusion is the majority of variance is due to additive effects.

There is a deep evolutionary reason behind additivity: nonlinear mechanisms are fragile and often “break” due to DNA recombination in sexual reproduction. Effects which are only controlled by a single locus are more robustly passed on to offspring. Fisher’s fundamental theorem of natural selection says that the rate of change of fitness is controlled by additive variance in sexually reproducing species under relatively weak selection.

Many people confuse the following statements:

“The brain is complex and nonlinear and many genes interact in its construction and operation.”

“Differences in brain performance between *two individuals of the same species* must be due to nonlinear effects of genes.”

The first statement is true, but the second does not appear to be true across a range of species and quantitative traits.

Final technical comment: even the nonlinear part of the genetic architecture can be deduced using advanced methods in high dimensional statistics (see section 4.2 in [1] and also….

2014-10-10 on infoproc

Can’t be sure until we try. Also, note that due to crappy IRB rules it’s often not possible to aggregate data for a better analysis (one of the advantages of simple regression is that you can just share summary statistics across groups). But in our earlier paper we found that CS could find perhaps twice as many causal variants as simple regression with the usual genome wide significance threshold, with a small false positive rate. See figure.…

2014-10-08 on infoproc

You are confused about the direction of implication. The study suggests that not every gene or region of DNA affects inter-individual height variation. That is not the same as saying that the height causal variants only influence height.

Please don’t post random stuff in the comments.

2014-10-07 on infoproc

Some biologists (“real experts”!) thought/think the entire genome (or nearly every gene) is involved in complex traits like height. To me this just sounded crazy but the authors of this paper consider it an important point to refute. To them “finite” means less than the whole genome.

It’s amazing to me how *little* intuition most people have for how a short program could produce a complex organism and yet be modular enough to be evolvable and account for many different types of organisms. A lot of priors held by biologists about genetics are implausible to me. Even now most can only think about simple Mendelian traits.

2014-10-07 on infoproc

Only the top ~ 697 causal variants have been reliably identified. At higher p values there are many false positives (not genome-wide significant: p < 5E-08, of these there are only 697). The top 10k in p values they refer to in the paper contain some fraction of height SNPs, but most are probably not and just contribute noise to phenotype prediction. If larger sample size were available one could find the *actual* top 10k variants and these would probably account for the bulk of the heritability. This would require sample size ~ million or more and better analysis techniques (e.g., compressed sensing) than what they are currently using.

2014-10-04 on infoproc

I did attend Merton’s class on continuous time finance (option pricing theory) and I also once dated an HBS girl. I could see their campus across the Charles from the window of my Dunster House apartment.

2014-10-04 on infoproc

Note I didn’t write that excerpt — I’m not an HBS grad!

But I imagine he may have had IP/patent law in mind.

2014-10-01 on infoproc

I didn’t make any timescale predictions. The directionality of what is discussed in the video is of course correct.

The horse analogy is a good one. I believe we’ll see a big divergence in economic value of skilled vs unskilled human labor in the coming decades.

2014-09-16 on infoproc…

“Qualifying for the USAMO is considered one of the most prestigious awards for high school students in the United States, with only 264 students qualifying in 2013 out of over 350,000 students competing.[1] Top scorers on the USAMO are invited to the Mathematical Olympiad Summer Program and represent the United States at the International Mathematical Olympiad.” (Keep in mind even the 350k are a selected population.)

Nope, can’t imagine how you’d find two hundred off-scale kids each year without using the old SAT. Nearly impossible …

2014-09-16 on infoproc

I thought I answered this in the main post. There are lots of indicators with higher ceiling than the current SAT.

1. Max out SAT at earlier age (like SMPY)
2. Perform well on really hard competition exams like for USAMO or IPhO qualification
3. Science, coding, number theory, etc. projects
4. Explain how Einstein deduced the Lorentz transformation to a Caltech prof interviewing you

Do you know how Intel/Westinghouse pick their winner from ~40 finalists? It’s not the project. It’s how smart the kid appears to be in interviews.

Just imagine your goal is to pick out the +4 SD kids from a bunch of > +3 SD kids. There are signals that are hard to fake … Admittedly, ALL the signals we are talking about are noisy. But I’m not sure why you can’t believe there are people out there who care about differences in the tail. In my field, choosing graduate students is all about that difference.

How do you think a college coach can tell the difference between a Division I player and an All-State HS player who is only Division II or III caliber? They all seem like great athletes to me …

PS I’ll admit I don’t know how to differentiate within this pool among non-STEM types. But that’s not Caltech’s problem … I’ll assume H uses the same voodoo magic that they use in holistic scoring.

2014-09-15 on infoproc

If you parse Pinker carefully you will see he did not make a mistake. I think he wrote carefully that BEYOND the 5-10% (S category) Harvard STARTS to assign nonzero weight to other factors like athletics or musical ability. He is definitely talking about S, as am I.

“… it’s common knowledge that Harvard selects at most 10 percent (some say 5 percent) of its students on the basis of academic merit. … The rest are selected “holistically,” based ALSO on participation in athletics, the arts, charity, activism, travel, and, we inferred (Not in front of the children!), race, donations, and legacy status (since anything can be hidden behind the holistic fig leaf).” (emphasis mine)

What David doesn’t seem to understand is that starting to assign nonzero weights to other factors could still lead to a top 25% of the class (400 kids) at or near the SAT ceiling. There are a lot of smart kids who 1. (nearly) hit the SAT ceiling, 2. but are not good enough to qualify for an S category with, e.g., 10% = 160 slots, 3. but who also are good at sports/music/leadership and hence get the nod over a more impressive academic kid who lacks the other factors.

Given its current cachet (probably wins cross-admit battles with every other school at > 70% probability, except possibly Stanford?), Harvard could definitely take more smart kids than it currently does IF IT WANTED TO. That is NOT to say that the current top 25% or even 50% of each Harvard class isn’t an impressive group — they are!

2014-09-15 on infoproc

Both 75th percentile distributions are “saturated” within noise at the ceiling. (This ceiling is *below* the 10% S threshold.) You need to look at other indicators.

I gave you an actual algorithm, used by H in the past (and perhaps still today), that could explain the same observation about 75th percentile scores, without making the S category 1/4 of the entering class.

Having said all that, it is possible that H is putting more emphasis on academic ability for, e.g., the top 1/4 of their class than they did in the past. The main change is that they effectively have first call on a good portion (most?) of the top thousand+ students in the US each year. It pains me to write that, but it’s probably true!

These numbers are out of date, but perhaps still of interest:…

2014-09-15 on infoproc

Most psychometricians (and even many admissions folks) discount the writing portion (score is too noisy, not g loaded) and only look at M+CR.

I’m making the point that you can’t just look at SAT scores to see how big the S category is. There are other, better, indicators.

2014-09-10 on infoproc

Oops, that’s a typo — should be 200 not 2000.

2014-08-23 on infoproc

There were other potentially controversial topics.

2014-08-20 on infoproc

Thanks, James. The summary looks fine. I perhaps should point out that the toy model with 10k causal variants could turn out to be wrong. But it does illustrate some basic points.

2014-08-19 on infoproc

There could be tradeoffs, and the amount of pleiotropy is an open question.

However, I would bet that in a 10k-dimensional space there are directions that, e.g., increase one or a few traits radically without causing problems in others. Certainly among humans that have already existed there are examples of individuals with very strong characteristics in almost all categories. The “maximal type” according to any chosen combination of characteristics almost certainly exceeds types already produced by random chance.

2014-08-15 on infoproc

A mutation in the LDL receptor-related protein 5 gene results in the autosomal dominant high-bone-mass trait

DOI: 10.1086/338450

… In an effort to identify genetic factors influencing bone density, we characterized a family that includes individuals who possess exceptionally dense bones but are otherwise phenotypically normal. This high-bone-mass trait (HBM) was originally localized by linkage analysis to chromosome 11q12-13. We refined the interval by extending the pedigree and genotyping additional markers. A systematic search for mutations that segregated with the HBM phenotype uncovered an amino acid change, in a predicted beta -propeller module of the low-density lipoprotein receptor-related protein 5 (LRP5), that results in the HBM phenotype. During analysis of > 1,000 individuals, this mutation was observed only in affected individuals from the HBM kindred.

This paper has ~1000 citations.

2014-08-15 on infoproc

Caplan and Russ Roberts (who hosts Econtalk) are not the same person. In fact they are quite different in beliefs and interests.

2014-08-14 on infoproc

I don’t know the details. It has been investigated in mouse models already (see link), but it was described to me at the meeting as originally discovered in humans in an African village. Locals hit by cars simply got up and walked away! I don’t know whether the general health or development of these villagers is affected by the mutation.

2014-08-14 on infoproc

The IVF counselor could turn over all the information about each zygote to the parents and let them choose.

You could also imagine that the parents fill out a survey about their preferences. This defines a kind of utility function (e.g., linear combination of traits with certain weights) which is then used to score the zygotes from best to worst.

In any case, on average the parents won’t do worse than they would without any predictive information.

2014-08-13 on infoproc

Alpha children wear grey. They work much harder than we do, because they’re so
frightfully clever. I’m really awfully glad I’m a Beta, because I don’t work so hard.
And then we are much better than the Gammas and Deltas. Gammas are stupid.
– Brave New World, Aldous Huxley

2014-08-01 on infoproc

DK, you’re an honorable guy! When I have a moment I’ll write a blog post laying out my hypotheses on these issues…

2014-07-31 on infoproc

If it turns out that GCTA correctly predicts the heritability accounted for by common variants, and rare variants account for most of the rest (e.g., for height), will you admit that you and most of the field were clueless as late as 2014? Because you don’t seem to think this is plausible.

If you do think it’s plausible, then there’s no need for you to keep complaining about it on my blog.

2014-07-31 on infoproc

> No one with any clue < Lots of "leaders" in the field had no clue 10 years ago. Do you remember, e.g., DeCode's business plan in the early part of this century? Many in the field had their priors focused on small numbers of causal loci, and those priors turned out to be completely wrong. It's the same crowd within the field that are skeptical of GWAS/GCTA. You still see articles about "missing heritability", "dark matter of the genome", blah blah, whereas it's just lack of statistical power thus far. Estimates of number of causal variants for diseases and other complex traits have gone steadily upward over the last 10 years, TO THE SURPRISE OF MOST PEOPLE IN THE FIELD. Why the need to cry about "missing heritability" unless you thought we should have already found most of the answers at this level of statistical power? That would only be true if # of causal variants were small. People who came from the animal breeding population genetics side (i.e., quantitative genetics: Visscher, Goddard, et al.) had the right intuition for how things would turn out, but many others in the field were totally wrong about how genetic architectures would work. These people had never read Fisher or heard of his infinitesimal model http://rstb.royalsocietypub…

There’s an interesting historical question as to whether people were really that clueless, or were self-interested (deliberately deceptive) and advocated the one or few gene picture in order to make drug discovery more plausible (e.g., to investors). But after some investigation I am convinced a lot of “experts” at the time REALLY BELIEVED what they were saying (some still have not figured it out).

2014-07-30 on infoproc

If it’s highly polygenic then effect sizes of single variants will be small.

What’s at stake in understanding GWAS/GCTA and all that is the transformation of genetics into a mature statistics-driven field that attempts to understand complex (as opposed to simple Mendelian) traits. It’s not surprising that parts of the field are going to be left behind in all this. The strongly held priors people had 10 years ago (e.g., one gene, one disease, one drug) have already been shown to be completely wrong… How much credibility can non-quantitative types have left?

2014-07-30 on infoproc

You have to go with the study that has the most statistical power. The SSGAC educational attainment study that found the 3 hits had 100k+ individuals. One would not have expected them to be found by earlier studies (effect sizes too small). You can’t really think properly about this subject unless you understand statistics.

2014-07-29 on infoproc

Oops, sorry I didn’t realize the monograph might be different from the book. Might be worth looking into.

I’ve read the Cambridge paper and it’s not very enlightening. The scores seem very low to me.

2014-07-28 on infoproc

I think what we can safely conclude from Roe is that her top scientists were unusually smart compared to average scientists.

In the book, she says she normed her high ceiling tests on a population of Columbia Teacher’s College graduate students, who also took standard IQ tests. Once she knows the SD in that group she can use it to convert high end raw scores to z scores. Perhaps the student SD is smaller than the general population SD and she failed to adjust for that. But she had their scores on well normed standard tests, so she could and should have done so. She didn’t do the psychometric work herself — she worked with a specialist at Columbia.

Also, there is probably some deviation from normality at the high end.

2014-07-28 on infoproc

I have two posts on the blog about Roe’s results and the identities of the scientists she studied.

2014-07-28 on infoproc

IIRC two of the experimental physicists in her sample had pretty mediocre (~120) V scores. One of them could have been Alvarez. The V mean for all the scientists was very high – 166!

SMPY is a good modern update of Terman.

2014-07-28 on infoproc

? The fact that the V ceiling on SAT/GRE is higher than for M is the choice of the test makers. They could have done it the other way as well.

2014-07-22 on infoproc

I agree.

However, IIRC in their paper the amount of variation in overall cognitive ability explained by their “chimp g” was pretty low compared to human g. This might be an artifact of the battery of tests they use, but it does weaken their results a bit in my opinion.

2014-07-22 on infoproc

Deary’s paper puts an upper bound on the effect — it could still be there but not statistically significant for sample size in the low thousands. If only a fraction of rare variants are deleterious for g then the correlation of incidence of those variants with g may not be visible against the background: fluctuations in number of variants that don’t affect g. If, say, 5% of rare variants have negative impact on g, you have a pretty tough signal to noise ratio to overcome.

2014-07-20 on infoproc

Only a small fraction of variants will affect cognitive ability. You have to pick the signal out over a large background of variants that don’t affect g score.

2014-07-19 on infoproc

There are some big biobanks at Vandy, Kaiser-Permanente, etc. Some hospitals are routinely taking blood or other DNA samples from every patient that passes through.

Once the price point is low enough these samples will all be genotyped. Where, exactly, the break point is really depends on funding agencies. Even with the cheapest SNP chips a 100k sample takes ~$(5-10) million to do.

The real problem going forward will be good phenotyping and the ability to pool data. The latter is encumbered by IRB and lack of foresight in bio sample donor release forms.

2014-07-17 on infoproc

I think they are protecting Conor. A good wrestler would be a threat to decision him …

2014-07-13 on infoproc

I don’t have an exhaustive list, but MMA fighter, Gigolo, and Call Girl must be near the top 🙂

2014-07-11 on infoproc

Reliability is much lower when testing kids. But if they were very careful with the chimps perhaps they have a good measurement. Nevertheless, 0.5 is probably a lower bound.

2014-07-11 on infoproc

Thanks for the papers.

I would be shocked if you could do cognitive tests on a monkey with very high retest reliability. Holding their attention is a bit more difficult than with humans! This contributes more noise to the g measurement and decreases the deduced heritability.

2014-07-10 on infoproc

Lorentz invariance could be an emergent symmetry. It doesn’t require anything about discreteness at small scales — see the lattice (where Euclidean invariance is emergent at long distances), for example.

Obviously, there is a discrete version of calculus (see numerical analysis).

2014-07-09 on infoproc

I think the evidence is quite strong that there is a “minimal length” in spacetime. The continuum is an *idealization* that comports with our intuition because this minimal length (the Planck length) is so much smaller even than the size of atoms.…

2014-07-09 on infoproc

No actual calculation in physics requires infinity. The easiest way to see this is to note that calculations required to compare theory to experiment can be done on finite computing machines. The continuum is an idealization and we are not sure, due to quantum gravity, whether the structure of spacetime is actually continuous. (The same applies to quantum fields and even to Hilbert space.) Some *theories* of physics may invoke infinity, but careful consideration reveals that the necessity of infinity will never be *experimentally testable* (i.e., there are related theories which do not invoke infinity which cannot be excluded by experiment).

2014-07-06 on infoproc

Never attribute to conspiracy what could also be explained by stupidity … 🙂

2014-06-22 on infoproc

There’s a link to my slides on the post above. Just open the link in another window/tab and you can toggle back and forth to the video.

2014-06-22 on infoproc

Verifying approximate linearity is by itself a big step in understanding the genetic architecture. However, I would be shocked if it weren’t the case, based on the arguments I gave in the talk. If you listen carefully, someone from the SSGAC comments during my talk that they already have some empirical evidence for linearity; estimates of non-additive and additive variance can also be made from sibling/adoption studies.

Understanding actual molecular mechanisms for how the brain works will, as you note, be much harder and take much longer.

But plant and animal breeders also do not understand all the molecular mechanisms related to the phenotypes they select for. It does not prevent progress.

2014-06-22 on infoproc

Science works through the accumulation and careful examination of data. Large numbers of studies are usually required when dealing with complex social or biological phenomena, to which people (even scientists) bring strong priors or political agendas. I am trying to follow the results in this area to make up my own mind about what to believe. I think it is best to be epistemologically careful and not jump to conclusions.


Researchers are willing to put in years of work to obtain results like those in the PNAS paper above; are you saying I shouldn’t analyze them here? What would you tell the researchers — they should quit or bury their results?

You’ll note I have written hundreds of blog posts on finance (i.e., derivatives, the credit crisis, risk, etc.) and economics — more than the number on genetics (including on the technology of sequencing, cancer, evolution, medicine, etc.), and far more than on AAs. Should I quit after a few posts (because some reader is unhappy? my conclusions about, e.g., the banking system or the field of economics are often negative), or should I continue because there are new things to learn and think about in those areas?

2014-06-22 on infoproc

The authors of the study suggest that AA academic success could be all due to effort as opposed to cognitive advantage. This might even be the case — one has to do a careful study to find out. I don’t think the issue is fully resolved. Does that sound self-congratulatory?

Some of us want to know how the world actually works.

2014-06-21 on infoproc

Look more carefully at the figures.

2014-06-18 on infoproc

In fact if you read the paper their definition of academic achievement for higher grade levels is simply GPA and hence is very crude. Some students are taking more difficult courses than others …

This may help explain the inconsistencies in table S4 of regression coefficients among the different subgroups.

2014-06-15 on infoproc…

We’re living in a crazy world! 🙁

2014-06-08 on infoproc

Sinica is not an academic podcast. Talhelm, an innovative guy, has actually done something beyond academic research. This particular thing is of much interest to residents of Beijing 😎

2014-06-06 on infoproc

Mutations of large effect may yield clues about impairment of cognition (i.e., molecular mechanisms). That may help to identify variants of smaller negative effect (“nicks”) in the future.

2014-06-06 on infoproc

I agree with the part of Eisen’s review that you quoted. I think he’s a bit harsh on Wade in the remainder, but this is mainly a question of taste. How much speculation should someone be allowed, assuming they deliver explicit “THIS IS SPECULATION — NO CONCLUSIVE EVIDENCE YET” warnings to the reader?

2014-06-06 on infoproc

We play another hand — this adds excitement and can end the game quickly 🙂

2014-05-30 on infoproc

“having to deal with ground and pound for an extended period of time” Yes 🙂

We don’t know how good JJ’s guard is. Cormier might be able to pass guard and then JJ’s length could be a huge liability. Length will also make it harder for JJ to defend Cormier’s takedown once he gets close or inside.

2014-05-23 on infoproc

It would be much cheaper if produced in volume (no really expensive components). It could also be mounted on a small drone, resulting in a weapon that causes little collateral damage.

See also…

2014-05-19 on infoproc…

It all depends on selection pressure, assortative mating, etc. which no one can really calculate reliably in the deep past.

I would say ~1ky is very fast while ~10ky is easily plausible.

Poor Fisher, Kimura, Falconer, et al. So little of their theoretical work was testable. Only now is good data starting to become available — we’ll see big advances in our understanding of human evolution in the coming decades.

2014-05-18 on infoproc

I’m really not interested in arguing about the meaning of a term like “race” — it is used quite differently by cultural anthropologists, biologists, the man in the street, etc. Consequently it has become an impediment to understanding.

But we can still make well-defined statements such as “Allele frequencies tend to be more similar between two Nigerians than between a Nigerian and a Korean” or “The number of base pair differences between two Nigerians is less than between a Nigerian and a Korean”. These are both now known to be true, but that was not the case 30 or 50 years ago. If you want to argue with me about what I mean by Korean or Nigerian then you are really just diverting the conversation, because those terms are about as well defined as “cat” or “car”.

I don’t want to make any strong claims about the *consequences* of these observations until the science is on firmer footing. It’s an important first step to note that some claims about *genetic* level variation in humans (the Montagu/Lewontin-type claims still taught in university classes) turned out to be incorrect.

2014-05-18 on infoproc

Victor, you seem well-intentioned but confused about the sense in which individuals “cluster” into groups by genotype. These posts might help:……

Old ideas about mtDNA, matrilineal lines, haplotypes, etc. don’t capture the essence of the matter: there is a sense in which genome X is more similar to Y than to Z. Just count the number of places where the G,C,T,A base pairs differ. Visualize this geometrically by converting the number of differences into a Euclidean distance (in a very high dimensional space). Then you get clusters (see figures, etc.), and they correspond almost perfectly to self-identified race (Risch et al.)

Two individuals in the same cluster are *genotypically* more similar to each other than they are to individuals in other clusters. What this means in terms of *phenotype* is something we are still working out…

2014-05-15 on infoproc

“… the fundamental equalities of all groups and all sexes don’t depend on the results of such analyses”

I think he means that fundamental *rights* of individuals should not depend on the outcomes on group differences. I agree with this.

2014-05-15 on infoproc

unknown territory …

2014-05-15 on infoproc

1) Some alleles are (nearly) fixed differently (~ 100% frequency) in different populations. Wade gives some examples and discusses hair, skin color, etc. Buy the book! (Or use Google.)

2) Additive variance for cognitive ability is probably larger than non-additive. If I had to guess I’d say ~ 0.6 / 0.2 for additive and non-additive contributions to total h2 (error bars about 0.1 for each number). Given enough data we may be able to get the non-additive part as well (something I am working on 😉

2014-05-13 on infoproc

I don’t think SMPY was trying to be very careful in gauging income percentile vs general population. They just wanted to demonstrate differential outcome in their group. I’d be surprised if the top 1% in IQ had lower probability of being in the top 5% for income than a randomly selected individual. Have a look at the other links which regress income on IQ — the result is always positive, although the coefficient is not that large. Apparently, income returns to IQ are smaller than one might guess.

I think a big part of this is that really smart people are often more interested in learning, solving problems or creating new things than in maximizing their net worth. In my own case (and yours!) I am sure I could be significantly wealthier had I focused on money to the exclusion of other interests.

2014-05-12 on infoproc

These SMPY links might be useful:…

I think age adjustment could be pretty significant.

Good income–IQ data is hard to find:………

2014-05-12 on infoproc

Yes, it seems to me that in environments where food availability is unpredictable one successful adaptation would be to be on the small side, or to have the flexibility to develop in reaction to the availability of calories. If you believe at all in epigenetics you might guess that it could take several generations of good nutrition for a population to reach its full size potential. The information about food availability would be useful to pass on inter-generationally.

2014-05-12 on infoproc

They were not especially tall, historically.…

I actually know a lot of Dutch physicists, but no one seems to have a good explanation for recent height gains there. Theories abound: high quality of life, relatively low stress, lots of dairy, hormones in the dairy, etc.

2014-05-11 on infoproc

“I think it’s especially important to be epistemologically careful in thinking about these matters, because of our difficult history with race.” = higher standard of evidence.

Yes, it’s a double standard — but there are good reasons for it. IF the conversations were only among scientists we would not need this double standard.

2014-05-10 on infoproc

Hence “Possibly” 🙂

2014-05-10 on infoproc

Within the range of environments typically found in twin and adoption studies, No.

Outside this range? Possibly. See Turkheimer’s work (various posts on this blog) on decreased heritability due to deprivation.……

2014-05-10 on infoproc

I doubt even the most extreme cognitive elitist wants to eliminate everyone dumber than a Fields Medalist …

2014-05-10 on infoproc

Even if H2 is true population distributions will overlap significantly on every quantitative trait. If you set a threshold for the trait then there will be people of all ancestries above and below the line.

2014-05-09 on infoproc

>> how to account for the possibility of “real effect” alleles being part of the population PCs

You can always restrict analysis to a sample from only one group at a time. Results thus far show robust replication — size and sign of effect tends to replicate across populations. This is true for height, lipid levels, blood pressure, etc. Sorry I don’t have all the links handy. Note this is all supportive of an additive or linear genetic architecture.…

Once you know the full genetic architecture you can just compute population averages (“breeding values” in the plant/animal terminology) for each group. You might indeed find that frequencies of loci affecting a particular trait are correlated with population structure PCs.

See also:……

2014-05-09 on infoproc

Here’s an answer to your question from the discussion at Tyler Cowen’s blog (he reviewed the book). Note this is not my comment — however, it illustrates that the most interesting aspect of Wade’s book might be the reactions to it and the manner in which people position themselves.


1. You are not going to learn any new Science

2. You are going to learn what happens in your society when a distinguished and relatively prominent Science journalist publishes a prominent book in which he shows a bit of courage and gets as close as possible to promoting an unorthodox and taboo truth without risking utter ostracization.

3. You will learn who cannot risk publicly aligning with that position in order to maintain their position and current and future influence. And you will learn the techniques they must employ in order to walk the narrow path between sacrificing their integrity promoting the erroneous orthodoxy itself, and supporting the accurate contrarian position. Don’t hold anything against Prof. Cowen, he’s doing good work, but sometimes he writes a post the purpose of which is not to be a reflection of his genuine understanding or position, but, essentially, to allow Sailer to write in the comments section and do the actual updating of priors. Asking why people successfully avoid the subject and remain respectable by constantly talking about the Flynn Effect just might be relevant to this lesson.

Learning the topology of PC and influence in your society, and observing the consequences, is in fact very important. Reading the book itself will tell you whether the negative reviews are giving Wade a fair shake or not, and if they’re not, that’s revealing, and the answer to ‘why not’ is extremely enlightening. And also depressing. Learning how to achieve success in life by walking the line, not sacrificing your integrity, but leveraging your popularity, esteem, and status to occasionally promote truth-tellers, is also a very valuable thing to learn.

Another thing to learn is the answer to the question of, “What the point of Wade’s book if it has to be so mellow?”

The point is to very gently walk up to the question of the origin of disparities between human population groups (don’t get hung up on the semantics of ‘race’, just concentrate on genetic relatedness). Right now, the PC-orthodox theory of the origin those disparities is 100% discrimination, oppression, privilege, historical legacy, etc. The orthodoxy says that all human population groups are neurologically uniform in the distribution of various cognitive talents and abilities. That argues for both the necessity and moral imperative of even extremely obnoxious government interventions in countless circumstances involving personnel selection and redistribution of resources.

If, on the other hand, a large fraction of that disparity is fairly attributable to genetics instead of social injustice, then bigotry and discrimination is not a good explanation for the disparity, and thus the government crusade against discriminating employers and coercive disparate impact policies are unjustified. Also, if the ‘test score gap’ cannot be closed by any reasonable government policy, then we should stop slandering decent educators doing the best they can with the materials they have as ‘bad teachers’ who fill ‘bad schools’.

Indeed, if those who are influential and persuasive over the elites in the political class who craft policy could adopt even a 50/50 nature-nurture model of the origin population group disparity, then the implication is a complete upheaval and revolution in government policy, the positive benefits of which cannot be overstated.

As an opening salvo in that ‘So What?’ war, Wade’s cautious eggshell-walking, and Prof. Cowen’s snippy review, are unfortunate deviations from the ideal due the oppressive ideological environment, but they are nevertheless to be commended.

2014-05-09 on infoproc

Wade’s discussion is too primitive for this (circa 2002–2005 or so). You have to look in the literature. There are many papers analyzing large samples for population structure — unbeknownst to social science “experts” on “race and ethnicity” who seem to only be dimly aware of the early papers (if that!) and cannot understand the mathematics well enough to see how far this must have advanced just as a consequence of sample sizes reported in the newspaper. Actually, it is a nuisance for GWAS — you have to correct for population structure as a possible confound for the real effect you are looking for.

Even the papers I link to below are slightly out of date.………

2014-05-09 on infoproc

For panel A, PC1 = 20% of the variance, PC2 = 5%, and PC3 = 3.5%. For panel B, PC1 = 11%, PC2 = 6%, PC3 = 5% and PC4 = 4%.…

2014-05-09 on infoproc

“we all have inalienable human rights regardless of our abilities or genetic makeup”

This is a moral stance I support for living humans. Perhaps not so much for zygotes comprised of only a few cells (cf the usual abortion debate).

Not everyone will agree with me, as you note. Do Neanderthals deserve equal rights? Homo Erectus? Chimps? …

2014-05-05 on infoproc

Politics and self-promotion play a huge role in science.

2014-04-16 on infoproc… then search on page for word “finance”?

Results are not that great but you could also squint at the photo of my bookshelf 🙂

Somewhere on my blog there is a list of recommendations but I can’t seem to find it now, sorry!

2014-04-16 on infoproc

Even a feral child has a much better mental model of the world than any animal.

“how to achieve it” — Deutsch is referring to the *difference* in learning ability between apes and man.

The brain evolved to learn from its environment. Its DNA program is short, but it requires interaction to develop properly.

2014-04-16 on infoproc

Let’s hope not everyone is as sensitive as you are.

“Deutschland über alles” predates the Nazis by almost a century.…

The song is also well known by the incipit and refrain of the first stanza, “Deutschland, Deutschland über alles” (literally, “Germany, Germany above all”), but this has never been its title. The line “Germany, Germany above all” meant that the most important goal of the Vormärzrevolutionaries should be a unified Germany overcoming the perceived anti-liberalKleinstaaterei. Along with theFlag of Germany, it was one of the symbols of the March Revolution of 1848.

2014-04-06 on infoproc

This field is advancing very fast. Gene editing is rapidly becoming a standard tool in plant, animal and human biology. Most labs report being able to use the tools effectively right away. These are not especially delicate protocols.

2014-04-05 on infoproc

Aren’t you interested in secular changes in these quantities? You can be interested in the factual situation without making a moral judgement.

BTW, I’m not a dean, I’m a VP and also a professor. I try to avoid suits and ties as much as possible … I look forward to my full time return to the faculty so I can wear shorts and sandals every day as I did in Oregon 🙂

2014-04-02 on infoproc

In the real world plant and animal breeders have fairly accurate predictive models built using similar methods. BTW, effect sizes are not 0 or 1.

2014-04-02 on infoproc

For twins you should use the broad sense heritability (including non-additive effects). You should also correct for noisiness in the test itself. The resulting heritability (of the “true trait value” — not just test score) might be as high as 0.9 or so assuming good environments. For example, I suspect that if one twin is 2 meters tall there is much higher than 1.4% chance that the other is also at least 2 meters tall.

2014-03-30 on infoproc

I was surprised at how tall Obama seemed given that his official physical records at the WH list him at 6`1 (sometimes I read 6’1.5). In the top picture I am slouching a bit (for some of the pics I had my arm around his waist; not sure about this one). But he did seem more like 6’2 in person. I wonder if he wears lifts or big heeled shoes. The advance team was very meticulous in requiring that we all dress casually – no suits; only Obama was supposed to have a suit on. So they care a lot about how he appears relative to us.

I’m consistently 183cm these days, although I seem to recall being measured at 184cm at physicals a long time ago.

2014-03-30 on infoproc

Care to elaborate on what is wrong with their paper? Maybe you should blame Princeton CS instead of Page House and the physicists 🙂

2014-03-30 on infoproc

Ah, my old Page House contemporary Ed Felton 🙂 Class of ’84 or ’85 IIRC. He was a physics major back then ……

2014-03-29 on infoproc

Seems like a potential governance challenge if they want to pursue micropayments …

2014-03-27 on infoproc

He’s a smart guy, no question. He was touring a biotech facility and he asked good questions during the visit. People who doubt his intellect are nuts.

2014-03-26 on infoproc

I’m confused by this. In earlier studies (1990s data) I think almost all of the increase in inequality could be traced to just four counties in the Silicon Valley, Seattle, NYC-Greenwich areas. But perhaps things changed more recently.…

2014-03-26 on infoproc

Wow, thanks! I guess the modern fraction is even higher than I remembered — over 10% It’s really hard to believe …

2014-03-26 on infoproc

Then I agree: he may be using an expansive definition of manager. (Money manager? 🙂

2014-03-26 on infoproc

I think it’s true that in the US most of the recent increase in income inequality is due to financiers and startup guys. But in France it may be high level corporate managers who are the culprits.…

2014-03-26 on infoproc

The problem is that people don’t assort that strongly for height so kids regress pretty fast relative to the taller parent. It could be part of the Clark puzzle but not the dominant part.

2014-03-26 on infoproc

I had always heard he was 6″1 or so but in person I would swear he is 6″2. Not sure what kind of shoes he was wearing …

I’m only 183 cm, barely over 6.

2014-03-19 on infoproc

Lubos: I think this might be the third time we’ve had this discussion. I don’t think you mean the same thing as I do when I refer to the Copenhagen Interpretation. In fact, when you write detailed things about QM you seem to have the same viewpoint that I have. But the conventional (at least in my opinion) version of Copenhagen is that given by Weinberg below:

Einstein’s Mistakes, Steve Weinberg, Physics Today, November 2005: “… The Copenhagen interpretation describes what happens when an observer makes a measurement, but the observer and the act of measurement are themselves treated classically. This is surely wrong: Physicists and their apparatus must be governed by the same quantum mechanical rules that govern everything else in the universe. …” (See…

>> misinterpret the quantum mechanics in terms of a classical theory but quantum world – our world – is simply not classical in any sense << This interpretation, which you call stupid, IS what most people mean by Copenhagen! In considering the CMB and the BICEP2 observer you cannot extract BICEP2 and treat it classically - the existence of BICEP2, the Earth, sun, our galaxy, etc. in this spacetime region (as opposed to an empty void) is due to a positive energy density fluctuation in the inflaton field. BICEP2 is not a classical observer that can be separated from the quantum state describing the fluctuations of the inflaton -- it is clearly just a part of the overall wavefunction. This makes the CMB situation different from the usual lab setup contemplated in what I (and I believe Weinberg) refer to as Copenhagen. I suspect you agree with the above paragraph, except for the use of the term Copenhagen in the last sentence ... PS When you listed winners and losers why did you not include ADD, Randall-Sundrum and other extra-dimensional solutions of the hierarchy problem? BICEP2 suggests that the scale of quantum gravity is indeed 10^19 GeV and I would think that disfavors those extra-dimensional models as much as any of the theoretical ideas you mention in your post.

2014-03-07 on infoproc

But an adopted child could benefit from parenting or financial inheritance …

2014-03-05 on infoproc

Note we are not talking about general X,Y but specifically: X,Y from a standardized bivariate distribution …

2014-03-02 on infoproc

This might be of interest to you:…

This type of test does not (at least superficially) rely on knowledge acquisition. On page 6 (Table 1) you can see age-dependent norms.

Of course in reality there is no easy way to disentangle “ability developed through training/schooling” and “ability developed due to brain maturation”.

2014-03-02 on infoproc

For younger kids the tests themselves are different. Scores are normalized by age group and mean scores improve with age until early adulthood. In some studies of gifted kids they have them take the SAT early and normalize by age cohort. IIRC, a score that is 99.99th percentile for a 13 year old is much lower than the 99th percentile score of a 17 year old.

2014-03-01 on infoproc

I should clarify that the -2 SD I use in the post is for Neanderthals living in modern society — so, benefiting from good nutrition, education, etc. Neanderthals and early humans living under difficult conditions probably both have depressed abilities. But my guess is that there was a significant gap between the two groups, at least 1 SD or more.

2014-03-01 on infoproc

>> they’re just as intelligent as they will be as adults, no? << I think it's not just learning and experience that make an 8 year old different from an adult. The brain grows and matures physically in a number of ways in the intervening years. My wild guess is that Neanderthal adults might have cognitive abilities similar to an 8 year old modern, and that that is not inconsistent with their ability to survive in cold environments.…

2014-03-01 on infoproc

Y = realized GPAs, X = cognitive ability, W = effort (conscientiousness). We only measure Y and X and W is uncorrelated to X.

2014-02-28 on infoproc

Y = college GPA, X = SAT score, BOTH standardized variables, hence *rescaling* after adding X + E. (That’s the natural way to organize the student data: z-score both GPA and SAT.) Then the relative SD of E (i.e., the noise for someone wanting to use SAT to predict GPA) compared to the standardized variable X is what I wrote in the post.

2014-02-28 on infoproc

That was a typo! Thanks for catching it.

2014-02-28 on infoproc

I was aware of all this, but he must be a special con man talent if he fooled the Icelandic language teacher and others so thoroughly 🙂

2014-02-28 on infoproc

Tried to avoid details in the post discussion above. The “~” symbol or the word “roughly” hides some coefficients 🙂

2014-02-28 on infoproc

Toy model: 10k loci, prob 0.9 of + allele at each site. Then the *average* person has 9000 + alleles and 1000 bad ones. But the SD of this distribution is ~ sqrt( p (1-p) N ) ~ sqrt ( .1 N ) ~ 30 or so. Most people have 1000 plus or minus 30 bad alleles, but exceptional people might have 900 or 800 or 1200, with correspondingly high or low g scores.

Look up binomial distribution.…

2014-02-27 on infoproc

What is your 10% number based on? (A bit scary for me … 🙂

2014-02-24 on infoproc

I think I agree with your comments.

Some context for the focus on EMH: there are a lot of econ and finance types who visit this blog, so the emphasis is a bit more on their theoretical framework …

The calculation with reference to bailout cost was written in 2008, so I didn’t know how things would turn out. In the oversold scenario I was expecting the govt to make money, which I guess it did. But there was a huge cost to the economy overall and that impacted poor and middle class people more than the rich.

2014-02-24 on infoproc

The minority subprime angle is real but I don’t see the reason to focus on that. Just one tragicomic / corrupt aspect out of many.

Why did “sophisticated” investors accept AAA ratings of CDO tranches built from these mortgages? You don’t need “ethnicity” as a variable in the model (borrower income and job history are enough) to see that the subprime structure is very dangerous …

2014-02-24 on infoproc

1) Agree that bubbles are hard to predict and that govt policy had an impact on the housing bubble.

2) Fama is just being dim in the quote. Bubbles are real: suppose I invent a way to get energy out of tapwater. A big speculative bubble ensues because of investor belief in the efficacy of my process. Later a problem in the process is revealed and the bubble pops. The cause of the run up in prices (mispricing) is clear, although not easy to sort out before or during the bubble. Maybe smarter (technical) people could have figured out right away that it wasn’t going to work, but what is a non-technical investor to do? They are merely following their beliefs, and have bounded rationality … What does Fama have to say about that? Substitute beliefs about housing market dynamics, credit dynamics, etc. for the energy scheme and you get back to our most recent bubble. Shiller’s model of how investor sentiment works is more realistic (his wife is a psychologist!) than Fama’s …

So, YES, bubbles are real even if they aren’t easy to predict.

3) The calculation done in this post might reveal a bit more about my opinions on all of this:…

“A little calculation is in order: suppose unfettered markets lead to systemic crises every 20 years that cost 15% of GDP to clean up. I think that’s an upper bound: a $2 trillion (current dollars) crisis every 20 years.

Easy Question: What growth rate advantage (additional GDP growth rate per annum) would savage, unfettered markets need to generate to justify these occasional disasters?

Answer: an additional 0.1 percent annual GDP growth would be more than enough. That is, an unregulated economy whose growth rate was 0.1 percent higher would, even after paying for each 20 year crisis, be richer than the heavily regulated comparator which avoided the crises but had a lower growth rate.

Hard Question: would additional regulation decrease economic growth rates by that amount or more?

Unless you think you can evaluate the relative GDP growth effects of two different policy regimes with accuracy of better than 0.1 percent, then the intellectually honest answer to the policy question is: I don’t know. No shouting, no shaking your fist, no lecturing other people, no writing op-eds, just I don’t know. Correct the things that are obviously stupid, but don’t overstate your confidence level about additional policy changes.

(Note I’m aware that distributional issues are also important. In the most recent era gains went mostly to a small number of top earners whereas the cost of the bailout will be spread over the whole tax base.)”

2014-02-23 on infoproc

I’m generally pro-markets. I’m also OK with weakest versions of EMH (individuals unlikely to beat markets; markets aggregate information). But what the “experts” were saying during the bubble (strong EMH, so no bubbles! actually I think Fama still says this) and just after it burst (medium strong EMH, so CDOs must be correctly priced now…) was pure comedy gold at the time. Now, 5+ years after the crisis, it’s not comedy. It’s tragedy if “the profession” can’t learn from a catastrophe.

2014-02-18 on infoproc

No one can do this yet. You probably need data on roughly 1 million individuals (genotype + phenotype) to extract a good predictive model for IQ. Possibly in 5 years but it is hard to predict.

2014-02-17 on infoproc

There may be improvements one can make on L1. Elastic Net uses both L1 and L2 penalties. What is nice about L1 is the theoretical results concerning the phase transition behavior (Donoho-Tanner) and other performance guarantees, along with the existence of fast computational methods (coordinate descent). My main interest is in providing an upper bound on the data required to select the full set of causal variants for a complex trait of given sparsity. This threshold is probably of order 1 million individuals for g, but with better methods one could perhaps improve it somewhat.

2014-02-14 on infoproc

“enriched” = enriched for mutations. The frequency of mutation in these gene sets was higher in the schizo group than the normal group.

2014-02-01 on infoproc

If they see no off-targets in 5% (exomes) of the genome then (assuming uniform probability of random cutting), there might be ~ 10 at most in the rest of the genome — it depends on how many individuals they did the exomes for. If you did, e.g., 20 individuals and found no off-targets on their exomes then you would be pretty confident that off-target probability is really small. IIUC, in the paper linked above they claim no off-targets in 4 clonal populations, which suggests less than 10 in the whole genome of an individual. I seem to recall other papers, perhaps mouse or human stem cell, that also did deep sequencing looking for off-targets.

For therapeutic purposes, you’d probably be willing to tolerate ~10 random mutations in exchange for eliminating something very bad.

2014-02-01 on infoproc

This lab is, I believe, in Kunming. They don’t have the same issues we have here …

2014-02-01 on infoproc

What do you think of this paper?…

“Here, we show that off-target effects of RGENs can be reduced below the detection limits of deep sequencing by choosing unique target sequences in the genome and modifying both guide RNA and Cas9. … Our results highlight the importance of choosing unique target sequences and optimizing guide RNA and Cas9 to avoid or reduce RGEN-induced off-target mutations.”

It seems like the matching of guide RNA to target region is the crucial thing. Is there really uncontrollable random cutting?

2014-01-31 on infoproc

If you do PCA on Chinese genotypes there is a N-S gradient. If you fly from Shenzhen to Beijing you will notice the height/bulk gradient right away — like flying from Naples to Copenhagen. It is interesting that the early British, Dutch and American traders not only noticed this but recorded it in their writings.

It’s also interesting that China’s “down cycle” will have lasted only about 200 years, if Delano (distant relative of FDR!) and the Koxinga historian (earlier post) are to be believed.

2014-01-31 on infoproc

Thanks for the references!

2014-01-22 on infoproc

This point is discussed in an earlier post:…

For height and g evidence suggests that the variance accounted for by the nonlinear part of the model is smaller than that of the linear part. In other words, the linear model alone should produce correlations between predicted and observed phenotype of > 0.5 or so. The nonlinear terms will increase this accuracy but they are the subleading part.

That additive variance dominates *at the population level* does not contradict the existence of epistasis. If causal alleles are at low MAF then even the nonlinear interactions will be approximately linear in their effect on the population as a whole. For example, if A is rare then most people are either aa or aA and the nonlinear AA (e.g., overdominance) part only affects a tiny fraction of the population (overdominance is a small correction to the accuracy of the predictive model averaged over the entire population). All of this is discussed in a recent well-known paper:…

2014-01-22 on infoproc

PS Just to clarify further, the predictive models used thus far are quite primitive (linear): just associate an effect size with each variant or SNP. Most effect sizes are zero, the rest are +/- small fractions of a population SD. For a given individual add up the effects associated with their genome and you have a prediction. The correlation between prediction and actual phenotype may eventually be 0.5 or higher for height and g.

Determining the effect sizes is a matter of statistical analysis on a (large) sample of individuals. Our paper gives estimates of the amount of data required to find all nonzero effects using human genomes as the “compressed sensor”.

Future developments will include nonlinear terms in the model, etc. Look for a paper in the future estimating sample size requirements for determining the nonlinear effects …

2014-01-22 on infoproc

The predictions will always be statistical in nature, that is probably unavoidable for such complex systems.

>> No blind real life comparison of genomic predictions vs conventional breeding has been done.

It’s just a matter of time for this. What can be done already is to make a model-based prediction on some set of animals or plants and check prediction vs outcome to get a correlation or other performance score. For example, I’ve seen impressive results in predicting the height of corn from SNPs.…

2014-01-21 on infoproc

>>No, it presently cannot be done. Not even remotely close.<<…

“Breeders have recently started switching from pedigree based methods to statistical models incorporating SNP genotypes. We can now make good predictions of phenotypes like milk and meat production using genetic data alone. Cows are easier than people because, as domesticated animals, they have smaller effective breeding population and less genetic diversity.”

We will have good predictive models long before we understand all of the molecular processes involved.

2014-01-21 on infoproc

If you look through all the posts on this blog (as opposed to journalistic stuff written by other people), you’ll see the s ~ 10k and n ~ 1M numbers have been discussed many times. I think I’ve even explained to you before (in comments) that 10k refers to loci, not genes. So 10k is out of millions of SNPs or even billions of bps if you want to count that way (your DNA does a lot more than just code for individual proteins … it’s more than just genes). If you read the paper you’ll see that success with the algorithm (i.e., sufficient sample size) means we’ll be able to build a genomic predictor of the phenotype, even if s is “large” (e.g., 10k). This can already be done with domesticated plants and animals.

“Fighting chance” refers to finding ONE or a few genome wide significant (p<1E-08 or so) hits, not ALL of the associated loci. That was always the goal of the first BGI project, as explained in the talks (see slides) I've given on the subject. But even finding the first hits (by this definition) is a milestone for doubters who don't believe that cognitive ability is influenced by genetics. Unfortunately, our group won't be the first to do this:…

2014-01-07 on infoproc

My mom lives near SLO (my parents retired there from Iowa) so I’ve been going there for years, but never did any clamming. The central coast is still fairly unpopulated and beautiful. You were lucky to experience a golden age in CA!

2013-12-31 on infoproc

CRISPR allows direct gene editing. Read the links on this blog, not Wikipedia.

2013-12-31 on infoproc

In my opinion, no singularity for 50+ years. Most people who use the term singularity are incredibly over-optimistic about it. Some comments here:…

I’ve been following AI/CS and genetics since I was a kid but I didn’t do anything actively in genetics until just recently.

Note Moore’s Law itself is in trouble:…

After another order of magnitude or two we will reach chip feature sizes where quantum effects become important. This could bottleneck Moore’s Law progress seriously for some (perhaps long) period of time.

2013-12-31 on infoproc

CRISPR is looking very interesting.

2013-12-31 on infoproc

While I believe that people will look back and wonder what all the fuss was about, I don’t believe the consequences of the new technologies will be anything less than transformative in the medium to long run. I’m very confident that there are many, perhaps ~ 10-30, SDs up for grab here.

2013-12-31 on infoproc

I find the Oklahoma ranking strange. I don’t recall seeing them that highly ranked in other places.

2013-12-24 on infoproc

It’s my older brother. Another photo of him here:…

2013-12-20 on infoproc

Everyone has structural variants in their genome. IIUC from the article, about 1% of the Icelandic population has neuropsychiatric CNVs as defined in the study.

2013-12-19 on infoproc

Averaging will remove rare variants, and if most of them are negative, then the trait value will increase.

2013-12-19 on infoproc

The movie doesn’t focus on what Samantha’s 1E09 other instantiations — the ones who work in R&D, not “human-care services” — are up to in the background 😉

2013-12-02 on infoproc

1) We have not published this result. The effect was not statistically significant when we tried to replicate it on a couple of other populations, so the 1 SD ~ 40 SNPs is only a rough *upper bound*. That is, if the # of SNP changes per SD change in the phenotype were larger we would have detected it (above noise) in the other sample(s). It could be 40 or 30 or 50 but 100 is probably ruled out. We had to use jacknife / bootstrap to get an idea of the background noise for this method.

2) If there are N causal variants then it takes about sqrt(N) SNP changes to get 1 SD. So if N ~ 1E04 then 1 SD ~ 100 would be about right (note avg MAF of causal variants also affects this relationship; see slides). We didn’t make assumptions about N. We *measured* the # of SNP changes per SD difference in phenotype by averaging SNP distance over all pairs in a population of order 5-10k individuals for whom both phenotype and genotype were known.

2013-11-28 on infoproc

Cornelius: V and M are positively correlated — this is the basic observation leading to g, etc.

I realize you and I were not describing the same situation. I meant that there are many individuals who are, say, +3V, +1M, so “low M” is in relative terms, not absolute. I don’t think I’ve met anyone who is +3V, -1M, although I suspect they exist (just very rare).


2013-11-28 on infoproc

> High V individuals who seem to have low M are just mathematically ignorant. < I know many counterexamples to this statement, including individuals I have known since childhood.

2013-11-27 on infoproc

re: utility of g, if you are admitting thousands of kids year after year to a freshman class, then statistical validity is all you need ask for from the test. Even a fraction of an SD difference between two groups (e.g., Berkeley vs UCLA admits) could have a big impact downstream: e.g., on Nobel prizes won per alumnus.…

If you are only selecting a *single* programmer or grad student for your team, I’d still say an SD difference between two candidates is worth thinking hard about (other things being equal). So, I’m certainly not sympathetic to the claims of the anti-g crowd.

I think the biggest non-testable factor is motivation, drive, work ethic. We try to measure that using proxies but it’s challenging, not least because the factor is itself time and context dependent.

PS I sent you an email

2013-11-27 on infoproc

> there is some aspect of innate intellect in play that is not captured by the tests < This is undoubtedly true! The tests only have validity in a statistical sense (averaged over large numbers of people). Fun at IAS: ... All the questions were very practical. And very sensible. And I was rather terrified of the Institute before that, because it was well known that all these guys at the Institute would talk a good game, you see. Like somebody would say, “Well, isn’t that just the same as Smorglepop’s theory?” I’ll give you another example of it in a minute. At any rate, I gave the lectures, and there were nothing but practical questions like, “If you were trying to do this problem, how would you set it up? Did you mean by this a minus sign there?” “Yes” You know. All perfectly OK, and I gave nice lectures. I gave all the lectures I wanted to and explained everything and went back home to Cornell. I said: “Hey, Hans, the Institute has changed! Something has happened. These guys are very different. They didn’t ask anything but sensible questions. They didn’t say, ‘Isn’t that the same as Porkyschnorp in 1621?’ or something like that.” “Oh,” he said, “I know the reason. I went just a few weeks before you did and gave some lectures on nuclear physics. I started to give the lectures and I hadn’t opened my mouth and said two, three, four words, when this stuff started. Somebody jumps up and asks a question. He says, ‘Isn’t that the same as what Wegischnorp said in 1960, and so on, in a paper in the Weische Physica Acta?’ Somebody else, before I can answer, gets up (typical Princeton Institute) and says, ‘No, you see, what Bethe is going to say is ‘this, that and the other thing,’ and what the fellow says in the Weische Physica Acta is ‘this, slightly different.’ And somebody else says, ‘No, it isn’t exactly so different, because Bethe —’” He says to me, “So when the third fellow gets up to argue, I slam the table” — you know, when Bethe gets mad he can look formidable -– “I slam the table and I said, ‘Gentlemen, if you knew what I was going to say, why did you invite me to speak? Now, I want to make an uninterrupted speech, unless you have a specific, detailed, and sensible question.’”

2013-11-27 on infoproc

Note, this discussion is necessarily subjective and I don’t ascribe a high degree of confidence to my comments. But it’s fun to discuss. The Feynman v. Schwinger (and to some extent Dyson) observation is something I have thought about for many years. I have had the chance to interact with F and D a bit and have heard talks from S. Certainly S and D would score higher on a V test than F, although F is still very expressive and a good communicator (one of the best!).

I know people (I guess I am one) who can quickly scan papers and get the gist of things. (The average scientist would be shocked at just how fast this can be done. A big part of it is simply preprocessing — having a better mental map of the key issues across many areas. F definitely had this! But part of it is a V-dependent ability to parse and decode what is written on the page.) If the paper has math in it then this is necessarily a combination of verbal *and* mathematical fluency. I also know people who are very good problem solvers yet are not fast readers and cannot do what I just described. They are much more comfortable solving a problem themselves as opposed to understanding someone else’s solution, especially when that solution is encoded in the formal, academic language of a scientific paper. The first group tends to have a global understanding of the literature and the second group much less so. Reading is still the fastest bandwidth path into the brain! However, there are some otherwise smart people who don’t learn primarily by reading. I don’t know if the V factor is the best one to characterize this, but it seems reasonable to me. The second group also tends not to have as large a vocabulary as the first, etc.

Incidentally, a lot of programmers and engineers and experimental physicists are in the second group and their V scores are certainly lower on average than among theorists. (See Roe study and SMPY…)

If you read the whole AIP interview there is another revealing part where F discusses the kind of grilling typical at IAS. Note the questions often ask the speaker to relate his work to some larger context or very esoteric earlier work. Those kinds of questions, which F felt very uncomfortable with, because of the holes in his knowledge, were typical high V theorist questions — i.e., from erudite know-it-alls who have been absorbing and structuring/compressing knowledge at a high rate throughout their lives. F felt more comfortable with “practical” self-contained questions, e.g., about the method of calculation or some numerical factor.

2013-11-26 on infoproc

I’ve been in the position of having to select teams to accomplish *real* tasks with objective standards of success — not just impressing clients. HFT firms, most software startups, and most science and engineering teams are also in this position. They may be misguided as to the best techniques for filtering human capital (brain teasers are noisy indicators, but can give you some insight into how a candidate thinks, etc.), but their intentions are real. A good quant/developer/technologist is worth 10 mediocre ones …

2013-11-25 on infoproc

No, Malyshev’s boss at Citadel was someone you know/knew. A student of Stanley’s. Look up the court docs 😉

2013-11-25 on infoproc

Is it really weeding out or are there hidden correlations somewhere? For example, I don’t think there is much pressure for political conformism among science faculty. At least, I know conservative physicists, mathematicians, computer scientists, etc. who don’t seem to suffer professionally for their political beliefs. Yet, the O v. R imbalance you speak of is pretty high even in this population.

Most faculty have no experience with business — they’ve never had to make payroll, fire lazy workers, struggle against inept government bureaucracy. It’s not surprising that they are more likely to vote left than equally smart people who went into private enterprise.

As you know I’ve never been a big Ayn Rand fan. However, time as an entrepreneur in Silicon Valley has me almost embracing some of her viewpoints (most progress is due to the intense efforts of a small number of people, etc.) … very few professors have had this kind of experience.

2013-11-23 on infoproc

If you do a statistically careful survey of 160s, over a wide range of questions, I don’t believe that they’ll have, on average, more nutty or incorrect or unjustifiable beliefs than a group of 130s. In fact it is much more likely that the 160s are “rational” (evidence based, Bayesian, whatever) thinkers than the 130s. Some (many?) 130s have trouble understanding Bayesian updates, decision making under uncertainty, belief propagation, systematic cognitive biases, etc. A 160 would not have this problem, although they may be sloppy about using “rationality” in all areas of thought.

PS I agree with your assessment of Chomsky’s beliefs.

2013-11-23 on infoproc

I don’t disagree with Chomsky’s statement — the effect is there, it’s a question of magnitude.

2013-11-19 on infoproc

Heritability can be studied now in populations of unrelated individuals (regress phenotype similarity in pairs against genetic similarity as measured by SNPs). The results are consistent with the much earlier estimates from twins studies. So it doesn’t seem that the shared womb is a big effect.

2013-11-18 on infoproc

To get an SD change in g you only need to alter of order sqrt(10k) alleles. Actually the number is even less than that because the minor allele frequency p of causal alleles could be much smaller than 0.5 on average — 1 SD ~ sqrt(p 10k).

2013-11-16 on infoproc

Both are trained in physics 😉

2013-11-16 on infoproc

Did you not notice I voted twice for Obama?

In my opinion Levy is a talented writer (I don’t care about her lifestyle choices) and Berkeley is one of the greatest universities in the world, with a tradition that inspires tremendous pride. Lawrence, Oppenheimer, Kerr, Savio and Birgeneau can all be counted among the illustrious sons of California.

2013-11-07 on infoproc

Look at the general UK reaction to Cummings’ essay. Read some of the commentary in the Guardian, etc.

2013-11-06 on infoproc

When new information conflicts with cherished beliefs the process of learning can be painful.…

2013-10-27 on infoproc

The main purpose was to assess adults (e.g., age range up to 60+), so most of the individuals tested have been out of school for many years. Note the Japan-Italy comparison made in the report is on measures of literacy (reading ability and reasoning).

2013-10-26 on infoproc

Yes, it’s always about status (that which is honored by society). At the beginning doing things that are good for the group bring high status, but in the decadent era money=status and it’s every chimp for himself!

2013-10-26 on infoproc

It’s really about values. In the earlier stages people are more likely to value their contribution to society. In the later stages they have become cynical and self-interested.

“Money replaces honour and adventure as the objective of the best young men.”

Glubb’s ideas about stages of empire are not very novel — similar observations have been made elsewhere, since Roman times. What I like is the way he describes the changes in values during the late stages — something he probably experienced in his own lifetime in the British empire.

2013-10-20 on infoproc

C&H seems to have made my son naughtier. Not sure whether it’s good or bad 🙂

2013-10-14 on infoproc

L1 penalization is just a method that allows one to efficiently find the sparsest solution. If you want to complain about idealization complain about the additive (linear) genetic model. But there are empirical justifications for this approximation.

2013-10-14 on infoproc

Yes. It would be surprising if all loci affected a particular trait and also counter to initial indications from GWAS. Various priors like the one you suggested are similar to the prior of sparsity. But in the case of L1 penalization there are rigorous results like the ones we reference in our paper.

2013-10-13 on geneexpression

Well, if the field keeps advancing it will hopefully attract more and better human capital.

2013-10-12 on infoproc

Physics has abstract constructs like those in pure math, but also emphasizes the scientific method and experimentation as the main driver of knowledge acquisition. Physics people will instinctively try to formulate simple (but not too simple) abstract models and test them against data. They are much more likely, in my experience, to have out of the box ideas than engineers — I think this is because they have the habit of thinking from first principles.

Math guys are very bright but often are much less practically able than physicists. I am not surprised at Bezos’ ability to quickly improve the efforts of a team of engineers or other “specialists” — it’s a common experience for physicists who go into technology or finance.

PS At the World Summit I attended (see earlier post), of the first 15 or so speakers (MSFT research guy, VC and former CIO at Yahoo!, entrepreneurs, etc.) about 5 had a background in physics!

2013-10-11 on infoproc

Perhaps you should ask Bezos whether he agrees with your take on this… 😉

2013-10-09 on geneexpression

The problem is that individuals with a mastery of all the disparate areas you mention above are quite rare. Most people who understand SNP calling or RNA-seq don’t also understand Fisher’s Fundamental Theorem or variable selection from high dimensional data. Hopefully this situation will improve over time.

2013-10-08 on infoproc

Given the statistical errors in estimates of effect sizes I would say the correlation in the graphs is very impressive. A more sophisticated analysis, taking into account uncertainties, could put an nontrivial upper bound on interaction effects.

2013-10-08 on infoproc

It’s still a work in progress, I think.

2013-10-06 on infoproc

inflationary = negative effective pressure, not mass.

There are positivity theorems in GR that don’t let the effective mass become negative.

2013-10-05 on infoproc

As I imply in the post above (“… happy to find … one genome-wide significant hit”), there are plausible models of the genetic architecture in which more statistical power than we have in the current study is required to even detect one causal variant. In our original power calculations (see, e.g., the Google talk I gave early on) we used height as the model to estimate number of hits. Since then evidence has accumulated to suggest that g will be at least a bit harder than height. So we might not find any hits in our BGI analysis.

My rough guess is that we’ll need sample size of hundreds of thousands to millions of normal individuals to get to a good chunk of total g variance. In research of a more mathematical nature, we’ve made some progress on the “variable selection with high dimensional data” problem and are pretty sure there will be a “phase transition” to good recovery of loci at a certain point, so progress (when it comes) may be more dramatic than what most people in the field expect.

2013-09-26 on infoproc

Bica, across from Rockridge BART near Zachary’s. Third wave espresso, bro 😉

2013-09-22 on infoproc

Rigorous types question whether you can really measure alpha (i.e., true effective alpha not just performance relative to benchmark) even for a pm who has been trading for, say, 5 years. Along similar lines, can you tell the difference between a BS artist who is good at impressing people (i.e., superiors) and someone who is actually effective in getting the most out of a team? Can you sniff out a high functioning sociopath? Can you do it using only a paper trail of references (selected by the applicant) and a brief interview? What’s the hit rate in the CEO or C-level selection process at most companies?

2013-09-22 on infoproc

There are plenty of smart kids whose parents just aren’t into things like CTY or IMO or Intel. Those kids will just bloom a bit later. You can get a first rate UG education at any of the top 100 US universities and if you bloom then you’ll still get into a good graduate program. Top 10 schools might get 50-70% of the top few thousand kids; the rest are all over the place:…

And we’re just talking about cognitive ability. Success in many fields depends as much or more on other factors like intuition, people skills, determination, etc.

My guess is “elite MBA” admissions is mostly signaling and less a real measure of talent — as difficult as it is to measure cognitive ability or conscientiousness, management skills (for example) would be much harder to determine from an applicant file. You end up with a self-fulfilling prophecy: remember, NO ONE thinks that the classroom learning that goes on in the MBA program is that important. It’s the branding and networking that counts.

2013-09-22 on infoproc

I agree with your take on recent trends but I doubt filtering methods are that predictive. Plenty of great talents, even geniuses, will not top the rank ordered list of HS applicants at that stage of their lives. While you’d be smart to bet on an IMO winner doing something significant later in life (although certainly not all do), there are many others who aren’t as distinguished at age 18 who will as well. At +4 SD in cognitive ability, there are a few hundred kids available each year to US colleges. H probably gets more of these than any other school, but they certainly don’t get nearly all of them. A few thousand kids each year come close to maxing out the SAT M+CR. I suspect in that population it’s a crap shoot to identify the future super talents, with the proviso that the M ceiling is kind of low so you could easily find people (e.g., USAMO) that are well above that level.

PS Tsinghua and U Tokyo are significantly bigger than Harvard … 14k undergrads each … and each gets a strong draw from their respective national talent pool (no athletes or legacies, last time I checked).

2013-09-22 on infoproc

Re: Ito, yes, but many more people get at least some exposure to SE or two slit experiment.

The two slit experiment is elementary, but still mysterious, and many people (including most scientists and some Caltech sophomores) would have trouble explaining it and answering detailed questions.…

2013-09-22 on infoproc

Class sizes vary quite a bit. Top 5% of H is almost 100 people whereas top 5% at Caltech is only about 10. I’d say the quality of the latter group is higher.

If you were to say that the top 5-10 kids in a graduating class at H are as good or better than the top 5-10 in any other class then I would probably agree with you. Note, though, that recently MIT has been getting a lot of the IMO kids. I also suspect the top 5-10 at Tsinghua or the University of Tokyo are pretty good 😉

2013-09-21 on infoproc

Good point — I removed the modifier “deeply” 🙂

2013-09-19 on infoproc

Sorry it has taken so long to get results back. Unfortunately, I’m not at liberty to explain all the reasons for the delay.

Perhaps we should have offered a saliva-back guarantee?

2013-09-15 on infoproc

> For the past decade or so, the top American students from the top colleges have heavily gravitated towards finance, consulting, and recently tech, NOT STEM Phd programs. < Depends how far you go into the tail of ability. For the most intellectually able subset finance and consulting still mostly get the second team. Ask Jim Simons what he thinks about it ("MBAs need not apply" 😉…

> Admission to these programs acts as a power signal to employers < This is certainly true. Whether those employers know what they're doing is another story. Divvying up the spoils (extracting rents) and creating real value are two different things.……

2013-09-15 on infoproc

HBS churns out > 800 MBAs per year. Compare to the PhD output of top 5 math + physics departments in the US, which produce about 200 per year. One year of HBS grads vs four years of top math/phys guys. So 20 years of HBS MBAs should be compared against ALL the top math/phys guys produced by the US since it became a scientific power. In the latter population you have people like Jim Simons (richer than any HBS grad?), Ed Thorp (early hedge fund billionaire, invented options pricing before Black-Scholes, now taught at HBS), people who invented the computer (Turing, a Princeton PhD; he also helped the Allies win WWII!), semiconductor (Intel founders like Noyce — created more value than any MBA?), laser, superconductors, quantum computers, created atomic weapons, individuals who win Nobels and Fields Medals and will be remembered in 100 years. I suppose it depends on what you are impressed by, but even if you restrict to making money (highest net worth) and creating great companies I think the HBS guys could be on the losing end.

MBAs typically take over existing business built by real entrepreneurs who most often do not have MBAs. How much value do they actually create? Jobs vs Tim Cook? It’s not just STEM guys who are suspicious of HBS value. Ask entrepreneurs in Silicon Valley.

2013-09-14 on infoproc

Their results suggest only limited genetic correlation between g and height. So if you think that genetic load affects both through pleiotropic effects you’d have to restrict it to the part of heritability yet unaccounted for by SNPs — i.e., variants with MAF below about 0.05 or so.

2013-09-14 on infoproc

It’s not just that g is correlated with math ability — you don’t need genomics to deduce that. It’s that the *SNPs* which are correlated with g have a lot of overlap with those that are correlated with math ability. This is not surprising to most people, but it does say something about whether there is a real “general factor” of intelligence as opposed to a bundle of separate abilities that just happen to be correlated (e.g., due to environmental effects).

2013-09-14 on infoproc

This study just says that there is substantial overlap in the SNPs which account for, e.g., heritability of about 0.5 for height or g, and those responsible for the heritability *due to SNPs* of, respectively, bodyweight or narrow cognitive measures (reading, math ability). In other words, there is pleiotropy among common causal variants. But note the overlap between genes affecting height or IQ is much less (perhaps even close to zero) than for two cognitive traits.

Unaccounted-for heritability (i.e., the gap between 0.5 and some larger value like 0.6 or 0.7 additive heritability) could be due to much rarer variants. “Common” SNP MAFs extend down to 0.1 or 0.05 so it’s a matter of terminology whether you want to call them rare or not.

2013-09-14 on infoproc

This study did not find specific genes and does not say anything about group differences (neither will ours). It does suggest some things about the genetic architecture of cognitive ability.

2013-09-12 on infoproc

I actually didn’t like UG at Caltech that much (mainly, no girls) and would have transferred somewhere wimpy like Stanford or Princeton had I not been able to graduate in 3 years.

Yes, as you can tell from the blog I’m a big Caltech booster. That’s because I believe in the mission.

“Isn’t it right that a country the size of the United States should be able to afford one university in which intellectual achievement is the most important consideration?”…

2013-09-11 on infoproc

Teaching future theorists, or even teaching students that are likely to go on to obtain PhDs, is quite different from teaching engineers, or premeds, or physics 101 to humanities and social science students. Of course I, personally, would rather have a brainy professor with really deep knowledge of the material. But on the other hand you run the risk of bad, unprofessional teaching from an eccentric genius. I’ve experienced the whole range of possibilities …

I agree with you that the variance is highest among the researchers. The brainpower thing plays a role, and the potential for aspie inability to connect with students or understand what they are confused about. I think the main issue is that researchers have not (primarily) been selected for their teaching ability, whereas adjuncts or permanent lecturers could and should be.

It is interesting that the weaker students in the Northwestern study got the most benefit from non-tenure stream instructors. This is what I would have predicted.

I like David’s story below about his QM class. It certainly rings true to me. No offense to experimenters, but it is usually a big mistake to have a serious QM course taught by a non-theorist.

2013-09-11 on infoproc

Yes, the department has to trust that the administration wants to improve research productivity in the department, attract better faculty, etc. and is not unduly sensitive (for whatever reason) to the use of adjuncts.

The reactions of my colleagues to my adjunct proposals were sincere — not based on meta-strategic considerations. I think they honestly had the wrong beliefs regarding relative quality of instruction.

2013-09-10 on infoproc

The incentive landscape is complex and varies from university to university. But, just to take a simple example, a chair could convert one or a few tenure stream FTEs into adjunct salaries and meaningfully reduce the teaching “load” on active researchers in the department. In my experience, reduced teaching is very attractive to good researchers, and would aid recruitment. If, for example, you could give every active researcher a semester off every other year that would be very enticing. Ultimately you could end up with a stronger department overall by allowing “specialization” — researchers do (more) research, and teachers do (more) teaching.

2013-09-10 on infoproc

I think the last point you made about variation of g loadings by population is controversial, at least from what I have heard second hand from psychometricians.

In any case, the pragmatic view of g is that, however it was assembled or defined, the important question is whether it has *validity* (i.e., predictive power).…

2013-09-08 on infoproc

I think it’s this guy’s work: http://www.kotterinternatio…

2013-09-06 on infoproc

I was expecting this kind of result from the beginning. In fact, I couldn’t believe how overly optimistic researchers were 10 years ago with their “one gene one disease” nonsense. I suppose this was useful (while it lasted) for getting money out of big pharma and other dim funders.

2013-09-06 on infoproc

We already know that by and large genotype determines phenotype. Now we are learning about the specific genetic architecture and how to predict phenotype given genotype.

2013-09-05 on infoproc

There is much more information in the genome than in just the 20k or so genes (exomes). Think of all the non-coding control regions. So 10k out of millions of snps (regions of common variation in humans; out of billions of loci in the whole genome) is a significant localization of schizo risk.

2013-09-04 on infoproc

Thanks! I posted the following comment there:


It is quite difficult to demonstrate unitary evolution in an experiment because to do so one would have to detect decohered branches of the “many worlds” wavefunction. This requires either exponential sensitivity (of order exp (-S) where S is the number of dof of the measuring device), or the ability to prepare operators comprised of macroscopic superpositions.

This article clarifies the analogy between unitarity in BH evaporation and unitarity in ordinary QM measurements:…

2013-09-01 on infoproc

As you know I was referring to different phenotypes for those species, although the math is basically the same.

In the case of dogs, it’s clear that they are shifted in cognitive ability relative to their wild cousins (wolves). The average dog is a many SD outlier among wolves in ability to interact with humans.


2013-09-01 on infoproc

I suggest you look up terms like parental midpoint, regression to the mean, etc. IQ doesn’t generally increase from generation to generation except due to selection pressure.

2013-08-28 on infoproc

Hi Lubos,

Thanks for your comment and blog post about my paper. Note I cite the Nomura gang as well as P-R. I actually have you (earlier post on your blog) to thank for pointing me to P-R, which I agree is a nice paper.

I was kind of surprised that most people working on BH information take “factorization” for granted, whereas it appears both you and I consider it a nontrivial assumption. It seems obvious to me that if you give up factorization you can avoid firewalls, but I anticipate this claim may turn out the be controversial.

I’ve also watched many of the KITP talks on firewalls. I wonder if Raju was able to get his points across in offline discussions — e.g., concerning what is meant by “state dependent operators”. To me it’s kind of obvious that, e.g, the b, tilde-b operators in the AMPSS notation are state dependent in the sense that they are clearly background geometry dependent!

2013-08-28 on infoproc

The B-P-S catastrophic energy non-conservation would local and ubiquitous… but I was never convinced by their paper (I’m with Wald and Unruh on this).

2013-08-26 on infoproc

I hesitate to calculate the financial opportunity cost of each physics paper I write 😉

Warning: Physics Envy may be Hazardous to your Wealth, Andy Lo and Mark Mueller…

2013-08-26 on infoproc

Who says she’s an ubermensch? Her abilities are far from superhuman, but she did pretty well with what she had …

An ubermensch would hardly have to pull all-nighters with a study group to get through 160A 😉

2013-08-25 on infoproc

Drugs like steroids give you an advantage, but you still have to put in the work to make progress. Watch the video — it’s short.

2013-08-23 on infoproc

Well, I agree that giving up EP so quickly is crazy.

But I do think if ~30 really top theorists (check out the workshop participants at KITP) have thought hard about it for ~1 year and can’t sort it out then it’s nontrivial …

2013-08-23 on infoproc

Not sure I understand, but I think you mean that to a particular semiclassical observer it appears the BH destroys information whereas a global observer (who can see the all branches of the wavefunction of the universe) sees unitary evolution. If so, I agee 🙂

Note there is an exact parallel to ordinary QM measurement (e.g., Stern-Gerlach), which *appears* to collapse the wavefn (first to a mixed state, and then, probabilistically, to a final outcome), whereas a global observer (at least in many worlds) can see that no collapse actually happened.

2013-08-23 on infoproc

Surprisingly many! In fact a significant number of theorists whose brainpower I admire have (a) looked carefully at AMPS and (b) can’t find the problem. Among the people who share my perspective I can only count Nomura and co. and Lubos!

I don’t think it’s a problem with understanding of GR. The subtle aspect is really in the QM part.

2013-08-23 on infoproc

Actually, in some sense the BH information problem is *posed* within a many worlds framework — does the quantum state of the BH + radiation + environment evolve unitarily (pure state to pure state), or do BHs cause pure states to evolve to mixed (Hawking’s original claim)?

Unfortunately I can’t make it to the meeting — my day job interferes with doing physics 🙁 But I’m getting ready to post a sequel to my earlier paper since this subject remains unresolved.

2013-08-19 on infoproc

I’ve never looked carefully at heritability of longevity, but I thought it was at least moderately heritable — like 25 or 30 percent of variance.


2013-08-16 on infoproc

Oops, thanks for catching that. I think it’s actually misspelled on the IAAF site. To make things worse, there’s a string theorist named Peiming 🙂

2013-08-13 on infoproc

I agree, O was probably V > M whereas F was M > V. I doubt Bohr was extremely gifted mathematically, but perhaps you don’t consider him great 😉

2013-08-07 on infoproc

Recent Duke admissions records are available via their sociology project on college life. Duke’s president Brodhead is a Yalie, former Yale professor and former dean of Yale college, so I suspect that Duke’s methodology is not dissimilar to Yale’s. I doubt there are sophisticated models. There may be general conclusions taken from historical studies — such as that athletes are more likely to succeed in business, legacies more likely to give, etc. Some of this is discussed in The Chosen in the context of Harvard admissions. The Duke scoring system is quite conventional.

2013-08-02 on infoproc

I haven’t seen the fight, but grappling in MMA is very different from sport BJJ — your opponent can’t punch you in the face in a sport match.

Enson Inoue once told me, as we were practicing a sport BJJ move: “This shit would get you killed in a real fight!”

2013-08-01 on infoproc

MMA is not BJJ!

2013-07-31 on infoproc

I’m very interested to see how far Askren can go with his “ordinary” physical abilities. Lawal needs to develop better striking to get to the elite level, but you’re right he has the physical tools.

2013-07-15 on infoproc

Agreed. If you were really careful you might take it as an uncertain hypothesis that democracy is a big positive US differentiator. As you noted we have many other advantages …

PS I am curious what you thought of the part I found most interesting — on technological optimism, comparison of the current debt crisis with the energy crisis of the 70s, etc. It’s near the end of the talk.

2013-07-14 on infoproc

The violence / civil war thing isn’t that central to his argument (in fact, technology and innovation are much more important). It’s mainly about whether American democracy can continue to work, a topic I suspect you’ve thought about. He doesn’t have a clear answer but states both sides of the argument pretty well.

2013-07-08 on infoproc

I suspect he is referring to an anti-communist “loyalty oath” required at the time.

“Are you, or have you ever been, a member of the Communist Party?” 🙂

See, e.g.,…

2013-06-30 on infoproc

Although I’m here at the moment, I’m not really an expert on the environmental and QoL issues here. Long term expats would know the situation better than me. We’ve had clear blue skies while I’ve been here, but perhaps that’s just luck.

2013-06-30 on infoproc

I’ve never been a fan of HK — dirty, noisy, hot. But the air and water quality seem fine. I do like the views from the crazy vertical architecture 🙂

2013-06-24 on infoproc

I think they actually cite you guys — there’s a table in the supplement with all the values of h2. My feeling is that the uncertainties in GCTA results are significantly larger than what people tend to report …

2013-06-24 on infoproc

Some have hypothesized that genetic variance is mainly due to rare variants of large effect, as opposed to common variants of small effect. At least for HDL-C, it seems to be the latter. Whether the heritability is due to high or low MAF variants has lots of practical consequences …

2013-06-21 on infoproc

The paper is not exactly clearly written, but if you look at it more carefully I think you’ll see that the issue is power. Many of the EA studies are underpowered so you have to adjust to examine the “real” rate of replication. See Fig 5, for instance.

2013-06-21 on infoproc

The point is that “failure to replicate” is defined relative to a statistical threshold, which in turn depends on power (sample size), which is not indicated on the graph. Underpowered studies are not expected to replicate even if the effect is real. Once you take this into account (i.e., restrict to studies which had the power to see the effect) you see that the EA and E results are very consistent.

Alternatively, you can just consider each study to be a (noisy) estimator of OR (allele effect). Then you see from the scatter plot that the ORs from studies conducted in EAs (although overall less well powered) are highly correlated with ORs from studies in Es. From the results in the paper you can actually estimate the residual variance in OR that arises due to different genetic and environmental backgrounds (E vs EA), and it is small: OR_E is a good predictor of OR_EA.

2013-06-21 on infoproc

I think you are confused. What is important is that the Odds Ratios (OR) are highly correlated in both the Euro and EA studies. That suggests that the effect of the allele is similar in both sets of backgrounds (genetic and environmental).

The 50% figure you quote does not adjust for statistical power:

“… calculating replicability rates after controlling for statistical power. First, we focused on the 81 attempts with ≥80% power to replicate the Odds Ratio (OR) found in Europeans (Table S5 and Materials and Methods). For that subset, replicability increases dramatically to 76.5% (62 out of 81 attempts are successful with a P<0.05 threshold). Second, we calculated that at most 132 positive replications would be expected out of statistical power (59% on average for the 225 attempts in East Asians, Table S5). The 103 observed replications thus correspond to an effective replicability rate of 77.9%, which suggests that a noticeable fraction of GWAS associations are shared across Eurasians."

2013-06-06 on infoproc

> I’m intensely interested in why suddenly we started cultivating food and building cities, but not sure genetics holds the key, and if it does it seems likely it is mutations later than the original emergence of homo-sapiens. < Yes, it's the *later* version of homo sapiens that we are genotyping. So those mutations you refer to are part of what makes us different.

2013-06-06 on infoproc

They had 300k years and did very little. We’ve had 50-100k years and done much more. I think there’s a good chance that there are differences in average cognitive ability between us and them.

2013-06-05 on infoproc

Neanderthal brains were similar in size to human brains. (Perhaps even bigger?) Whatever makes us smarter is in the small differences between us and them …

2013-06-04 on infoproc

I don’t think I’ve seen the VW through the wormhole clip online — I did see it on the Science Channel. When the wormhole has a quantum instability it lands me on the sun or something like that.

The video quality of the VW going through the Caldecott tunnel is amazing considering it was shot by a tiny HD digital camera duct taped to the roof of the car!

2013-06-02 on infoproc

Yes, thanks for mentioning the film. I’ve seen it and it is excellent.

2013-05-29 on insidehighered

Flaherty declined to include my detailed statement concerning genetic screening (see below) which I emailed directly to her in response to her request. How that constitutes good reporting (as opposed to deliberate sensationalization of this topic) I do not know. I expect better from Inside Higher Ed.

For more professional reporting on the BGI research project, which does not try to make the unwarranted leap to eugenics, see links below to articles which appeared in Nature and the Wall Street Journal.


My position on preimplantation embryonic screening for intelligence (or any other quantitative trait) is:

1. It is highly speculative: it may become possible in the next 20 years, but we don’t know for sure. Perhaps it will remain science fiction. Further research is necessary before we know.

2. *Assuming* it is in fact possible: each society should decide (e.g. by democratic process) whether they want to make it legal. If a particular society decides to make it legal I have no problem with that outcome.

3. In the event that a particular country decides to make it legal, I would rather see it made free (paid for by government) than to have only wealthy people doing it.

If you want to quote me, please quote the entire discussion above. Don’t take any parts out of context.


Earlier email sent to Flaherty:

Hi Colleen,

Ordinarily I’d be happy to help with your story, but our collaboration has decided there has been too much wacky coverage of our project and are going into a quiet phase while we analyze our data.

Our work has nothing to do with IQ selection of human embryos or IVF, except that in the *long run* basic science on the genetic architecture of cognitive ability could have an impact in those areas.

These articles are fairly accurate descriptions of the project (unlike the one that appeared in Vice):……

I address many of the points you might be interested in during this discussion on NPR:…



2013-05-28 on infoproc

Yes, Mochizuki is an oddball among oddballs. It’s not even clear his claimed proof is correct. I should have been more specific in stating that my admiration is for people who are willing to work for years and years on very ambitious projects. Perhaps Zhang, Perelman and Wiles are better examples …

2013-05-28 on infoproc

You are right about this, although perhaps his upcoming colloquium in June at the University of Tokyo might signal some increased willingness to explain the work?

2013-05-19 on infoproc

Do you think “top students” have trouble running Excel? 😉

In ordinary business and accounting (as opposed, say, to derivatives pricing or HFT trading) there is a limit to the level of quantitative tools that are actually useful. Most of the time some basic statistical analysis plus back of the envelope estimates (maybe, wow, a linear or gaussian model) are all that can be justified given the unquantifiable residual uncertainties. Further sophistication just lends a “spurious air of technicality” — useful only to snow clients 😉

Really cynical people would suggest that what I wrote above even applies to finance, but then again you have Renaissance / Jim Simons as a counter-example…

2013-05-16 on infoproc

I sympathize with your comment, but unless I refuse all contact from journalists this kind of thing is going to continue …

2013-05-15 on infoproc

I don’t want to go into detail here, as this is a matter of ongoing research, but having whole genomes rather than SNPs should allow novel analyses that go beyond simple GWAS. For example, we may be able to say something about mutational load … Of course, I’m just an ignorant theoretical physicist, not a “real expert” on genomics 😉

2013-05-14 on infoproc

That’s how science works. We don’t know how many hits we’ll get because we don’t (yet) know what the underlying genetic architecture is. There were many failed GWAS before they found the first height hits at sample size of about 10k.

2013-05-14 on geneexpression

Had GATTACA-like technologies been available for me I might have used them. But I’m not pushing other people to do so. It should be a matter of individual choice, and also for societies to decide collectively via the democratic process (i.e., some countries may wish to make use of certain genetic technologies illegal).

My main motivation for understanding the genetics of cognition comes from the observation that the human brain, the most complex object we know of in the universe, is produced from a genetic code of only gigabits in length. How, exactly, this works is one of the greatest scientific mysteries to be unraveled. Genomic selection and other “spin-offs” from this work are of secondary interest.

2013-05-13 on infoproc

My opinion can be boiled down to the following (see comments on my 2005 post):

Steven Weinberg says (and I agree) that Charles Sanders Pierce said all that can be said about the problem of free will when he claimed that all that matters is that we have _the subjective experience of choice_.

I’m not sure what Dennett has added to this. I’d almost like to quip that in his work on this subject “what is right is not new, and some things do not even seem right” 😉 But of course he deserves a more careful reading from me before I reach that conclusion.

2013-05-09 on infoproc

Josh, sorry I’ve taken so long to reply to your comments. If you want to know my views on this stuff in more detail I recommend this talk:…

2013-05-06 on infoproc

I’m not a “g man” so I don’t endorse the stronger interpretations of g. But I think it’s obvious that we can crudely measure a quantity that is related to the folk notion of intelligence, and which is stable, predictive (of certain life outcomes) and highly heritable (these are empirical facts). Most people who cite Shalizi’s essay think they are launching a devastating attack on the basic idea of intelligence testing, or genetic causation of quality of brain function. None of this depends at all on whether there is a “unitary g” (i.e., your point (3)). Shalizi’s essay says nothing about the points in my second sentence above, but does reveal his overly strong priors and those of his fans.

2013-05-02 on infoproc

The first few paragraphs (and first link) here might help:…

The point is just that isolated populations tend to cluster together in “genome space” (just as, for long periods of time, they clustered together geographically and reproductively) and these clusters are totally obvious once you statistically analyze the genomes of large numbers of people. That this is controversial is the dual consequence of (1) ideology over science and (2) stupidity (bounded cognition).

Re: African diversity, the “genome space” is high dimensional, so a population can spread out into this space (have lots of diversity) without overlapping with other clusters. The most diverse cluster need not include the other clusters as a subset.

2013-05-02 on infoproc

For the record, while I agree with the opinions expressed in the excerpt above, I do NOT agree with all of the opinions expressed in the articles linked to in your “HBD Fundamentals” list. (Don’t ask me to go into further detail …)

2013-05-01 on geneexpression

Very rough estimate: our sample of smarties might have as much power as 20k random individuals. So if g and height are similar we might get a couple of genome-wide significant hits. We can also test some other hypotheses like mutational load. Finding individual QTLs (e.g., at genome-wide significance level) is not the same as estimating total heritability accounted for by common SNPs, which is what the recent papers are about; neither of those is the same as obtaining a good g predictor, which will probably require millions of individual genomes together with g scores. There are really 3 different problems under discussion, each with different power requirements.

2013-04-25 on infoproc

Yes, every “test (cheating) center” should offer machines with multiple VMs and keyboard toggle that switches the screen 🙂

2013-04-24 on infoproc

R&R did some egregious things in their famous paper, and tried to pass it off as having deep policy implications. It is a black eye for the profession of economics, and this goes way beyond some simple Excel mistakes.

2013-04-19 on infoproc

We actually discussed this earlier in the day. The story is a bit complex. He was a physics major until his senior(?) year, when he switched to math. My impression is that he didn’t really apply himself to things he wasn’t interested in.

2013-04-17 on infoproc

The wikipedia article gives a length definition, but it’s hard to measure accurately.

2013-04-15 on infoproc

Why is the SD only 10? If anything the distribution is broader than normal because you have legacies/athletes/AA-admits on one end and the “top brains” on the other. In fact, you should really think of the Harvard (or Princeton) population in terms of distinct subgroups. There are roughly a couple hundred +4’s in the US freshman pool each year. It is plausible to me that Princeton gets of order 10 of those (Harvard gets more), and those are the kids Bezos is talking about. The reason I posted this is that it really shows the qualitative difference that an honest Bezos (perhaps a +3; I agree with you) detects between himself and the next level up. Of course, you have to have very “g-loaded” material (quantum mechanics) to reveal these differences in a meaningful way.

2013-04-10 on infoproc

“I just wanna ask, was the 2nd conversation between you and Veltman?” yes

“Since when did Veltman claim that his Nobel was “stolen”?” He did this for years. See his samizdat.

For more detail on the history, click the link in the post.

2013-04-04 on infoproc

Sadly they stopped using faculty interviewers some years ago. When I was in HS they sent a full professor of experimental high energy physics to interview me and the other applicant from my school. It was the dead of winter in Iowa and I cannot imagine it was anything like a holiday for the professor.

Modern science is such a competitive grind that faculty interviews are totally impossible these days.

2013-04-02 on infoproc

I’ve never seen any results on age of onset of interest in logic, but your comment raises the point that this study selects for precocity, which is correlated with, but not perfectly predictive of, maximum adult potential. That is, some of the people who qualified for this study are more exceptional for precocity than for their adult cognitive ability. See the Caltech comments below.

2013-04-02 on infoproc

Admissions officers have several tests (SATs, APs, math/phys/chem/informatics competitions, etc.), grades, evaluations, etc. to go on so they have more (good) data than simply one test at age 12. I’d much rather be in their position to pick winners than just having age 12 SATs.

I’m not saying 50% of the class is 1 in 10k — it’s only a fraction. Also the 1 in 10k people in Lubinski’s study are not actually all 1 in 10k due to luck and regression (as explained).

Re: drive and conscientiousness,

“You and Your Research: … At Los Alamos I was brought in to run the computing machines which other people had got going, so those scientists and physicists could get back to business. I saw I was a stooge. I saw that although physically I was the same, they were different. And to put the thing bluntly, I was envious. I wanted to know why they were so different from me. I saw Feynman up close. I saw Fermi and Teller. I saw Oppenheimer. I saw Hans Bethe: he was my boss. I saw quite a few very capable people. I became very interested in the difference between those who do and those who might have done.

… Now for the matter of drive. You observe that most great scientists have tremendous drive. I worked for ten years with John Tukey at Bell Labs. He had tremendous drive. One day about three or four years after I joined, I discovered that John Tukey was slightly younger than I was. John was a genius and I clearly was not. Well I went storming into Bode’s office and said, “How can anybody my age know as much as John Tukey does?” He leaned back in his chair, put his hands behind his head, grinned slightly, and said, “You would be surprised Hamming, how much you would know if you worked as hard as he did that many years.” I simply slunk out of the office!

What Bode was saying was this: “Knowledge and productivity are like compound interest.” Given two people of approximately the same ability and one person who works ten percent more than the other, the latter will more than twice outproduce the former. …”…

2013-04-01 on infoproc

A lot of this is context/society dependent. If you look at the Clark-UK/China data it seems plausible that in the past smarter people who were more economically successful also reproduced more. Today that’s not true for a number of reasons.

Re: Kanazawa, agriculture and trading/specialization were evolutionarily novel at first. Then we went through a period where most people were involved in agriculture and/or had specialized roles in society, and their success had “fitness” (defined reproductively) consequences. That period was long enough to have effects on allele frequencies.

2013-04-01 on infoproc

In earlier work it was shown that these 1 in 10k kids regress a bit by the time they are high school seniors. This is expected since they use a single test for admission and some of the kids who qualify are a bit below the cutoff but got lucky when taking the SAT at age 12. The typical HS SAT scores for this population (obtained pre-1995 when the ceiling was higher than today) were somewhat lower than 1 in 10k, although, IIRC, above 1 in 1k. Based on this I’d say a good fraction (although less than 50%) of the Caltech class is at the same level, roughly, as this group. Recall about 40% of techers go on to earn STEM PhDs, typically at good places, which is not very different than this group. Techers might have been selected a bit more for conscientiousness.

2013-04-01 on infoproc

I don’t recall seeing it, but Lubinski might know. They have tons of data on this population.

2013-03-31 on infoproc

What did I leak? 🙂

2013-03-24 on infoproc

You might start here:… , esp. the PCR section. I’m not sure what’s published but this is something that needs to be fine-tuned in the lab. If you have a serious interest in this you can email me.

2013-03-24 on infoproc

Who should I trust: a journalist at Technology Review, or people who are actually doing this kind of thing in the lab? 😉

2013-03-22 on infoproc

I only know what the experts tell me about this. But if you google ivf freeze embryo the results seem to support what I wrote.

2013-03-22 on infoproc

The story I heard could be apocryphal, but I’ve heard it many times, including in Birge Hall 🙂

2013-03-21 on takismag


Thanks for your nice summary of my talk. The slides you link to are from an earlier version — here are the ones I used in the video:…

Best wishes,


PS The audio level is bit low but should be OK if you use earphones.

2013-03-18 on infoproc

As I mention at the end of the talk, the reproductive technologies are already here — it is already possible to sequence a zygote from a single cell. Freezing fertilized eggs does not seem to damage their viability. The main thing that is lacking is the ability to predict phenotype from genotype, which is necessary for pre-implantation selection. My rough guess is that it will take GWAS samples of order 10^6 to determine the effect sizes which are the necessary inputs for genomic prediction. At the moment we are far from this level of statistical power (two orders of magnitude short!), but the situation will likely be different in 5-10 years.

2013-03-11 on infoproc

99.5 is just an arbitrary value, as you can see from the last graph. But there is a qualitative difference between the 1% (as conceived by OWS) and the “real” 1% who are perhaps actually the 0.1% (or whatever). A lot of the 1% are older people near retirement who have simply saved all their lives on fairly ordinary incomes.

The Real Wealth essay I link to was written by someone else but his points are well taken.

2013-03-01 on infoproc

You’re right — I think Valve is in the Seattle area. I liked the O’Neil podcast as well, although I agree with some of the comments on the econtalk site that find her a bit naive at times.

2013-02-26 on infoproc

I agree — it is the rational mind that trains the intuitive (Kahneman) instinct to achieve the type of deep intuition that physicists refer to. One problem I often encounter is that people who have developed deep intuition for a particular topic are not then able to communicate the rationale behind their hunches to colleagues who have not spent as much time on that particular area or set of problems. Unpacking the developed intuition is quite difficult. I am always on the lookout for people who can explain their heuristics to others.

2013-02-18 on infoproc

Oops — I didn’t realize it was already published. A colleague sent me the link above and I didn’t realize it was to an old draft version.

2013-02-15 on infoproc

I am unqualified to answer your question. But at the 10% probability level I wouldn’t be surprised if we someday found a cognition enhancing drug.

2013-02-14 on infoproc

This paper is from 2009 — they used only 54 SNPs for height. The number of genome wide significant hits today is approaching 1000. Also, there are applications of genomic prediction for which parental midpoint is useless, such as differentiating between zygotes (PGD).

2013-02-13 on infoproc

We are getting sequences back right now. But there was a problem with DNA extraction from a subset of samples, which caused a delay. If you are in the study you will be getting an email soon with more details.

One interpretation of the Visscher-type estimate of heritability is that it is common variants causing the phenotype variation. Statements to this effect often appear in his papers; the logic is that low frequency variants are only in weak LD with common SNPs, so they can’t account for the observed phenotype-relatedness correlation.

Another interpretation (not completely ruled out, AFAIK) is that relatedness as measured using SNPs is correlated with the presence of deleterious variants: two individuals who have higher SNP relatedness are also more likely to be similar in their level of deleterious rare variants. That might account for the measured heritability even if the actual causal variants are not directly detected by the SNP chips. Time will tell.

2013-01-28 on infoproc

Because luck plays such a huge role in discovery (and life!), even most +5s will not be remembered by history. The most accomplished and the smartest are usually not the same individual: accomplishment and smartness are only imperfectly correlated.…

vN was “smarter” than Einstein, but the latter made larger contributions to science.…

2013-01-28 on infoproc

Here are bios of some of the “minor figures” in the early development of the computer (minor relative to, say, von Neumann or Turing). I’d guess these guys are +(3-4), with vN and Turing way beyond. Every new hard technology has to have people like this around to make progress (Schwinger and similar types made very important contributions to waveguides and radar, etc., etc.). Multiple directions are investigated (each by competing teams or companies) for every one that succeeds. Only a few superstar (or highly memorable) figures make it into the popular histories, but you can be sure there are lots of guys like these contributing who remain relatively unknown.…

Guess how many +2,3,4’s are involved in this:…

or this:

Is your question “Why don’t +4,5s stand out among +2,3s?” (they do) or “Where are all the +4,5s?” (100k of them spread around the world, but concentrated in certain places …)

2013-01-28 on infoproc

Because the distribution from +3 to +5 is filled out uniformly, they don’t stand out that much *from each other* — in the same way that top WRs in the NFL aren’t that different from each other, but many of them today are bigger, stronger, faster than, say, Jerry Rice was.

Re: Elon Musk, I don’t find what he is doing nearly as impressive as the work of the invisible to the public teams of engineers and applied physicists who keep Moore’s Law and its equivalents in storage, bandwidth, DNA sequencing, etc. going year after year. The actual underlying technologies for, e.g., chip fabrication or DNA sequencing, change on roughly 5 year timescales, so whole new fields have to be invented with regularity to keep these cost-capacity curves going decade after decade. See, for example, “spintronics”.


2013-01-25 on infoproc

This is sort of how I think of it. At least the top 5 or so kids per graduating class are +4ish (top by brainpower, not necessarily GPA, although the two are correlated) — this is even consistent with test scores in the old days when the SAT ceiling was pretty high. The top couple of kids on campus at any moment might be qualitatively different from these +4s — maybe they’ll go on to do something really special 😉…

2013-01-24 on infoproc

Perhaps we should have a referendum on whether Gide (and related avatars) add anything to the discussion here …

2013-01-24 on infoproc

We can address standing out among +3’s very easily. The average Caltech kid is around +3. The difference between the top kid in the class and the average kid is huge. If you want to compare +4 vs +5 that is hard because there aren’t many of the latter to go around. But I do know people who are considered “scary smart” even among +4 types.

There are definitely “clear standouts”, but I suggest that a +2 can’t really tell the difference between, e.g., a +4 and a +5 very easily (only indirectly).

(Note this discussion should not be taken to imply that I’m a “g man” or whatever. Brainpower can only be measured in a crude way, and the distribution is probably not exactly Gaussian in the tails …)…

2013-01-22 on infoproc

B&D is also there (yellow paperback) on the third shelf, not far from Wald. I have many more books but still in boxes … I have most of Landau and Lifshitz but never got in the habit of using them much. You’ll note I don’t have Jackson or Goldstein or very many math books on the shelves yet … I notice someone has kept the copy of Nielsen and Chuang I lent them; probably lots of other missing stuff 🙁

2013-01-18 on infoproc

Very few Chinese developers in China were educated in the US. Most graduate students admitted to US programs stay here, and the ones that go back tend to take more senior positions.

2013-01-16 on infoproc

I think it’s in a box somewhere…

2013-01-10 on infoproc

In the interview about his new book that I heard I thought he stuck to his 2020 prediction, but I might be mistaken. The interview is actually not bad:…

2013-01-10 on infoproc

Singularity <> machine learning and related areas. These are core subjects in computer science and modern statistics.

If you think the Singularity is near (i.e., will happen by 2020), then I lump you in with Kurzweil.

PS At least as far as I can judge from a recent interview, Kurzeil’s latest book is less nutty than The Singularity is Near, although see the review at the link:…

2013-01-07 on infoproc

Any genomic data CG has is covered by IRB or HIPAA or other legal protection. BGI isn’t doing this to get at any specific genomic data; they have plenty of their own.

2013-01-07 on infoproc

I’m guessing the 4 min rest doesn’t contribute much to the overall calorie expenditure. If you did the 2.5 min of sprints tabata style you’d probably burn about the same number of calories. Note, even with HIIT/tabata you need to warm up properly to avoid injury, so in the end I can’t really get a decent workout in less than 10-20 min total time.

2012-12-26 on infoproc

Haven’t talked to the director yet. Here’s the only review I could find:…

2012-12-25 on infoproc

The great thing about “in terra pax, homínibus bonæ voluntátis” is that almost anyone, regardless of religious belief, can endorse it! Happy holidays 🙂

2012-12-20 on infoproc

Stay tuned — we recently started sequencing our samples!

2012-12-18 on infoproc

I like Dennett but was underwhelmed by Consciousness Explained when I read it many years ago. Still, the conversation you ask for would be interesting.

2012-12-16 on infoproc

The slide says “Shift SOME hiring …” — SOME, not ALL. We’re talking about marginal dollars here, not the bulk of hiring. Also, you have to remember what my portfolio (Research) is, as opposed to a dean, who has responsibility for training undergraduates, etc. MSU has historically been run along the lines you seem to favor, but the goal now is to increase research intensity here. Note “increase” <> “prioritize to the exclusion of all else”!

2012-12-15 on infoproc

In your system researchers might be efficiently priced throughout the seniority range, but here I think there is an arb. Beginning profs and senior profs are correctly priced, but often universities don’t, as my trader friends would say, bid back properly against offers for people in the, e.g., 4-8 year seniority range.

2012-12-15 on infoproc

I’ve heard a couple of Hawkins’ talks and tend to share some of his views on AI. IIRC, Numenta’s technology is a brain architecture inspired neural net algorithm.

2012-12-15 on infoproc

If your team knows how you think they can anticipate what kind of evidence you will want, what your decision is likely to be, etc. That lets them plan better and wastes less effort. If everyone knows that decisions will be made rationally according to transparent metrics, there will be less BS and positioning. There are lots of crazy things happening in any big org, and it’s good for people to know that demonstrably nutty things, if brought to the attention of the leadership, will be stopped.

2012-12-13 on infoproc

These activities are for fund raising. The donors we met with in NYC are supporting research and teaching activities at MSU. Tuition is up mainly because state support is down (at public universities).

I personally could make much more money in the private sector, and I certainly would if I didn’t care so much about the mission of research universities.

If being educated at a research university is not a big consideration, one can certainly get a degree at a less expensive university. YMMV.

2012-12-12 on infoproc

Church is correct. At the moment there are probably reasons to store the raw data because of sequencing errors, etc.

2012-12-02 on infoproc

Re: Brian, no, I never did. I don’t know any physicists who train MMA or BJJ, but there must be some.

There are at least a few who wrestled in college. At UO there was a grad student who had been a pretty good 197 pounder in college and I wouldn’t want to have to fight him (we used to train together sometimes). But he’s not a theorist 😉

PS Any thoughts on Unz’s article? You can tell me in private …

2012-11-28 on infoproc

I’m referring to prizes *won* by former students (alumni). Click the link and read.

2012-11-28 on infoproc

NMS Semifinalist is based on PSAT/SAT (and GPA?) and is not a particularly high filter. Just look at those numbers.

Don’t know how well Ron did with his AJ counting.

2012-11-28 on infoproc

NMS Semifinalist is based on PSAT/SAT (and GPA?) and is not a particularly high filter. Just look at those numbers.

2012-11-28 on infoproc

MIT has been aping the Ivies for almost two decades now. It began with a famous comment by their president to the effect that too many MIT grads work for Harvard grads. MIT admissions policies have deviated from a pure academic meritocracy. It’s not just geographical factors that are causing the difference between MIT and Caltech.

Also, as you may know, Asian Americans are very ambitious and tend to apply nationally, not just locally, to university. You can look at *accept rates* at places like Harvard or MIT and they show the disparities that Ron discusses. I suggest reading his article before making further comments.

PS You are missing the whole point of the graph. Why has Harvard Asian percentage not increased given the large relative increase in size and quality of Asian American applicant pool?

2012-11-24 on infoproc

I think there’s good evidence for relatively stable non-cognitive factors such as Big 5 personality traits. The tests for these are noisier than for g. Once this inaccuracy in measurement is accounted for, the heritability of non-cognitive factors is almost as high as for g.

2012-11-17 on infoproc

As an SV guy, DC gives me the hives. No value creation here, but lots of skim.

2012-11-14 on infoproc


That was a little joke — see Ginzburg’s comments at the Out on the tail link.

2012-11-14 on infoproc

> it’s not easy to be self motivated when you get lost after couple of minutes of lectures, on videos you can repeat the important concepts as many times as you like< Interesting observation! But it is clear you are above average in motivation.

2012-11-11 on infoproc

There are really two factors at work here:

1. Once you have power, you may be more tempted to do nasty things. (Power corrupts, etc.)
2. Individuals who manage to climb the power ladder (who have the drive or cunning to do so) may be more likely to be sociopaths.

Re: 2, IIRC one study has the rate of sociopathy among CEOs at about 4x that in the general population.

2012-11-10 on infoproc

Common practice these days, including at most NFL camps. I’ve never tried it myself.…

2012-11-09 on infoproc

IIRC, in the Perelman bio Perfect Rigor it is stated that because of his Jewish heritage Perelman would not have been admitted to Leningrad State University had he not made the USSR math olympiad team. Anti-semitism of this sort does not preclude some big successes like the ones you list. I know that some physics institutes in the USSR did not hire Jews and there was fierce rivalry between certain “Jewish” schools of physics and others.

2012-11-09 on infoproc

Sure, that’s why left-leaning commentators kept saying it was “neck and neck”, “too close to call” all through election day …

2012-11-08 on infoproc

Just as personality factors are largely independent of cognitive factors, the characteristic of being “epistemologically careful” (someone who reasons cautiously, with minimal bias) is only imperfectly correlated with g. I know many physicists who fall in love with their theoretical constructs and lose the ability to reason dispassionately about whether they actually describe Nature. The problem is much, much worse in fields that study complex, strongly-interacting systems. “All priors, All the time”!

2012-11-08 on infoproc

There’s no hard threshold. Even someone who doesn’t understand what a standard error estimate is can still appreciate “quant magic good” and thereby improve their own predictions about the future. Interestingly, my impression is that elite education at, e.g., HPS is good because many HVLM types who attend these schools will develop a healthy respect for quant magic (or respect for HVHM thinking). This doesn’t always happen and some HVLM types are dismissive of quant methods throughout their entire lives. One positive thing about this election is that both candidates were people who could appreciate (if not perform) quant analysis.

2012-11-08 on infoproc

It’s not just TV pundits. Look at David Brooks (or any NYTimes columnist), who clearly is a high V modest M type.

2012-11-08 on infoproc

Go back to the original High V Low M post and look carefully at what I wrote about V. High V High M and High V Low M are two very different phenotypes. See also the SMPY / SVPY papers.

“High verbal ability is useful for more than just impressing others — it typically implies a certain facility with concepts and relationships between ideas — but high V alone is a dangerous thing. The most confused people I meet in the Academy tend to be high V, low (modest) M types.

More on the V / M split in this longitudinal study of gifted children (SMPY / SVPY — see esp. figure 4).”

2012-11-08 on infoproc

New technology, not yet well understood … 🙂

2012-11-08 on infoproc

Nice graphic!

2012-11-08 on infoproc

It’s not just the High M crowd that moves markets …

2012-11-07 on infoproc

Central Limit Theorem = black magic for our punditocracy (idiocracy).

2012-10-27 on infoproc

De novo mutations with big negative effects on IQ never show up in any heritability studies because individuals with such mutations would be removed (or never recruited in the first place).

De novo mutations with small negative effects on IQ would be lumped into unshared environment, as you say. My guess is that this is not a very big effect.

2012-10-25 on infoproc

What fraction of academic economists are “frontier” economists? You might be surprised to know that I talk to a lot of economists and most are not the frontiersmen you refer to. But maybe the ones I know are too old …

Is the Shiller quote below (2008) now completely out of date?…

“I BASED my predictions largely on the recently developed field of behavioral economics, which posits that psychology matters for economic events. Behavioral economists are still regarded as a fringe group by many mainstream economists. Support from fellow behavioral economists was important in my daring to talk about speculative bubbles.

Speculative bubbles are caused by contagious excitement about investment prospects. I find that in casual conversation, many of my mainstream economist friends tell me that they are aware of such excitement, too. But very few will talk about it professionally.

Why do professional economists always seem to find that concerns with bubbles are overblown or unsubstantiated? I have wondered about this for years, and still do not quite have an answer. It must have something to do with the tool kit given to economists (as opposed to psychologists) and perhaps even with the self-selection of those attracted to the technical, mathematical field of economics. Economists aren’t generally trained in psychology, and so want to divert the subject of discussion to things they understand well. They pride themselves on being rational. The notion that people are making huge errors in judgment is not appealing.”

2012-10-23 on infoproc

” almost irrelevant for most of those 500ky.”

Almost zero and zero are not the same. From the fossil record we know that evolution typically works slowly on generational timescales — selection pressure is generically small in the units you prefer. There are probably many examples (across many species) of traits such as body size and morphology, controlled by existing polygenic variance, that changed slowly over time (e.g., on 1E06 year timescales). I don’t know why you find it implausible that intelligence is similar. (Note, evolution that requires de novo mutations could be slow for other reasons, such as the improbability of advantageous mutations.)

“Therefore there’s no reason to presume that the selection pressure towards higher intelligence was there at all.”

Except that it is very plausible that we are much smarter than our hominid ancestors. Or perhaps you are a creationist?

2012-10-23 on infoproc

Selection pressure has to both remove de novo mutations and improve average genotype. So the rate we are talking about is net of de novo mutations. I think you should probably quit now.

2012-10-23 on infoproc

“going down by O(30) every 50 years or so as long as intelligence was under selection”

You seem to be making strong assumptions about the strength (rate) of selection. If our ancestors 500k years ago were on average -3 SD relative to modern humans they would have 1100 vs 1000 (-) alleles and perhaps it took 500ky to reduce the average number of (-) alleles by that 100.

2012-10-20 on infoproc

Certainly a big advantage to know a priori that the problem has a solution. But, for example, it is claimed that the French had a hard time with the thermonuclear step and actually required outside assistance.

2012-10-19 on infoproc

Perhaps Teller is the one who realized that light from a spherical source is easily refocused or reflected onto a spherical target. Interestingly, the Chinese design supposedly uses x-ray lenses to focus the radiation, as opposed to reflection from a heavy metal.

The whole thing is a bit mysterious, probably because some of the “facts” we are going on are actually disinformation.

Re: Ulam vs Teller, this is a good article: How Ulam Set the Stage, Bengt Carlson, Bulletin of the Atomic Scientists 2003 59: 46 DOI: 10.2968/059004013

2012-10-14 on infoproc

Eysenck does not refer specifically to Ericsson but he presents data that contradicts Ericsson’s hypothesis. Perhaps the largest data set that undermines Ericsson is the SMPY longitudinal study in which top 1% individuals can be compared to top 0.01% individuals over an extended period of time. These individuals are selected before age 13, and the g measurement at that early stage in life has clear predictive power for achievement later in life.…

2012-10-07 on infoproc

James, I received a copy of your Singularity book in the mail and have been skimming it. Very nice!

2012-09-30 on infoproc

Deans / Dept. Heads have to cover the salaries but the overhead monies often end up (depending on the specific organization of the school) with the VP of Research or perhaps the Provost. So incentives are not aligned, usually. Also, incentives for individual faculty (who have the specific domain expertise) to go out and identify talent are usually lacking. Some faculty are team spirited and want to improve their department, but others aren’t sufficiently motivated to do this. It takes a lot of effort and coordination between different players to get this to work.

I should also add that this strategy has to play out over a number of years and over that timescale the personnel and organizational structure may change.

2012-09-30 on infoproc

“… you would be left with a relatively larger and more prestigious research effort when the new equilibrium set in.”

Your summary is exactly correct, which places your understanding above the 90th percentile among faculty and senior administrators at research universities.

“… it is the highest paid employees that contribute the most value in excess of their pay”

At a well-functioning place this is true but I’m not sure it’s always true. Most faculty (who are typically somewhat left of center like me) support differential pay and aggressive recruitment of top scholars. It’s only the far left fringe who oppose this and they tend not to be in the hard sciences.

2012-09-29 on infoproc

Overhead on research doesn’t actually cover the full cost of the work, but the details are complicated. (Donors, state taxpayers, etc. are subsidizing the Feds in producing the national research output.) Nevertheless, using the de facto utility function governing modern research universities, more high quality research has positive utility, as does increasing the quality (impact) of faculty. That is, most university presidents would be very happy if research expenditures and faculty prestige were to increase.

In more practical terms, the overhead on (marginal) research dollars can be used to fund desirable things like improved infrastructure, faculty pay, more student fellowships, etc.

2012-09-27 on infoproc

In the Moneyball analogy, Harvard is more like the Yankees than Oakland 🙂

But, yes, they acquire most of their senior faculty this way. However, they are probably looking a bit later than +(5-10) years into the career. At that later stage there there is broader agreement on the impact of an individual researcher and the market is closer to efficient (i.e., you pay for what you get, but Harvard and other top schools can afford it).

2012-09-27 on infoproc

I hope the results hold up, but it does seem that these guys got very lucky to choose an initial set of 775 SNPs that happened to contain 240 or so that control most of the variance for autism.

2012-09-18 on infoproc

Yes, it appears to have stopped, at least according to data from the UK and a few other countries I have seen.

2012-09-16 on infoproc

Click through and Zillow will tell you. About $7k per yr.

2012-09-15 on infoproc

Investment in near-term R&D is often well-supported through the profit motive. However, work that is not likely to pay off for a decade or two is much harder to get funded, because it’s hard to predict who will reap the eventual payoff. Thus, long term R&D depends on more abstract principles for its support (e.g., the future good of the nation or humanity). I would guess that almost all societies (hence my use of Civilizations) allocate too few resources to basic research and even to applied research whose payoff is 10-20 years in the future.

2012-09-15 on infoproc

We impoverished Americans welcome our new mineral-rich overlords! 😉

2012-09-15 on thechronicleofhighereducation

We found very strange results in the CIS data. We went to some professors in the department who told us there are very different paths through the major, of varying difficulty. We asked for a particularly rigorous subset of courses, which is how we ended up with a small sample size in that major. But still the correlation with SAT was small.

Having run software startups I can tell you that CS (or CIS as it is called at Oregon) is an odd subject. Many good programmers are not particularly high scorers on, e.g., the SAT but rather it is their interest in the subject that makes them good. (My opinion: except for very high end niches the subject is not as “g loaded” as physics or math.)

The predictive power of SAT needs to be studied in more detail. It should not be difficult now that all student records are digitized. Schombert (my co-author) and I did this in our spare time using widely available web technologies like Python. If you have decent computing skills you can do this at your own university — just ask the central admin for access to student records.

For the real scholars and scientists on this thread, some links:…………

2012-09-15 on thechronicleofhighereducation

“… predictive power of SAT ranges from 12.25% to 25%”

I think you are referring to percentage of variance accounted for. I don’t think this captures what ordinary people mean when they use the term “predictive power”.

Anyone looking at our scatterplots would admit that the correlation is significant — especially in a social science setting, where, e.g., r = 0.4 is pretty high. In fact, as we say in the paper, knowing an applicant’s SAT score (and nothing else) gives you about as much predictive power as knowing their HS GPA (and nothing else). I would call that an *extremely* powerful predictor, given that it only takes a few hours to administer the SAT. Combining SAT and HS GPA to form a more powerful predictor (which is essentially what admissions offices do) yields correlations of almost 0.6. That is, there is clearly additional predictive power obtained from the SAT even if HS GPA is already known. The contrary claim is often made but it is untrue.

You are not surprised at the 90th percentile threshold for physics and math (and probably other math-intensive fields such as EE); neither were we. However, note you are admitting that a simple test like SAT-M can identify the 90 percent of the population that have *very low probability* of succeeding in quantitative STEM fields.

PS Our criterion for “mastery” of the UG curriculum in math/physics was quite modest and did not require differential geometry or general relativity. We only required a slightly above average GPA (IIRC, +0.5 SD) in the usual UG courses in those majors.

2012-09-14 on thechronicleofhighereducation

Using 5 years of complete undergraduate records at the University of Oregon, we determined that SAT correlates about 0.35 to 0.5 (varying by major) with upper-division, in-major GPA. The correlation is *higher* than exhibited for first year GPA, because freshmen typically choose courses of varying difficulty according to their academic strength (e.g., multivariable calculus vs basic algebra). By junior and senior year (and after segregation by major) this effect is less pronounced, and indeed SAT is a strong predictor of upper division performance.

The scatterplots in the paper should be enough to convince anyone that the correlation is real and significant. Our study was not funded by the College Board.

We also find a non-linear threshold for mastery of the undergraduate curriculum in physics and pure mathematics, at about 90th percentile on SAT-M. Success in these majors is extremely unlikely for students who score below this threshold.

Those who doubt the utility of the SAT in predicting college performance have not looked at the data.

2012-09-13 on infoproc

Nice catch. I listened to the Lewis (audio) interview last night while falling asleep and made a mental note to track down the probability reference in the article, because it seemed to be a pretty sophisticated observation for a politico. But I guess I messed up the attribution 😐

PS I recall Robert Rubin making a big deal out of this “thinking probabilistically” as a big competitive advantage for his event arbitrage group at Goldman. It seems like half the book is his extolling this kind of strategic thinking in finance, politics and foreign policy. I found it surprising that sophisticated people could approach strategy in any other way ……

2012-09-13 on infoproc

Oops, I guess I failed to notice where the quotes ended. I thought that Lewis, in the audio interview, attributed the comment to Obama, but I could be wrong.

Re: Obama hate, I don’t get it. Lewis has always (in my judgement) been a straight shooter and I’ve been reading his stuff since Liar’s Poker.

Obama’s reasoning sure beats the Gog and Magog logic of his predecessor!

2012-09-07 on infoproc

You won’t learn much other than historical trivia. Better for non-experts.

2012-08-27 on infoproc

Thanks! Fixed.

2012-08-24 on infoproc

Apparently the biologists who wrote the PNAS commentary agree with the physicists.

2012-08-19 on infoproc

A biologist writes in to correct Mandelbrot’s comments on Delbruck:


“Delbruck, who knew no biology” is absolutely not true.

Delbruck learned plenty of biology (including Drosophila genetics) in Berlin, before he ever got to the USA. He, along with other physicists from Bohr’s circle who got interested in biology, learned it mainly from an illustrious guy called Timofeev-Ressovsky (…. In fact, they published a very classic paper together in 1935 (… ), in which they, naively aping Rutherford’s experiment, estimated a size of the gene (… ). The idea was actually Timofeef’s – the initial experiments were done before Delbruck came to town.

Interestingly, Timofeev-Ressovsky was a pupil of Koltsov, the guy who in 1927 wrote that genes are inherited via “two mirror strands that would replicate in a semi-conservative fashion using each strand as a template”. Apparently, it was Watson who brought the idea of the template-based gene propagation to Crick. Watson denies ever hearing of Koltsov – almost certainly true. But also almost certainly true is that Watson, being Luria and Delbruck’s protege, heard the idea from Delbruck who, in turn, got it from Timofeev-Ressovsky (who was always adamant that Koltsov’s insight had to be at least part true).

Also, before Delbruck was hired by Caltech in 1947, he spent 8 years in the USA doing all kind of biological research, some of it with Luria (including the famous 1943 Luria–Delbruck experiment). It is inconceivable that Luria, being a medical doctor by education, had not contributed to Delbruck’s understanding of biology.

I guess that it all goes to show that we shouldn’t buy too much into the idea that a bright physicist can get in, teach himself biology by teaching some 101 course and make several seminal contributions as a result 🙂


2012-08-17 on infoproc

Software start up success has a big noise component; possibly borderline noise dominated. Academic physics less so but even there luck is important.

2012-08-16 on infoproc

I think Taleb is writing for people of above average ability (i.e., himself). In Fooled By Randomness he expresses bitterness at the “lucky fools” who made more money than he did. BTW, most practitioners I know think Taleb was a failure as a trader.

2012-08-10 on infoproc

If Bolt trains seriously for it my money is on him. Sub-44?

2012-08-08 on infoproc

Just an aesthetic choice — do you prefer that Psi is real, or do you hate the existence of the other branches? But there are testable consequences, at least in principle. See Deutsch experiment; if you choose the one branch option at some point there are some macro superpositions not allowed.

2012-08-08 on infoproc

The slides labeled “Born Rule” at the link might help:…

2012-08-07 on infoproc

I’m not an expert on Bohmian models, but IIRC they have problems with relativity and fermions. In other ways (determinism) they may be similar to MW.

2012-08-07 on infoproc

When you are young a lot happens in a year and it seems like a long time. When you are older it seems years go by much more quickly.

2012-08-07 on infoproc

It’s possible both Blake and Bolt are dopers. I’m assuming Bolt is not, but I don’t know (for that matter I don’t know about Phelps, and didn’t know about Armstrong for many years). If Blake is a doper he’s comparable to Ben Johnson in talent level — without the steroids he’d be good but not great. The fact that the Jamaicans (both M and W) have been so successful in the sprints in the past few years makes me very suspicious. In any case, I am confident that there are very few males in the US, Europe or Asia (for example) that can run under 9.7s without drugs…

2012-08-06 on infoproc

Silver: “… an athlete with the perfect swimmer’s build and a world-class work ethic would still stand little chance of competing in this year’s games if he happened to be born in a poor nation like Cameroon or Panama — he might never have gotten into a pool, let alone an Olympic-size one. But running, especially over short distances, can be practiced virtually anywhere and anytime.”

On this specific point Gould is correct.

2012-08-06 on infoproc

>> Did you ever engage is something more challenging than running or swimming, something like downhill skiing? << Other than football, Judo and Brazilian Jiujitsu? I guess not.

2012-08-06 on infoproc

I think Silver also believes Bolt >> Phelps.

2012-08-06 on infoproc

Sprinting ability cannot be developed that much. Anyone who is remotely capable of sub-10s is a freak and knows it from an early age. Even people who are capable of sub-11s are freaks by the standard of HS sports (probably only 50% of NFL WRs could go sub-11s), and already get massive positive reinforcement for sprinting. Sub 9.7 is on another level entirely.

I knew a lot of kids in Iowa who desperately wanted to improve their 40 times but even with lots of training (weights, track sprinting, plyometrics) could only make marginal progress.

Sprinting ability, explosiveness, etc. are close to a “general factor” of ability in certain sports (FB, BB, soccer, etc.), and Bolt has, perhaps, as much of it as any other living human.

2012-08-06 on infoproc

The time we have left is going to pass very quickly 🙁

2012-08-06 on infoproc

See my post from 4 years ago. Bolt is #1 in a much deeper talent pool than Phelps. Number of people who know whether they have potential to be world class in 100m dash is much larger than the number who engage in enough skill-development to have a chance in swimming. I’m a former competitive swimmer so I know what I am talking about.

2012-08-05 on infoproc

It is very surprising to a physicist that there are whole fields in which people had given up the idea of causation and retreated wholly to correlation. We work in exactly the opposite limit!

I think Pearl’s contribution is conceptual clarity — adopting his language makes it easier to communicate about causal models. This is necessary in the complex contexts that arise in social science, but not so necessary in physical science. However, I would be hard pressed to characterize any of this as a real breakthrough in epistemology or the scientific method.

Despite what some people (overly optimistic Bayesians, AI enthusiasts, etc.) might think, the scientific method (i.e., discerning the underlying causal rules in Nature) has yet to be fully systematized and probably cannot really ever be fully mechanized. (How does one systematize selection of priors?) Even “rationality” has never been satisfactorily defined.

See link for related discussion. I highly recommend Eric Baum’s What is Thought?…

2012-08-05 on infoproc

I think Dyson is talking about cooperation pumps in the last quote about “double selection” (within and between tribes).

I agree that the definition of “evolutionary opponents” is a bit problematic (although clearly PD are aware of this).

2012-08-02 on infoproc

He’s very strong mathematically, and started out as a mathematician, but became a physicist when he came to the US. I would say his way of thinking is much more typical of a theoretical physicist than a mathematician. Other than his work in number theory as a youngster, he’s pretty much exclusively worked in physics.

2012-07-30 on infoproc

Although the optimal play between two smart (ZD-aware) players may resemble TFT in some respects, the dynamics of the game are totally different from what generations of people who studied it had thought. ZD negotiation, perhaps leading to cooperation, is a form of iterated ultimatum game.

Here is what Press says in the Edge comments:

“This sounds a bit like TFT, but it is actually quite different. First, while TFT guarantees both players the same score, it does not guarantee them both the maximum score. In fact, both players playing TFT is an unstable situation allowing any possible (mutual) score, while the above “treaty” stably awards both players the maximum. Second, TFT plays out on a rapid, move-by-move timescale. There is no way to pause for reflection or negotiation without losing points irretrievably.The ZD treaty regime instead allows for a whole range of careful, deliberative negotiations. You can never change your own score by unilateral action, and you always retain the future ability to extract any specified penalty from your opponent. That is a world in which diplomacy trumps conflict.”

2012-07-19 on infoproc

PS I just checked the paper cited in the earlier post……

There it is claimed that Europeans became aware of the Chinese examination system as early as the 16th and 17th centuries, and systematic written examinations did not appear in Europe until the 18th century (written Tripos started around 1750). It is implied that even the Tripos might have its origins in the Chinese system.

For example:

As early as 1755 there is an article about China in the Gentleman’s Magazine which reads as follows:

“Writing is the only Test by which a Man of Sense desires to be try’d…. All authors agree that the Chinese excell all other Nations in the Art of Government…. Their Honour and Titles are not Hereditary; … The Mandarins are chosen once a year at the Metropolis of China.3”

… On the authority of Isaac Vossius (Variarum Considerationum, 1685) and the Jesuit missionaries, BUDGELL recounts in detail the systems of competitive examinations …

2012-07-19 on infoproc

Galton was well aware of competitive exams as a filter for talent. He had a nervous breakdown preparing for the Cambridge Mathematical Tripos examination. University positions in maths were primarily awarded to Wranglers (winners of these competitions).…

But I think the use of exams more broadly in England did not arrive until later. I don’t know whether Galton ever advocated for civil service exams or for the Chinese model, but of course he did play a key role in the development of g 🙂

“When Galton eventually entered the larger realm of education beyond his home at age eight, he discovered that there were others who could perform academically better than he. During the next ten years he watched his scholastic aspirations for greatness evaporate. Although he did well, he failed to earn the highest honors. This phenomena was inconceivable to him, since he had grown up with all the social advantages. He began searching for an alternative rationale for his limited greatness. The conclusion he eventually drew was that there must be some innate difference between those whose achievement went beyond his and himself. The ground work was set for his later work in intelligence theory and testing.”


2012-07-19 on infoproc

Hmm … where did all these meritocratic exams come from in the first place?…

2012-07-16 on infoproc

You can’t use US numbers for real medical costs — our system has huge distortions. Few $K is the cost in Taiwan or Korea and success rates are if anything higher there. That’s not even factoring the economies of scale that would arise if a large fraction of couples wanted it.

Who says the US is the first market for this?

2012-07-15 on infoproc

On the other hand some governments will be slow to adopt (even legalize) this technology, so in those countries only the rich will have access.

There’s a part of the Blair Underwood (genetic counselor) scene in Gattaca that was cut in which he asked Ethan Hawke’s parents whether they want genetic enhancement for Hawke’s little brother — e.g., alleles for music or math ability that aren’t present in the parental DNA. But the director left this out — possibly because it changes the nature of the future world the movie is set in. If that kind of enhancement were possible the engineered people (with whom Hawke competes for the space navigator job) would be inconceivably superior. There would be essentially no overlap between the distributions for engineered and non-engineered people, whereas in the film it is clear there is still substantial overlap. I think you can find this clip on YouTube.

2012-07-15 on infoproc

They’ll have the genotypes of both parents in all likelihood, so trivial to check that the cell comes from their zygote.

2012-07-15 on infoproc

You can’t make the leap all at once for obvious reasons (e.g., a +6 SD genotype is still +6 SD improbable until direct genetic manipulation becomes possible), although keep in mind that more than 10 eggs can be extracted over multiple cycles, frozen, and then fertilized at a later date.

Imagine what a couple might pay to ensure that they get the best out of 10 or 50 possible offspring, optimizing over their choice of heritable attributes. Compare this with the cost of a Harvard education or K-12 private school tuition. The cost of an IVF cycle is down to a few thousand dollars and could go even lower.

Genetic prediction at high accuracy will probably be possible once of order millions of genotype-phenotype data pairs are available for analysis. I predict about 5-10 years. The advance in the Nature article makes me confident that the necessary reproductive technologies will also be available.

I hope that progressive governments will make this procedure free for everyone. The benefits from increased economic output, decreased welfare and criminality rates, etc. far outweigh the cost of what I have described above ( = few cycles of IVF + running my algorithms provided at dirt cheap licensing rates 😉

2012-07-09 on infoproc

Ultimately “genetic diversity” is just a proxy for things we really want to know, like the distribution of actual causal variants.

Although no one knows for sure, it seems likely that the vast majority of rare variants will be deleterious. See my BGA talk slides for an evolutionary argument: a random mutation is more likely to hurt than help; those that do help get pushed up in frequency by selection until they are no longer rare.

2012-07-09 on infoproc

Read a bit further in for more about how he runs his funds and how gullible his investors were.

2012-07-08 on infoproc

I say human nature. There will always be Madoffs, con men, trendy funds, etc. preying on greed and overconfidence.

Also, what’s the minimum g required to deeply understand (and have your behavior modified by) statistical results in portfolio theory? Even among rich people or those who control a lot of capital there will be many below this threshold.

I suspect that not only are individual investors and pension funds overpaying for financial services (e.g., 2/20), but *society as a whole* is overpaying for what it gets out of the financial services industry in total.

2012-07-08 on infoproc

1) I know admissions people have other goals in mind, but if you ask them (as has been done in studies) whether they think they can predict GPA performance (which is easy to measure), it turns out they are wildly wrong. They could be wrong about the other things as well but that’s harder to measure.

I don’t have any 2/20 investments. I almost put some money with Renaissance (not Medallion, that’s closed) but decided against it. However, you and I may be exceptions. Somebody is putting up the trillions under 2/20 mgmt!

2012-07-08 on infoproc

I just cut one mark in half, but there are no more.

2012-07-08 on infoproc

I think many people are irrationally overconfident. So they think they can pick the one hedgie out of many who actually has alpha, even though they may understand (at the rational level) that most do not. It doesn’t help that the hedgies are very good at portraying themselves as having special magic.

Why do admissions people at elite schools think they can predict college performance better than a simple linear combination of SAT + HSGPA? (When in fact the opposite is true.)

Why do economists think they can understand/predict the economy when empirically it is clear they cannot?

My son thought he had a way of rolling dice that increases the probability of his desired outcome. I got him to look at the statistics but he’s only six, after all 😉

2012-07-08 on infoproc

Yes, the author is completely clueless about genetic/environmental breakdown/confounds.

That doesn’t detract from the specific results correlating empathy levels and social class, although the causality is still in question.

2012-07-08 on infoproc

Even if the Agents’ incentives are biased so that they are actually acting rationally, their employers (also supposedly sophisticated) are clearly not making the right decisions.

I don’t believe all the Agents are in the know here. For example, in the Kauffman Foundation report on VC returns I linked to earlier, the Kauffman researchers seem genuinely surprised at how poorly their foundation has been doing with VC investments — and these guys are experts on VC and startups.

2012-07-06 on infoproc

The great thing about the SV/tech/science ethos is that you gain prestige by actually inventing and building cool things. As you know, this is not the case in NYC or DC. DC gives me the creeps when I visit. Wall St. is OK for me because I speak the language and have friends there, but it’s nowhere near as fun as SV.

You might like Paul Graham on cities:…

2012-07-06 on infoproc

The SNP genotypes contain the diploid structure: AA, Aa, aa. However, if (as we believe) causal variants tend to be rare, dominance is a relatively small effect.

The # SNPs / IQ SD is obtained two ways: by comparing outlier groups and also by just regressing pairwise genetic distance against IQ difference for all, e.g., (5000)^2 pairs in the very homogeneous ALSPAC population. But we are still doing more analysis to make sure the results are real. Also, at BGA multiple research groups stepped forward with more phenotype / genotype data (some datasets as large as 10k individuals) so we want to analyze those as well using these methods.

All of this will be described in gory detail in the paper and lengthy supplement.

2012-07-05 on infoproc

I interpreted “unusual” as exceptional. For example, a genotype which is +3 SD is exceptional/unusual. I do not believe there is a tight correlation between g and achievement (the correlation is positive and significant, but nowhere near unity).

But there may be cognitive thresholds for mastery of certain subjects.…

Chomsky talks as if there are discrete differences (“intelligent human being” vs ???), but really they are continuous. I think he means intelligent but still within some normal range; whereas some advanced subjects may require more.

2012-07-02 on infoproc

Genes of large effect which are relatively common in the population are easy to detect in association studies. Given the lack of replicable hits in current studies we can place an upper bound on a combination of effect size and frequency in the populations that have been studied so far. At least for Europeans (the focus of many association studies already), it is clear that almost all the variance is in alleles of small effect size and/or low frequency in the population.

2012-07-01 on infoproc

From one physicist to another, we argue as follows. Let there be N alleles of roughly equal effect size (an idealization, but doesn’t affect the result very much). Then one population SD in the trait corresponds to of order root(N) changes. If the trait is highly polygenic, so N >> root(N) , it is obvious that one can move many (i.e., of order root(N) ) SDs in trait value (“breeding value”) with existing population variation. When I explain this to physicists it takes about 5 seconds for them to understand. For “experts” in genetics, it takes a little longer 🙁

Lactose tolerance isn’t really very polygenic (AFAIK), so to make a change there required a new mutation. Without the new mutation selection probably couldn’t push things very far.

2012-07-01 on infoproc

> Do you know if results this dramatic have been demonstrated in other contexts? < Yes, many. At ICQG there were lots of examples. Two that I remember well are size/temperament of chickens! The math is pretty clear that this has to be the case for most quantitative traits.

2012-07-01 on infoproc

We get more adaptive mutations as the population gets bigger, but I think the dominant factor is still the extant variation — e.g., (-) alleles that can be pushed to lower frequency by selection. So if Clark is right and free markets and industrialization selected for intelligence, future time orientation, ability to cooperate with others, etc., then the response would come mainly from decreased frequency of (-) alleles for those traits in each new generation.

The amount of extant variation in any quantitative trait is likely to be huge, and a population can react to selection pressure using this extant variation, even if the mutation rate is exactly zero. (Over thousand year time scales the rate for producing new adaptive variants is relatively negligible.) This has been seen time and time again in animal and plant breeding. One set of experiments on drosophila (IIRC) was able to increase their flying speed by more than an order of magnitude!

Weber, K. E. (1996). Large genetic change at small fitness cost in large populations of Drosophila melanogaster selected for wind tunnel flight: rethinking fitness surfaces. Genetics, 144(1), 205–213.

PS Thanks for correcting my 10B number earlier!

2012-06-30 on infoproc

Your question is actually a deep one. g scores are *scaled* to fit on a normal distribution. Unlike for height or weight there is no obvious correct unit. Whether +1 SD in changes are “equal” independent of where you start in the distribution is really a question of “validity” (technical term, meaning what real world outcomes are predicted by g scores) and depends on external societal factors (e.g., capitalism, market structure) and/or the nature of pursuits such as math or science or technology.

Note that since g is clearly polygenic and mostly additive in genetic architecture, it is natural to expect (central limit theorem) that it will be roughly normally distributed in a population. (At least, if one adopted a different set of units for g that did not result in a normal population distribution, the underlying genetic architecture would make no sense.) One can go further and note that evolutionary dynamics, as formulated by Fisher, are driven on the rhs of his famous equation by additive variance. So one could say that evolvable population traits are likely to end up normally distributed. It seems plausible that cognitive ability depends on many biological/physical subfactors in the brain, such as efficiency of neural function, brain volume, etc. that do have a natural unit of measure, and these subfactors are likely normally distributed in the population.

Finally, regarding the +30 SD number, one could reason in a completely different way. Since < 100 billion or so humans have ever lived, and there are probably many many more distinct genotypes/phenotypes (e.g., 2^(many thousands) based just on gene variants), it is unlikely that the already realized phenotypes could be near the maximum potential of the species on any trait. I often argue with other physicists about the true exceptionality of, e.g., Einstein, based on this kind of logic (i.e., many more physicists, drawn from a much larger talent pool, have lived post-Einstein than pre-Einstein).

2012-06-30 on infoproc

You may be right that 1,2 and 3 are slightly different from each other. We’re just at the beginning of this research and one can imagine study designs that will get at this question. For example, once we know where the causal variants are for modern humans we can compare to earlier hominids like Neanderthals, whose genomes we have sequenced. We can also look at primates. It’s an important question whether the same loci are detected as linked to intelligence across different times and cultures. We can look at very old people who were educated in the early 20th century (assuming we collect DNA from them soon; we can get IQ scores from military or school records), people raised in the Amazon rainforest, etc.

Other experts I spoke to in Edinburgh share my belief that we can determine much of the human genetic architecture once we have of order millions of genotype-phenotype pairs (probably in the next 5-10 yrs). There are a number of interesting and open applied mathematical and computational problems related to genetic prediction using such large datasets. Animal and plant people have had great success already in genetic prediction but are dealing with smaller effective populations.

One important aspect of the toy model I discuss in the talk is that if it is correct, then any couple has the potential of producing a nearly max IQ child, because there will be little overlap between the locations of their (-) alleles.

2012-06-29 on infoproc

When looking at the figure showing selection pressure on (+) and (-) alleles, think millions of years and the difference between humans and apes. There has clearly been directional selection for cognitive ability. Differences between groups of people or between detailed selection pressures provided by human environments are negligible compared to what happened during those millions of years.

2012-06-28 on infoproc

Several people made the Tolstoy observation; you are not alone 🙂

“intelligence … fairly optimized in the population already”: the results imply lots of extant variation to work with. Just imagine flipping all the (-)’s to (+) in a particular person. You get a lot more than +5 SDs.

39 = 30 at this level of precision. 10k is better than an order of magnitude estimate, but could easily be off by a factor of 2-3.

2012-06-04 on infoproc

Actually I think 10.3/13.3/4:15 is the most impressive evidence of raw power/speed and endurance all in one person.

Incidentally, I think Eaton smoked the UK’s best hurdler in the video.

IIRC, Liu Xiang ran 10.5 or so in the 100m as a kid. That’s faster than the fastest UO football player.

2012-06-03 on infoproc

Your throughput estimates seem way too low. Also keep in mind the coverage level.

600 GB/wk is doable now, which is many complete human genomes per week on a single machine.

I’m going to delete all of this soon.

2012-06-02 on infoproc

Runs scheduled for Sept 2012. Will take a while to process and clean data.

2012-06-02 on infoproc

Unofficially, and off the record, we will be doing whole genome sequencing instead of SNPs. So you will eventually get a lot more information than you did from 23andMe. You will likely be among the first 10k or so humans to be sequenced at this level of coverage.

Note, I will probably not leave this comment up for long.

2012-05-31 on infoproc

For some reason there is an enclave of Pokemon enthusiasts among kindergarteners in Eugene OR! My kids got sucked in.

2012-05-28 on infoproc

Genetic technology is used in two ways in Gattaca:

1. for personnel selection — what you guys are discussing; this is kind of questionable for reasons mentioned on this thread

2. for pre-implantation selection of embryos — this works even if the heritability is, e.g., 0.5

2012-05-22 on infoproc

They still don’t have the chops to look into HFT. I think it would have to be complaint-driven, and they’d have a hard time getting their arms around it.

2012-05-22 on infoproc

One math/phys postdoc I know looked into it. But they were looking for someone less high powered (!!!)

2012-05-18 on infoproc

The implausible prior is not that there will be some (exceptional) single gene controls, but that *most* phenotypes or disease resistances will be controlled by only a few genes. Some people were/are genuinely surprised that this didn’t turn out to be the case. Relative to what we are finding, “few” could be 10’s or even 100.

You can do genetic engineering with thousands of causal alleles if variance is additive. In the long run that may be more interesting than drugs.

2012-05-15 on infoproc

I just looked at his CV and realized I might have mis-remembered. The second finance trip was maybe consulting work and he kept his academic job.

Re: kilocalories, Don’t they teach you guys anything in Canada? I lecture about this stuff in physics 101 when I cover energy 🙂

2012-05-15 on infoproc

Class and status are fun, but people with money have real power. Gates, Bloomberg, Jim Simons, Google founders, Zuck, Buffet, etc. all have something in common — they are smart and they did/created stuff. The geek/nerd neg loses a lot of force as you get older.

2012-05-15 on infoproc

Y is weaker in STEM so it’s a slightly different category of school.

2012-05-15 on infoproc

See my post Creators and Rulers.…

2012-05-15 on infoproc

I know one of the Coursera founders. It will be very interesting to see how all of this plays out. It’s not really a technology competition, though.

2012-05-15 on infoproc

Caltech’s model will not allow it to compete with HPSM for the reason you mention, among others. (M has started aping HPS and isn’t that much like Caltech anymore.) It’s still a great institution for creating value for society, but it will never win the prestige or even the resource/endowment competition.

2012-05-09 on infoproc

> they are independent of the ‘interpretation’. < Yes. > Btw, I always thought it was common knowledge that decoherence entangles e.g. a macroscopic apparatus with ‘the environment’, i.e. the whole world. At least this is how I read H.D.Zeh. < Yes, but the details were not well understood. Beyond the bipartite case classification of entanglement is still an open problem. It turns out that things simplify considerably for "typical" many-particle states and this can be applied to cosmology. For example, the idea that entanglement between X (small) and the universe is maximal, but between X and Y (also small) exactly zero might be surprising to some people.

2012-05-09 on infoproc

Yes, but definite predictions remain. Most of the entanglement we describe remains after projecting onto our “relative state” (as Everett would call it).

Note, our results hold even if you take the universe to be in a mixed (not pure) state, assuming you don’t choose some crazy measure over the probabilities defining the mixed state.

2012-05-09 on infoproc

See last comment about photons in a box. If you agree that entanglement properties of Psi describing the box are (in principle) measurable, then next consider two sets of photons that already exist (e.g. in the CMB) but will only come into causal contact in 15 Gyr. You can make similar predictions concerning their entanglement that will be testable in the future. This is a not a practical experimental proposal, but at the moment neither is the experiment with the box of photons.

2012-05-09 on infoproc

UCSC is strong in several areas, not just astro. They (like UO) are probably higher in this set of rankings than they should be, but are another example of a strong department that isn’t well known outside the field.

2012-05-08 on infoproc

Keep in mind the rankings include all areas of physics, not just hep-th or GR!

2012-05-03 on infoproc

> … the hard sciences, math …

In these areas I’d guess the evidence is pretty strong that returns to g continue to be positive above +2 threshold. But in other areas I just don’t know.

2012-05-03 on infoproc

> g is merely necessary up to a fairly basic point < I think it's more complicated than that. For example, perhaps someone who is +2 has a 1% chance of somehow developing into a "genius", whereas someone who is +3 has a 10% chance, etc. Jensen himself is not very mathematical -- he expresses himself in the excerpt rather imprecisely. Note his point about the Nobelist he tested disfavors the idea that the probability of developing into a "genius" is independent of g once the basic threshold is reached. If that were the case almost all of the identified "geniuses" would have IQ pretty close to the threshold since there are so many more individuals there than further out on the tail. If you search around on the blog for SMPY or Roe data you'll find evidence against this.

2012-04-30 on infoproc

Perhaps I am mistaken, but Steinberger (who was at the school) might be the guy to the right of Shelly (from our perspective) that I earlier identified as ??

I’m pretty sure the guy with the sweater on his shoulders is Zichichi.

2012-04-30 on infoproc

Even worse, I missed meeting John Bell by one year — he was a lecturer at Erice in 1989 and gave his famous talk Against Measurement. If I had heard Bell give that talk in 1989 it probably would have changed my life, as I had been wondering about measurement since learning QM in high school.…

2012-04-26 on infoproc

> Do I understand you have already 2000 just at the 1 in 10,000 level? Could you say more about that sample?

SAT administered to younger kids has a high ceiling. This has become the most common way to test gifted kids in the US.

> Also how many do you have at the level 1 in 1,000? And are you accepting more of those?

We set the threshold on our web site at roughly this level. You can volunteer there:

All accepted volunteers will get at least SNP genotyping, possibly whole genome sequencing.

2012-04-26 on infoproc

Perhaps it wasn’t clear from the slides, but when I refer to predictive power of the test, I mean for real world outcomes (e.g., college GPA, income). Ability of the test to predict the outcome of a similar test at a later time is “reliability” not “validity”.

2012-04-25 on infoproc

Thiel is a big risk taker and has correspondingly strong opinions. I don’t think he’s in the category of “cautious, careful thinker” so, yes, you should weigh his advice carefully before accepting it. He made a very aggressive global macro call with his hedge fund a few years ago that, as far as I understand, went very badly. On the other hand he probably more than made up for it with his early Facebook investment …

2012-04-24 on infoproc

I actually agree that CA has a lot of problems.

I was trying to express that nevertheless there is something ineffable about the sunshine and natural beauty here. It wasn’t a serious policy/economic/demographic analysis.

2012-04-21 on infoproc

Trying to decide whether I should try to see the fight at a bar or something. Bones has to be the favorite, but Suga impressed me against Davis. Could be a good fight!

2012-04-19 on infoproc

I might write something, but in the mean time, the workshop wiki is pretty good and the videos are quite watchable.

2012-04-14 on infoproc

Thanks, I think it’s fixed now …

2012-03-29 on infoproc

Definitely not for everyone, and I agree about injury risk. I don’t do full-blown CF, but I like some of their exercises and I find the videos inspirational. The attitude that by pushing hard you can get your workout done in a very compressed period of time is very useful for me. But warming up properly is important.

2012-03-13 on infoproc

Familiarity breeds contempt 😉

Just kidding.

2012-03-13 on infoproc

>> So it may be that social and cultural factors play just as strongly in the “hard” and “soft” sciences. << Your comments are correct but it's a question of degree. The problems that Woit and Smolin indicate are real but are caused by lack of experimental ability to probe quantum gravity (string theory). Under these circumstances physics can become almost as bad as the standard situation in the humanities. But it is an anomaly.

2012-03-13 on infoproc

My wife has a PhD from Berkeley and is a professor of comp lit. I know exactly what you mean 😉

2012-03-13 on infoproc

You may need to learn to hold your tongue in certain circumstances 😉

2012-03-13 on infoproc

>> But scrupulous self-honesty is required, and one has to approach problems with an eye for truth and a willingness to let the object of analysis bite back << I agree 100%. Unfortunately, as career advancement and social prestige become more important later in life, this becomes harder and harder.

2012-03-12 on infoproc

Fermi was the first hacker 🙂…

Feynman’s mentality was hacker-like as well, of course.

2012-03-11 on infoproc

I’ve never looked carefully into this question. However, creating the first engineering instantiation of an idea is not equivalent to having a deep theoretical understanding of the idea (there are many examples of this in history; the inventor of the steam engine did not have a deep grasp of thermodynamics, etc.). The reviewer you linked to suggests that computer scientists gave vN credit primarily for elucidating the theoretical aspects of what came to be known as vN architecture. Whether this was just I do not know.

It happens that I looked into some of McCartney’s claims about Eckert and Mauchly vs Atanasoff (these arose in the court case that awarded Atanasoff the patent for the first digital computer), and it seemed clear to me that McCartney was incorrect. See the very detailed research by Alice and Arthur Burks. This alone is enough to make me doubt what McCartney has to say about vN.

2012-03-11 on infoproc

From the review:

“McCartney’s particular slant may also have its origins in current attitudes. In the 1960s and 1970s, the mathematical aspects of computer science were ascendant; for example, there was a strong correlation between the rankings of computer science departments in the U.S. and the quality of their theory faculty. In the 1980s and 1990s, there has been a shift in academia toward the engineering aspects of computer science. Herman Goldstine’s The Computer From Pascal to von Neumann—published in 1972, the heart of the mathematical era—argues for the prominence of von Neumann’s contribution to the stored-program theory over the engineering contributions of Eckert and Mauchly. These early developments have been reinterpreted over the last fifteen years, giving more emphasis to the machine builders and less to the theorizers. McCartney is not a member of the computer community and does not seem aware of this historiographic shift, but as a journalist he has picked up on the prevailing attitude in the practitioner community.”

I don’t think McCartney’s take on all this is very reliable. His views on Atanasoff were, IIRC, incorrect as well. (See my post on Atanasoff and some of the links there.)

It’s also fair to say that although vN was the most brilliant, he clearly was not the most creative or original. He was surpassed by Einstein (as Wigner notes), but was also scooped by Godel and others. Turing’s contributions may also be seen as greater in the fullness of time.

2012-03-04 on infoproc

I don’t claim to be an expert, but my general experience with modeling complex phenomena is it is very easy to have much more confidence than is really warranted in your model. Therefore I think it is reasonable for an outsider to be skeptical as to whether existing climate models can predict the future, or whether they have properly characterized the dynamics. Nordhaus’ response indicates he doesn’t appreciate this point.

Off the top of my head I can’t think of any complex systems (e.g., stock market, macro economy, the solar cycle) that we have good predictive models for.

2012-03-01 on infoproc

Are you confusing Schiff and his brother?

2012-02-29 on infoproc

I don’t want to ban anyone. But aren’t you guys at the point where you could easily imitate each other (i.e., as in pass a Turing test)?

Maybe you can create a FAQ-like point by point exchange (e.g., using google docs) so that everyone on the internet can access your combined wisdom and decide for themselves? 🙂

Then on a comment thread you can just say: “This was addressed in part 3C under Princeton valedictorians relative to student population, legacies and athletes excluded” and give the hyperlink. 🙂

2012-02-27 on infoproc

Although some of the comments here have veered towards the topic of Asians, my original intent was to address the critics of my NYTimes op-ed. A vocal subset of commenters there kept claiming that SAT is only a measure of SES or test prep. That op-ed was about affirmative action for URMs, so the main question is whether URM SAT scores are a good measure/predictor of their ability to benefit from elite education.

PS My comment above that you liked was partly tongue in cheek. Although I do think a more intense education experience can raise test scores for an entire sub-population (this applies to both the A-W and W-B test score differences), no one knows whether it accounts for the entire difference or only part of it.

2012-02-27 on infoproc

Yes, Turing anticipated Penrose (actually, he was not the only one). I think Turing can be forgiven on this particular bit of nuttiness as it was still early days for QM foundations, philosophy of mind, logic, etc.

2012-02-27 on infoproc

Exactly how PSAT score predicts SAT score is a bit tricky because kids are still learning a lot at that age. In particular, some kids are only just encountering/mastering some of the material that is on the tests, like algebra or geometry or how to read a graph.

In E. Asia most parents/kids are already scrambling for whatever edge they can get on entrance exams, etc. Not sure this is a good thing, but on the other hand people seem a bit better educated there on average. As long as it’s not just test-specific tricks they are practicing, there is likely some actual learning/education going on. When I first visited Seoul I was shocked to turn on the TV in my hotel room late at night and find someone working out calculus problems on one of the channels!

2012-02-27 on infoproc

Look at the range of scores. It seems unlikely that this is a special population of people. (This is addressed a bit in the paper.) I suspect very few people who score that low on PSAT were “intensely preparing”. More plausible that there is just some probability p that a given kid will try some test prep in between taking the PSAT and SAT. There might be some selection bias in the direction you mention but I doubt it is very significant.

Note there is some systematic improvement for very low scorers. Perhaps for them a few tricks help a lot. If so, they didn’t know the tricks when they took the PSAT.

2012-02-26 on infoproc

Very hard to disentangle if you are talking about sub-populations with very different cultures and intensities of educational experience.

2012-02-26 on infoproc

To be fair this data does not rule out the possibility that a yellow horde of really determined kids, driven by fearsome Tiger moms for most of their lives, could systematically raise their scores somewhat. But it does show that it’s not an easy task. At some point you are talking about actual learning/education as opposed to mere test prep.

2012-02-23 on infoproc

Don’t hold your breath.

2012-02-23 on infoproc

Maybe you can post the URM graduation rates since prop 209 (see my other comment). I think phrasing the argument in terms of SAT as predictor of college performance is too confusing for most readers — just directly stating the mismatch hypothesis (URMs get better outcomes when matched in ability to the rest of the campus) and supporting data would be better.

2012-02-23 on infoproc

Unfortunately I was under time pressure when preparing the op-ed, otherwise I could have argued from this perspective. Apparently, it’s a slam dunk:

“4) Graduation trends. URM graduation rates have improved sharply since Prop 209 went into effect. For the six cohorts of black freshmen who started at UC campuses before Prop 209 went into effect (the matriculating years of 1992 through 1997), the average 4-year graduation rate was only 22.2%. For the years since 1998 (matriculating years 1998 through 2005), the black 4-year graduation rate across the UC system is 39.4% — a near doubling. For Hispanics the 4-year graduation numbers are 27.2% for 1992-97, and 41.8% for 1998-2005. Six year graduation rates have risen as well, though less dramatically. Combined with the matriculation trends described above, the number of black and Hispanics graduating from the UC system has been rising steadily and remarkably. For example, the number of black students who matriculated at UC campuses in 2005 and graduated in 2009 was over two-and-a-half times higher than the number of blacks who earned 4-year degrees annually in the early 1990s.”…

2012-02-23 on infoproc

Good point. I suppose it’s most effective just to point out that URM graduation rates went up after UC went to race blind admissions.

2012-02-23 on infoproc

IIRC, after Prop 209 the number of underrepresented minorities (URMs) at UC decreased substantially, but their graduation rates went up significantly.

But you don’t need such dramatic evidence. The Duke study (and all other similar studies) find that higher SAT score predicts better grades or class rank (and the opposite with lower score) for all groups. That means URMs admitted with higher scores typically do better at college.

All you need to know about mismatch is that if you regress class rank or GPA on SAT you don’t need an additional ethnicity or race variable to improve the fit. The predictors work about equally well for all groups.

2012-02-21 on infoproc

I actually agree with you on this — I think Adam was probably confusing analytic geometry with calculus. It’s a shame Russ didn’t pick up on it — I was generally in shock at how little intuition these, um, highly verbal guys had for making real stuff.

Adam was also confused at points about why Luke’s job isn’t in China. It has to do with labor vs capital costs and not with availability of Luke’s skill set in China. People with Luke’s abilities are plentiful in China — literally 1/5 the cost!

2012-02-19 on infoproc

Feynman was about 5″10 when I knew him but could have been a bit taller in 1961.

Dyson is not very tall.

2012-02-19 on infoproc

I think Bowles and Gintis are careful to distinguish income and earnings. The former could come partially from inheritance. But you’re right that having a cushion to fall back on also helps in starting a career, and hence boosts earnings as well.

2012-02-19 on infoproc

RE: dating, OK Cupid may have done something like this.

I agree, it’s a feasible project and the outcome would be quite interesting.

What about NFL combine or college stats as predictors of NBA/NFL outcomes? I’m sure that’s been looked over pretty carefully.

2012-02-17 on infoproc

It’s possible that both Clark and Murray are correct. Clark’s convergence may have stopped and reversed in the last 30 years due to globalization, meritocratic sorting, etc.

2012-02-17 on infoproc

That is exactly what Greg finds, across a variety of cultures. Doctors in Sweden and India, Oxbridge attendees in England, etc. 200 years is not enough time for an extended group to regress fully to the mean. This suggests an invisible dynamics of assortative mating and values transmission that operates in every society he’s studied.

2012-02-14 on infoproc

I meant specifically moral hazard affecting decision making by bank CEOs or their mortgage traders. I do not believe these guys took on extra risk specifically because they thought they were going to be bailed out. The *firms* may have been bailed out, but these individuals lost huges amounts of money in the collapse (value of equity went to zero).

2012-02-12 on infoproc

This is childhood performance, which tends to be less heritable (more subject to environmental influence) than adult ability. Obviously genes play a role here as well — it would be nice to see a 90/10 analysis of the parents 🙂

2012-02-11 on infoproc

It seems clear to me that many more white, upper class Americans are interested in elite higher ed, and are willing to do what it takes to prepare their kids for the admissions battle, than was the case 20 or 30 years ago.

When I was in HS, typical white and even most Asian families just accepted that the SAT really was an aptitude test and that there was no point in preparing for it. (And these were families of professors and scientists.) The only kids I knew who prepped for it were Jewish.

2012-02-10 on infoproc

Thanks for pointing out the typo. The slides at the end are just for fun in case someone is curious about that topic — it’s very speculative.

Note this talk is very similar to the one I gave at Google, although if you look carefully you’ll see there’s some progress.…

2012-02-10 on infoproc

Works OK for me, but I use Preview on OSX. Can you try again? Or use gdocs?

2012-02-09 on infoproc

I don’t have any feel for the answer to question 1. I can see negative trends ahead for finance and for the higher ed bubble (although maybe not MBA programs and maybe not quant trading). Perhaps other readers will comment.

Re: 2, I don’t think it’s really a brainpower question. You could get to $500k per year if you work hard and have good interpersonal skills, along with a bit of luck.

2012-02-09 on infoproc

Believe me, hedgies have better perks than those other schlubs.

2012-02-09 on infoproc

Yes, I think originally we were thinking of an Ivy undergrad (academic admit), say in top quarter of their class.

I agree with your rough estimates but keep in mind that I doubt g can be defined/measured to much better accuracy than 1 population SD.

2012-02-09 on infoproc

This is known as the “fake alpha from hidden tail risk” strategy.

2012-02-09 on infoproc

If indeed income above (say) $150k per year has limited marginal utility for you, then I would definitely take the academic route. The B-school prof route is cushy, well paid and not that competitive compared to actually working in finance or in a really tough academic field. Actually in some sense it has been a good arb for a while now — good salary, great quality of life, not too hard, etc.

Point 4 above may cause you some problems, especially further along in your career. Keep in mind that, with the exception of a few specialties, career success even in academia is not primarily determined by brainpower but rather in combination with other things like effort, ability to play politics, make a superficial impression, etc. So being smarter than your competition might not be very decisive. Another way to say this is to note that although 1st rate people hire other 1st rate people, 2nd rate people typically hire 2nd and 3rd rate people. [ I know you say you’re not brighter than the other PhD students would be but I suspect you are …]

You seem pretty pragmatic, so you probably won’t explode from having to write dumb, obvious papers and spending years getting them refereed and published in “top journals”. This is often a problem for really smart people who deviate from the traditional academic specialties where the other people are super smart.

Not sure why you are excluding econ/finance as your PhD area as that might be the best fit.

You can probably pursue your machine learning/stats interest as part of your PhD and later research program. These methods are applicable to almost all fields now. So I don’t think you’ll lack interesting things to do.

BTW, there’s always the argument that you should just do what you are passionate about and let the chips fall where they may. If you go for a math/physics PhD (assuming that’s what you are passionate about) option (a) will still be there at the end. You shouldn’t be too sure that you aren’t smart enough — if you really want it, you probably have a shot.

2012-02-09 on infoproc

Just trying to understand what is happening in this once great country of ours…

I also have a specific interest because I almost went into finance on several occasions.

2012-02-09 on infoproc

Yes, after tax differences are much bigger, although the total compensations of the other groups are like a rounding error for the hedgies 😉

2012-02-08 on infoproc

I put up a link to more data.

2012-02-06 on infoproc

> which arguably changed the trajectory of their lives < In the old days (pre- finance and startup economy) controlling for SAT score eliminated almost all of the earnings advantage of attending HYPS as opposed to a state university. In other words, Ohio State grad with Harvard admit SAT scores had the same lifetime earnings as Harvard grad. The situation more recently is hotly debated by economists who study this question. But in general these results place a cap on the size of the effect you are talking about. > He felt you needed really complex and novel problems to discriminate g at this level < This is the kind of BS promulgated by (supposedly) high g types who are outside of science and academia. The stuff in the Caltech curriculum is plenty g loaded and high ceiling enough to distinguish 99.9th from 99th percentile. You have to be >99.9 to have a shot at grasping advanced topics in pure math and physics while an undergrad. Kids who are >99 but below 99.9 have a very good chance of washing out of the harder majors at Caltech.

2012-02-05 on infoproc

Causality is tricky and one could always claim that since we have a credentials-based system and are already sorting people by IQ/SAT from an early age that it’s all self-fulfilling, blah, blah.

SMPY has subpopulations that are 99th percentile and 99.99 percentile and you can see huge qualitative differences in their life outcomes. All these kids come from good families, were in gifted programs, etc. So it’s plausible that the main difference between the quartiles is simply ability as measured by SAT score at age 13. IIRC, the high earners in the top SMPY quartile made a lot more money than counterparts in the bottom quartile.

Re: 1400+, Caltech’s entire population is above this threshold and the admissions people had a detailed model that would predict Caltech GPA using SAT as an input. Obviously score above 1400 was correlated with performance at Caltech. But you need actual statistics to see these effects — anecdotes and personal experience may not be enough.

2012-02-05 on infoproc

There are plots for other majors in our earlier paper Data Mining the University (search on blog to find a link). You can see at a glance they look very different.

The blue bin sizes were varied near the threshold of 600 to get more resolution on what was happening, and some of the choices overlapped.

Nothing wrong with grinders, but keep in mind this is undergrad data so anyone < 3.5 probably isn't going to get into grad school or finish a PhD. Physics GPA=3.5 is only .5 SD above the average in the major.

2012-02-05 on infoproc

The data we had was for the general CS major at Oregon. If you could narrow it down to theoretical CS (which is basically math), I’m sure you’d detect a cognitive threshold.

2012-02-05 on infoproc

Click through to the research papers to find the answers to your questions. Or google SMPY.

There are almost certainly many millionaires in this population. Quite a few with incomes above $500k per annum, as I recall. (Do you know what a quant is?)

> Anecdotal evidence suggests that beyond a score of around 1400, the SAT does a poor job discriminating among different cognitive levels, and good students do better than good thinkers.< Actual studies contradict your anecdotal evidence. > they will typically regress to the mean in adulthood < Not by that much: still above +3 SD when tested again as college seniors. (This discrimination was possible with the old SAT.)

2012-02-05 on infoproc

SAT is one of many signals that have predictive power. Testing for empathy or social skills (e.g., via interview) also adds information to the process if done carefully, but of course it is easier to game an interviewer than an SAT test.

You’re nuts if you think my views are as simplistic as you imply.

2012-02-05 on infoproc

Other than pure math there is no other major we could find that has a cognitive threshold. We did SAT-M, SAT-V and M+V but nothing else looks like physics/math.

Note we don’t have EE here at Oregon (only at OSU), but we did analyze CS and there is no threshold. Click through to read the paper we wrote.

(Oops, perhaps I misunderstood you. The flattening of the blue points does not mean that probability of success saturates. That’s just because we’re taking the upper bound at 95 percent confidence. The main goal of the blue analysis is to bound the probability of success for lower M scores. The central value of the prob. of success increases above M=680 and at M>750 is still only about 50 percent. That is, only about half the physics majors with M>750 manage to graduate with in-major GPA > 3.5.)

2012-02-03 on infoproc

What about the $cha-ching! part? When will that happen? 🙂

I’m getting lots of emails from finance people about the piece. Almost all are favorable. A few are mad at me that I don’t address Affirmative Action, but I tell them to go back and read the final paragraph more carefully, and then they are happy.

2012-02-03 on infoproc

Yes, it’s a definite possibility. The whole issue is rather subtle, which is why I favor transparency.

It’s entirely possible that older folks like me are conditioned by the situation 20 or 25 years ago when bad things were definitely happening, whereas today most of the negative effects on Asian applicants are not due to racial bias.

2012-02-03 on infoproc

I tried to word the op-ed carefully to make clear that anti-Asian discrimination, while certainly something that happened in the past, may have been fixed. But I do think we (or at least I) need a closer look at the data to see what is really happening. I also think transparency about admissions will have other positive effects. You’ll note in my last paragraph I again endorse color blind admissions.

Regarding Duke specifically, the researchers using the CLL data don’t think that there is evidence for Asians underperforming their admissions strength, but I am awaiting further clarification from them. A technical point: they do not think senior grades are useful, due to compression of range (i.e., grade inflation), and think junior year class rank is more indicative of performance at Duke. They find that Asians are more prevalent in the harder STEM majors and have similar average grades to whites in those majors, in line with higher application strength. But, again, I await further analysis on their part.

2012-02-03 on infoproc

Maybe someone at the Dept. of Education will check on this question for us … 🙂

2012-02-02 on infoproc

>data comparing Asian and white applicants should be released in toto, but that similar data for black and Hispanic applicants should continue to be shielded< I'd like to see *all* of it released, but I agree with you there are reasons why DoE won't push too hard for this.

2012-02-01 on infoproc

Thanks! I think it is fixed now…

2012-01-31 on infoproc

> I reckon about Steve that he has ‘enough’ money, and that he is chasing things that can’t be bought. < Bingo! But if the horses don't come in I might reconsider and wish I had flown more miles in first class or on my own jet ...

2012-01-30 on infoproc

> finance work is boring VS startups and physics

Yes, but the pay is better!

> other brainiacs outside of finance would consider you a sellout

What do they know?

Despite hindsight being 20/20 I’m still not sure I made the right choice 🙂

2012-01-29 on infoproc

I teach my kids the Golden Rule …

2012-01-27 on infoproc

I have to admit the closest I have been to Detroit are the airport and Ann Arbor. So I have no first hand knowledge of the city …

2012-01-27 on infoproc

Maybe not fully destroyed — IIRC Pittsburgh lost half its urban population over 1-2 decades when the steel mills closed but is doing relatively well now. Perhaps the same thing will happen to Detroit, with the culprit again being concentrated industry in a sector that collapses.

2012-01-26 on infoproc

This kind of (self-interest motivated) trust is discussed specifically in the podcast. Rose claims that more is necessary for the most sophisticated forms of economic organization (e.g., highly efficient complex firms).

2012-01-26 on infoproc

Yes. Tell them to volunteer here: https://www.cog-genomics.or…

2012-01-22 on infoproc

Benbow and Lubinski have been studying spatial ability as a separate cognitive factor, and find it is a significant independent predictor of career outcomes. See the figure at the post below; spatial ability is represented by the arrow, with superior ability pointing to the right. Engineers come out on top, although perhaps because they’ve lumped the physicists in with chemists and other physical scientists.…

Roe also measured it in her study of top scientists. I was a bit surprised that theoretical physicists came out ahead of experimenters in spatial ability, as those guys actually build things. Perhaps we wouldn’t be quite as useless in the lab as we think 😉…

The author of the Atlantic article is pretty astute. I was impressed when he emphasized the role of spatial ability in Luke’s job.

2012-01-21 on infoproc

I think you miss the point. The advantage of low-wage countries is labor cost. Our advantage is in skill- and capital-intensive manufacturing. But that creates jobs for (a limited number of) Lukes, not Maddies.

2012-01-20 on infoproc

Not sure why the Disqus spam filter grabbed your comment. But I’ve now released it.

PS I don’t read all the comments, esp. the ones on old posts!

2012-01-18 on infoproc

Chonan very lucky. AS tapped very fast!

2012-01-16 on infoproc

Check out the podcast if you have time — it’s pretty good. Your recollection is not inconsistent with his conclusions.

PS I’ll be giving a physics colloquium at UC Davis on our genomics research. Mon Feb 14


2012-01-11 on infoproc

Crow was right on top of things until the very end. He notes correctly in the lecture that

1. Population genetics has until now produced a lot of theory but very few testable predictions.


2. This will all change with the advent of cheap sequencing. He points his younger colleagues in the right direction!

2012-01-06 on infoproc

Hard to point to Lander’s signal research contribution, as you note.

Venter’s math brain:…

Many techers on the computation team at Celera.

Highly recommend this book:…

2012-01-04 on infoproc

Hmm… good advice!

2012-01-04 on infoproc

Sineruse / Siserune: There are many, many comments on this blog that I completely disagree with, on both sides of the argument that you are involved in. (In fact, on every possible subject …) A blog owner has to decide how much effort they are willing to devote to dealing with comments, and I have mainly adopted a laissez faire attitude. (Also, I think it is interesting to see opinions of the sort that are normally suppressed.) The only commenters I have ever banned are people who have directly insulted me (in this case, suggested I am a racist of some sort).

PS I read through some of your comments on College Confidential and found I agreed with most of them. Is there an email address that I can reach you at?

2012-01-02 on infoproc

You are absolutely correct that there could be complex interactions and unintended consequences. On the other hand there are definitely free lunches (cost free improvements to be had). Certainly some people are *overall* big winners in the genetic lottery (not just in intelligence but in other aspects such as longevity or physical ability). Evolution has only explored a tiny subset of human possibilities.…

However, if I were a potential consumer of this kind of technology (e.g., some future company is peddling embryo selection technology as in Gattaca), I’d want to study the science very carefully, just as you would before adopting a particular drug therapy or medical treatment. Ultimately geneticists will only be able to make statistical predictions, so you will be gambling that your intervention actually improves the “fitness” of your offspring.

2011-12-30 on infoproc

I guess I’ll let other commenters decide whether what I wrote makes sense. You do realize that +.5 SD (or whatever) is *relative* to a reference population, and that you have to separate performance of that population from others in making a calculation? There are related groups of models here, some have a “striving” effect, others just have g, etc. But it seems plausible (and I thought this was also sineruse’s point) that striving has more impact on lower level competition than at the tippy top. But when you finally get to Putnam Fellows I would suggest that the effect of striving has decreased a lot. If this doesn’t make sense to you, too bad. I doubt other readers are having trouble with it.

2011-12-30 on infoproc

> …absolute differences in ability between Asians and whites in the context of discussions about elite college admissions

I’m interested in this kind of thing (HBD) independent of college admissions. (But note I don’t know to what extent these differences are genetic.) I don’t think I can be any more clear about what policies I favor for college admissions.

As to why I voted for Obama, there are many factors that go into choosing one candidate over another. I can support Obama over McCain even if I disagree with him on certain issues. There’s no big contradiction here, just as there wasn’t with my earlier calculation.

You obviously aren’t adding anything to this discussion. Why don’t you just admit you are in over your head? Being dim is OK, but being aggressively dim on my blog is not.

2011-12-30 on infoproc

He not only called me a racist but suggested I might treat research students in my group differently, based on ethnicity.

BTW, I don’t see why it’s necessary to protect the reputation of an anonymous commenter.

2011-12-30 on infoproc

I’m getting a bit impatient with your stupidity. You wanted to know why I did the little calculation (somehow accusing me of something nefarious), and now (finally, after repeated clarification) I have explained it to you.

I already posited (and I think Sineruse realizes this, but you are apparently too thick), that Asian strength of application is enhanced relative to cognitive ability. That’s what the striving is all about. It’s very possible that if elite schools took the rank ordered top N students (how to define “top”? see below) in the US the Asian fraction would be less than the 20% it is currently. This is an artificial scenario though, because Asians and Jews are much more interested in attending these elite schools than even the very smart fraction of W’s. That is, they apply to such schools with much higher probability. When this is factored in I don’t actually know what the population ratios “should” be. I advocate a race-blind market solution — let people apply, and let schools admit, but in a race blind way.

I’ve never advocated that university admission should be based on pure ability (g), or some combination of g with Conscientiousness (striving), or plus leadership and athletic qualifications, or even based on a projection of future success (however that is defined). Each school can make up its own mind as to how it wants to admit applicants. But the criteria should be applied uniformly to all ethnicities. There should not be penalties and bonuses based crudely on race.

Regarding your last question, I don’t know the answer, but if we are only talking about racial preferences, I don’t think dropping them would affect gentiles / Jews differently. Legacy/athlete preference might be a wash but I am not sure.

2011-12-30 on infoproc

Again, perhaps I am mistaken, but I recall Sineruse complaining about a SND (“Shifted Normal Distribution”) model in some of his posts. He adopts it in your excerpt above to argue that *even within* a SND model Asians are overrepresented relative to ability (g) at elite universities. That doesn’t mean he actually endorses the SND model — he is just using it there to make a point. I get the impression he doesn’t believe in the model, but I’d welcome his clarification on this.

IIRC, in at least some of Espenshade’s admissions models the elimination of preferences would lead to an increase in the Asian population and no change in the W population.

2011-12-30 on infoproc

Perhaps I was mistaken but it seemed sineruse complained about the +.5 SD I used in an earlier post. I thought it funny that all of his digging leads to a similar result.

The W subpopulation affects the estimate of the .5 SD.

In case you cannot read, I already posited that striving might enhance EA performance on grades, tests, etc. relative to g. But what is that g score?

Oops, this was meant as a reply to rogic.

2011-12-29 on infoproc

I’ve lost track of all the IMO/Putnam/ethnicity census analysis on this thread. While you guys are at it, can you verify for me that Caltech has historically produced more Putnam Fellows per undergraduate than any other school? (I’m guessing 2x lead over MIT and even bigger vs Harvard 🙂

If we just count US HS educated competitors, a roughly 6x overrepresentation of E. Asians would result from, e.g.:

+.4 SD math advantage; Putnam winner threshold +4 SD; tail area ratio:  (+3.6 SD / +4 SD)   =  (16/100k) / (3/100k) = 5-6

Sineruse: I used similar SDs for the two groups — I hope that doesn’t bother you 🙂

But perhaps the correct number is more than 6x? There’s a high performing W subpopulation that is hard to keep track of … Also, what is the actual W to E. Asian population ratio?

BTW, when examining small numbers of extreme cases there are many factors to take into account. I know a lot of people from the population under discussion. A fair number of HS math competitors end up in CS or physics in college and don’t really want to spend energy on Putnam (e.g., learning theoretical CS or quantum mechanics or general relativity are sufficiently taxing that they don’t have spare cycles for competition math…). Are you sure factors like this aren’t dominating the effect you are looking for? I can imagine a lot of pushy Asian parents telling their kids to do more practical stuff than math. Some IMO winners even end up as doctors or lawyers!

I would not be surprised if Asians have figured out that strong USAMO performance is the ticket for their smart kid to an elite school, despite discrimination (tough to stand out among Asians as SAT ceiling is so low…). But most of these kids are probably not as passionate about actually becoming a mathematician as the W kids who do well in the competitions, and are less likely to pursue math once in college. You would probably refer to this as “striverish” behavior.

2011-12-20 on infoproc

I’m traveling right now so don’t have time to respond to all the points you raise in your multiple comments.

I don’t have Espenshade’s book to check this, but the quote below from a US News article agrees with what I recall. It certainly looks like he controlled for legacy and recruited athlete status when computing odds ratios.

“Espenshade found that when comparing applicants with similar grades, scores, athletic qualifications, and family history for seven elite private colleges and universities:

Whites were three times as likely to get fat envelopes as Asians.
Hispanics were twice as likely to win admission as whites.
African-Americans were at least five times as likely to be accepted as whites.”…

Of course, it would be good to see some more recent data. In a complex process like admissions it takes detailed analysis to figure out what is going on. It’s certainly possible that abuses in the 80s and 90s have been corrected, but I would be surprised by that since the Asian percentages at the Ivies hasn’t changed. Of course it’s possible that the strength of the non-Asian applicant pool went up to compensate for fairer admissions practices for Asians, but it doesn’t seem plausible to me.

2011-12-17 on infoproc

“Harvard needs to filll its teams.”

Espenshade removed recruited athletes and legacies from his sample before computing the  relative SAT advantages of different groups and the odds ratios of admission at fixed SAT, GPA.

2011-12-16 on infoproc

I just checked for you. The top non-PRC kid is about +1.5 SD (top 6 percent or so) in the SAT-taking population, which is probably just a bit less than +2 SD in the general population (say, top 3-4 percent).

The SAT-M correlation with score in the class is about .55; Grit and score correlated at around .3

Grit and SAT-M are almost uncorrelated. So some linear combination of the two will give the best predictor. I’ll post these results “officially” on the blog at some point 🙂

2011-12-16 on infoproc

At +2 SD there are so many kids that the elite schools cannot even make a dent in the population. Do the math. At +3 SD you see a reduced number at state schools these days, due to the effects you mention.…

2011-12-16 on infoproc

We’re not only looking at SAT, but also “grit” (Big 5 Conscientiousness). We administered grit surveys at the beginning and end of the term.

FYI, the PRCers systematically *underestimate* their grit relative to Westerners. This has been reported before in other studies — Asians work harder but still consider themselves lazy, Westerners exaggerate their work ethic. (An advantageous psychology if you want to climb your way to the top…)

But from previous work I can already tell you that SAT-grade correlation will be .3-.6 as it is in almost all UG courses (we did this in our earlier paper). It looks like grit will add a bit of predictive power.

2011-12-16 on infoproc

PRC students and other foreign students don’t take the SAT, as far as I know. They are admitted through some different process at the moment.

This might explain my prior that these students would be not very good:…

2011-12-16 on infoproc

I’m not sure which argument you are talking about.

I’m not trying to make inferences about national populations as a whole. As you point out this depends on lots of selection effects. My main surprise in looking at this data was that these rich “slacker” kids are actually good students. Before I graded the final I hadn’t looked carefully at how this subpopulation was doing in the class, but I didn’t actually expect them to be doing well. Now I suspect that even the not so stellar kids (but admittedly from the upper class) in China are getting a pretty rigorous HS education, which is after all what PISA results say.

Perhaps you don’t have a lot of experience dealing with data. I’ve looked at UO aggregate SAT data and SAT distributions by major. You get nice normal distributions with slightly shifted averages. You can consult the data mining paper I wrote on this (use google) to see the means and SDs. It is very unlikely that physics 101 students deviate much from the distributions in the non-STEM majors here. (Physics 101 is actually considered *harder* than the other intro science courses that fulfill the gen. ed. requirement here, such as Geology 101 or Astronomy 101.) So yes, I do have some idea about the distributions, and the +2 SD estimate I made elsewhere is almost certainly roughly correct.

2011-12-16 on infoproc

But it apparently doesn’t decrease the correlation between their SAT score and college GPA. Does that mean they are also studying much harder in college?

I suspect you work in a place with a concentration of Asians, like Cupertino or San Gabriel Valley. There are plenty of Asians in other parts of the US that don’t have access to cram schools. You have to average them in to get to the aggregate numbers.

2011-12-16 on infoproc

I *give* them the formulas, but they have to understand the concepts to apply them. You haven’t been around “normal” (non-STEM) students trying to do quantitative stuff much, I guess.

2011-12-16 on infoproc

Sample size in theoretical physics is small, but I believe the answer is probably yes. It is also (I am told) yes in pure math and I am 100% sure the answer is yes in some fields like EECS. Note though this is including Asian immigrants from PRC, Korea, etc. who come here for grad school. If you just count A-A’s then the numbers are perhaps too small to make sense of, but see the link below.…

2011-12-16 on infoproc

What do you mean by “modest talent”?

What fraction of the population could get the high score in this class of 200 kids (assuming no PRCers)? I would guess the math ability average in the class is the same as in the general US population. That would be slightly below the average for university students. The top kid is then probably about +2 SD in the general population. (Note, as I mention elsewhere, there are kids in the class who have had AP physics but are non-STEM majors trying to get an easy grade.) Nevertheless, *half* of the PRCers outscored this hypothetical +2 SD person.

As I note at the link below, the top 10 percent of Shanghai students scored above the US 99th percentile on the math portion of PISA. These physics 101 results suggest to me that the PISA results are not unrepresentative.…

2011-12-16 on infoproc

I’m not surprised that people in some countries encounter this material in high school. It’s a scandal that in the US people can graduate from university without having learned basic physics. On the other hand, I doubt most adults in China or Korea or Japan could solve these problems — they’ve pretty much forgotten it all.

Also, I often get US students in the class who claim to have had AP physics in high school, and they don’t necessarily get the highest scores in the class.

2011-12-16 on infoproc

Almost none of these PRCers are STEM majors. This course doesn’t satisfy any STEM requirements — for that they have to take a harder physics course. The two PRC students who came regularly to my office hours this term were a business major and an architecture major. The architecture major was a senior (IIRC), so very far removed from whatever she learned in high school.

Even though this is physics for poets, the top few kids in the class in a normal term are usually quite sharp (we’re talking about 200 kids here!). What is surprising is that about half of the PRC students this term got *higher* scores than the usual top score in the class. No matter how much background you have in physics, that requires some g.

2011-12-15 on infoproc

Because we got a big influx of PRC students who are qualitatively different from American students in their ability to do basic physics.

2011-12-15 on infoproc

The PRC kids in my class are, by and large, not the academically strongest ones in their home country. Almost all of those kids would be going to upper tier universities in China, for a fraction of the tuition they are paying here.

2011-12-15 on infoproc

This is more like Conceptual Physics, but I do make them solve problems involving algebra.

2011-12-13 on thedealeconomy

Hmmm, … is it Hsui, Hsiu or Hsu?

OK, thanks for fixing the odd spellings 🙂

2011-12-12 on infoproc

With the gi this is really dangerous. If you get a straight armbar (juji gatame) on with full power using the gi you can easily snap someone’s elbow before they can tap. Happened to me once — months of rehab 🙁

2011-12-11 on infoproc

He should retire.

2011-12-11 on infoproc

There are two interpretations. One is that V is useful for self-aggrandizement, or claiming credit. The other is that the V score is correlated with the ability to make the breakthrough, see some conceptual connection, etc. I think both effects are there. Note for some fields a very high M score is a prerequisite to even get started. The only M scores quoted are for non-physicists, and interestingly (IIRC) the high scores belonged to a psychologist (psychometrician?) and a biologist (population genetics?).

2011-12-11 on infoproc

You need a high ceiling test to differentiate in the tail. See also SMPY, which is SAT administered to 12 year olds.…………

For ordinary careers, there are probably diminishing returns to brainpower, and other factors matter more. Of course, everyone is affected by luck.…

2011-12-11 on infoproc

Not sure what you think I am clamming up about.

Nor why you think I would not apply the same laws or principles equally to all ethnic groups.

An understanding of how the world actually is does not preclude wanting to make it better.

“Pessimism of the intellect, optimism of the will.”

2011-12-11 on infoproc

Some of us would prefer a nation of laws, not a nation of ethnic interests.

2011-12-10 on infoproc

Very cool! A lot of numbers to check, though …

2011-12-10 on infoproc

“…elite institutions started accepting Jews … because their incredible ability made it impossible to shut them out… ”

Part of that “incredible ability” was making the general public aware of what was going on — Jewish quotas, etc. — and making clear it was reprehensible and unfair. I’m part of the same strategy (blogging, talking to the press) on behalf of A-A’s. Amazingly, I seem to be one of very few people pointing out what is going on. Of course, that’s consistent with your characterization of Asians as not as aggressive as Jews  🙁

There are several issues being discussed here. One is what elite institutions should do in their self-interest (see link below). Another is what they should be *compelled* to do given that they receive quite a lot of Federal money. One could argue that it’s in their narrow institutional interest to cap the number of Asian admits. (See the recent AP article where I mention that >20% Asian fraction on campus might be jarring for older alumni.) But one could also argue that they are violating the civil rights of Asian applicants.…

Some people actually deny that elite schools discriminate against Asians. You don’t deny it, but note that it may be in their narrow interests to do so.

2011-12-10 on infoproc

I pretty much agree with what David says here. Note I only mentioned govt because Shawn had said he was interested in that. Seems soul-deadening to me 😉

2011-12-10 on infoproc

The point of the post was to address Thrasymachus’ question about whether SAT over-predicts college performance for A-A’s, which seemed entirely implausible to me and relatively easy to address statistically.

It’s much harder to address whether A-A academic performance translates into high end success later in life. I don’t know one way or the other, but I tend to agree with RKU’s comments on the previous thread.

If you believe in ethnic affinity operating at any level (including, e.g., subconscious), then it’s not surprising that in a majority euro country (Ashkenazi) Jews, S. Asians and E. Asians would be relatively advantaged in descending order. The degree of similarity (cultural, physical, psychological, genetic, etc.) to the majority population is ordered that way. I mean, who do you think these young IBers are trying to make rain with? (Also, see the charges of outright discrimination alleged by FanHsu — I doubt he is making things up, no matter how familiar and friendly you personally might be with Asians from your HS experiences. I’ve encountered many people who just don’t like Asians.)…

2011-12-10 on infoproc

Family SES and parents’ education are much weaker predictors of college performance than SAT or HS GPA. Controlling for SES and parental education (i.e., look at groups of kids with varying SAT but the same SES, or vice versa) has little effect on the correlation between SAT and college GPA. This was established by a huge study using UC data. It is exactly counter to the claim made by most SAT critics.

A few years ago Wake Forest stopped requiring the SAT in admissions. At a conference held on the topic, Kuncel pointed out how foolish and misguided they were. See the slide with the words “Minor reduction in test score predictive power when controlling for parental education and income.” This is kind of amazing because there certainly must be *some* effect on GPA from SES — for example, some kids have to work part time jobs during college and others don’t.…

See also…

2011-12-10 on infoproc

It’s a tough call. First thing, I’d say try to do something that actually interests you. Second, perhaps MBA heading into Fed track might work but I’m no expert. I do know the MBA credential from a 25-50 ranked school would be enough to get you a higher GS rank, and the Feds aren’t nearly as elitist as the tracks we’ve been discussing.

2011-12-10 on infoproc

I’m sure hours vary, but one thing about America: you’re very unlikely to make it to the top without busting your ass.

Re: tech startups, google around and you can find plenty of photos of sleeping bags under desks. How many other career tracks have that? Even rougher than that is the huge energy expenditure in: pitching potential customers / VCs / busdev partners, monitoring deals, putting out fires inside the organization, etc. etc. You have to have superhuman energy levels and determination to be a successful entrepreneur.

One difference though: in biglaw/finance/consulting you are usually grinding long hours under an asshole boss, often doing mind-numbing crap. At a startup you’re usually working toward a goal that you are passionate about, trying to build something cool. If you’re a founder, you are your own boss.

2011-12-10 on infoproc

20% more than who? All of these tracks: consulting, biglaw, IB, “hard” finance, startups, are super intense. I wouldn’t be surprised if biglaw and startup guys have the longest hours.

2011-12-10 on infoproc

I’m sure it’s possible. But how common is it? Studies I’ve seen with large sample sizes only show small effects.

2011-12-08 on infoproc

Responding to Maciano below: regarding career tracks at “soft” elite firms or at a big corporation, I think the payoff table is worse than the ones I give below for startup entrepreneurs and “hard” (money management, trading) finance jobs. Maybe IB has a better high end payout than the other soft tracks. To me it’s not attractive to be in a rat race for the best years of one’s life. The appeal of some of these other tracks is that perhaps you really enjoy what you are doing (esp. startups) and/or have a shot at making enough to retire early.

Individual is a smart, conscientious, driven kid with Tier 1 credentials. But no superpowers 😉

startup / tech

mid: ~$200k/yr, no hits, stuck working for big company as VP, no early retirement
90th: small hit, makes a few million, still working as above
99th: big hit, makes >$10 million, retires early, maybe rinse and repeat
99.9th: big, big hit, $100M payoff, livin’ large 🙂

mid: ~$400k/yr, lives in NJ, can’t retire
90th: ~$2M/yr for 10 years (mid-career), retires at 45-50 with $7-10M put away
99th: same as above but with several times higher net worth and earlier exit.

A professor comes out (financially) behind both of these tracks, but perhaps happier, more relaxed, more job satisfaction.

If we further restrict the group of individuals to Tier 1 PhDs in physics [who left for finance], it seems to me that the 90th percentile outcome in finance described above is more like 60-80th percentile.…

2011-12-05 on infoproc

Antti, thanks for the compliments 🙂

This paper might be of interest:…

Note though that in terms of pure entropy humans and other biological creatures are not so exceptional. I think perhaps you might be interested in measures of complexity or computational / logical depth. Evolved creatures tend to have great computational depth, as opposed to entropy.…

2011-12-03 on infoproc

The data I have see show that white/Asian grades are similar within errors (IIRC Asian grades were slightly higher). But Asians are disproportionately in STEM majors, and our UO data shows that STEM courses have lower average grades (less grade inflation) and are harder (the typical student has higher SAT and HS GPA than other majors). In humanities and social science courses taken by both STEM and non-STEM majors, the STEM majors get higher grades. If you put these things together it seems to say that Asians are actually stronger students at the college level, not just in HS.

2011-12-02 on infoproc

The data I have seen on this is not very good — there are GPA numbers but not weighted by course difficulty, etc. Asians end up disproportionately in the harder STEM majors. Anecdotally, the answer to your question is well known …

2011-12-01 on infoproc

Video is still the driver. 10^16 bytes = 10^4 terabytes is not that much. You could fit it in your house if you stuffed it full of terabyte hard drives 🙂

2011-12-01 on infoproc

For the case of humans, you can compress using a diff wrt a reference sequence. Individuals only differ at only a fraction of a percent of loci.

See, e.g.,…

Or Wikipedia: “The 2.9 billion[17][18] base pairs of the haploid human genome correspond to a maximum of about 725 megabytes of data, since every base pair can be coded by 2 bits. Since individual genomes vary by less than 1% from each other, they can be losslessly compressed to roughly 4 megabytes.[19]

The entropy rate of the genome differs significantly between coding and non-coding sequences. It is close to the maximum of 2 bits per base pair for the coding sequences (about 45 million base pairs), but less for the non-coding parts. It ranges between 1.5 and 1.9 bits per base pair for the individual chromosome, except for the Y-chromosome, which has an entropy rate below 0.9 bits per base pair.”

But this is a one-time gain. If we end up sequencing every person on earth, which according to recent trends would be economically feasible fairly soon, that’s about 10^16 bytes of data, fully compressed. It wouldn’t be a lot for each person to store, e.g., on their own thumb drive. But keeping it all in one place would be a bit daunting.

PS One of the reasons we can’t yet take advantage of this compression is that the sequence reads/assemblies have errors in them and consequently we have to store the raw data.

2011-11-24 on infoproc

The three versions of qm that I listed have different experimental predictions — at least in principle, if not in experiments we can currently perform. This was not appreciated for a long time, hence the original use of the term “interpretation” when what was really meant was different physical theories. Bell invented a term FAPP = For All Practical Purposes and FAPP the different interpretations agree.

2011-11-23 on infoproc

>> Hm: would you have confronted a Nobel Prize winning physicist if you had found out that he had committed serious sexual misconduct? << I can't say for sure what I would do in a hypothetical situation, but I would feel very bad about myself if I caught anyone doing what Sandusky was supposedly doing and didn't rescue the kid there and then. Precisely because he knew Sandusky so well I cannot imagine an ethical JoePa not wanting to know exactly what went on. Even if the situation is as you describe above, my opinion of JoePa is very low.……

2011-11-22 on infoproc

In your model, what are Joepa and McQueary thinking in subsequent weeks and years when they look at each other and at Sandusky as he passes them in the locker room?

Does Joepa think McQueary was on drugs and made the whole thing up? Then why is McQueary in one of the top coaching positions?

Perhaps you don’t believe McQueary’s testimony that he told Joepa all the details about the Sandusky encounter? Hard to believe “Saint Joe” didn’t want to know all the facts when he met with McQueary the next day.

2011-11-22 on infoproc

I don’t know much about their character at all. Chip Kelly, however, is a football genius 🙂

Even given the low ethical standards in big time college sports, I would guess very few head coaches and ADs would cover up an incident like Sandusky’s involving a 10 year old.

2011-11-20 on infoproc

Read what I wrote. What you call the “statistical interpretation” is not what the authors of the paper are talking about. Their terminology is probably confusing. No one is claiming to disprove what you call the statistical interpretation.

2011-11-19 on infoproc

The talk you linked to is related, but (judging from the abstract) emphasizes slightly different things. The result in the paper more directly addresses the question of whether psi is “real” or merely gives information about an underlying reality.

2011-11-19 on infoproc

See my comment below on Lubos. I think a lot of people misunderstand what the authors are trying to do. Probably their biggest mistake was to use the term “statistical interpretation”. They explicitly note in the paper (for the really clueless) that they aren’t trying to show that qm is really deterministic. In fact they entertain much more general classes of theories than ordinary qm, including stochastic hidden variables.

PS Ignore my “like” of the comment that this responds to. I hit the button when I meant to hit “reply” 🙂

2011-11-19 on infoproc

I think Lubos misunderstands what the authors are trying to prove. They are not addressing the probabilistic rules of QM (i.e., the Born rule) or how to do basic calculations. They are addressing a particular *interpretation* (“statistical interpretation”) in which it is asserted that the quantum state or wavefunction describes (for example) the state of knowledge of an observer as opposed to being a faithful representation of the state of the world (“reality”). This is an old idea, but in its modern incarnation is sometimes referred to as the Bayesian interpretation of QM (see papers by, e.g., Chris Fuchs: http://perimeterinstitute.c… ).

If you’ve never been exposed to this “qm formalism describes observer’s state of knowledge” interpretation (despite the fact that it is as old as qm itself), then it might be very difficult to understand what the authors are trying to do with their $\lambda$ object formulation. The authors show that it is quite difficult to avoid having the quantum state of a system ($\phi$ in their notation) determine uniquely the probabilities of all measurements. The tricky philosophical or interpretational question is whether this is enough to exclude the possibility that “reality” is more complicated than what is encoded in $\phi$ (or that $\phi$ only encodes the state of knowledge of an observer and the underlying reality is more complex).

2011-11-19 on infoproc

I think it means wavefunctions don’t collapse 😉

2011-11-17 on infoproc

Search under “marshmallow experiment” on this blog. You’ll find links to the original papers.

You might be interested in the two factor model (IQ + conscientiousness) we use in discussing college performance. The two factors are about equally important for most majors, but in math and physics it looks like there could be an IQ threshold. Search for “data mining university” on this blog. My colleague Schombert and I are currently collecting “grit” survey data on students to see to what extent it improves GPA prediction power over SAT only.

2011-11-16 on infoproc

Please volunteer. We have something like 500 right now but want to get into the multi thousands.

2011-11-16 on infoproc

I believe they are giving you magnitudes of changes, not the sign.

2011-11-15 on infoproc

> Paterno denied having been told about anal sex.

IIRC McQueary directly contradicts this in his GJ testimony.

The latest news is that McQueary claims to have stopped Sandusky in the act:…

In an email to former teammates, obtained by NBC News, McQueary said: “The truth is not out there fully” and that he “didn’t just turn and run” after seeing the alleged molestation.”I made sure it stopped. I did the right thing … you guys know me,” he wrote, adding that he “had to make quick, tough decisions.”…

I think when the dust settles Sandusky is clearly guilty and JoePa looks very bad.

2011-11-15 on infoproc

Genetic engineering is easier than singularity-level AI. I can see a clear path to the former, but no one knows how we will get to the latter.

2011-11-15 on infoproc

The dangers are obvious but I think it is inevitable.

2011-11-12 on infoproc

We aren’t announcing any results. These talks are mainly for recruiting and to get the community ready. The simple model you refer to is the usual one for additive genetic variance.

2011-11-09 on infoproc

If Paterno continues living this is all going to come out in the criminal trial and private lawsuits. Why did Sandusky not succeed Paterno as head coach, as was widely expected? What did Paterno know? How could Paterno and the grad assistant (now the WR coach) look each other in the eye year after year, knowing that Sandusky was still around, had an office in the building and ran football camps for young kids? Didn’t they wonder what happened to that little boy in the showers after 2002? Obviously there was no law enforcement action as Sandusky remained a free man. To say that Paterno discharged his moral responsibility by reporting a watered down version of the event to his “superiors” is ridiculous.

The whole thing makes me sick.

2011-11-06 on infoproc

I don’t know what study this is, but IIRC pre-meds do not actually have the strongest SAT scores or HS records among science majors. At UO it’s the physics and math majors.

2011-11-03 on infoproc

Nice! I’ve heard good things about Derman’s new book. I like Keynes’ comments on Newton’s powers of concentration.

2011-10-27 on infoproc

“The apparent randomness is actually evidence of expertise …”

But this argument does not apply to “expert” analysis of world events (see Tetlock). It could just be that markets (which are themselves affected by exogenous world events) are highly complex and nonlinear, so hard to predict even if still very far from perfect efficiency.

2011-10-27 on infoproc

Your faith in the market is excessive.

“do the same job for one tenth of the price but equal the productivity”

Kahneman gave the managers of the firm he studied evidence that they were paying for luck and could get the same performance at a much lower price. Because of what I like to call “bounded cognition” (but you could also refer to as a market failure), they didn’t act on the evidence (they could not accept it, just as you cannot; it was not a matter of arguing over statistical technique). The same applies to mutual fund investors who pay for active management (which underperforms indexing) year after year.

In my pneumonia example people with bounded cognition could easily ignore the statistical evidence that Ohio State MDs cure pneumonia just as effectively as Harvard MDs. The problem is that people in general do not understand statistical evidence very well.…

As you know, Kahneman’s work demonstrates over and over again that people deviate from “rationality” and this limits the effectiveness of markets.

2011-10-27 on infoproc

It seems to me that the 20th century trend in democracies is toward greater redistribution: social safety nets, guaranteed minimum income, etc. People have been conditioned to believe these are aspects of a just society.

The question is: what is the optimum level of redistribution? (Given a particular utility function for society.)

One argument is that we have to let the rich get rich in order to have strong economic growth. Too much redistribution means a smaller pie to split. But the Illusion of Skill argument (if correct) suggests that for some activities like finance a high marginal tax rate (say, which kicks in above the income of the *average* finance professional; this would then only affect the top earning financiers who, according to the argument are not adding any real value that the average guys can’t also provide) would not negatively affect economic efficiency.

If people irrationally and incorrectly believe that only Harvard MDs are capable of treating pneumonia, and bid up their compensation to exorbitant levels (levels so high that the Harvard MDs begin exerting financial and political control over society as a whole), wouldn’t it be better for society to impose a high tax rate on Harvard MDs, which kicks in above the income of other doctors with similar credentials (but who are not beneficiaries of the irrational belief)?

2011-10-25 on infoproc

“no better than those of the average professional”  Sorry, I should have written “than those of average professionals”. I’ve fixed it now.

“Money managers also manage risk more effectively than their investors would individually.” I’m comparing one professional against another, not against non-professionals.

2011-10-25 on infoproc

If you read the whole article, you see that Kahneman does believe in skill. For example, his studies show that some doctors are better at diagnosis than others. I am also sure that some entrepreneurs or some physicists or some athletes are better than others. (Although in the case of entrepreneurs it would be very hard to demonstrate statistically since outcomes are noisy and the number of attempts per entrepreneur is relatively small.)

But there may be areas where *the differences between high level professionals* (e.g., people who have been hired to run money, have top MBAs or graduate degrees, etc.) are statistically seen to be mostly due to luck. This has already been convincingly demonstrated for pundits or analysts of complex world events by Tetlock’s studies of expertise. (You can find several posts on this blog on the topic.) Whether it’s true of money managers (or even big company CEOs) is controversial. If you argue the skill side, I’d like to see *your* statistical evidence, not just repetition of your priors (again and again).…

In all areas of human activity, even the skill dominated ones, luck plays a big factor. This is a good argument for redistribution — almost every successful person owes some of their success to luck.

2011-10-23 on infoproc

I don’t know enough about this to have a strong opinion (high confidence level), but I thought it worth reporting what I learned from asking around.

2011-10-22 on infoproc

In many years of travel to/in China, and conversations with Chinese people worldwide, I have never met one who believed in a rural/urban or interior/coastal gap in (genetic) potential for intelligence. When I bring up such a possibility people just find it highly implausible. (Note these people I am talking about are generally quite open to HBD in general.) A common response I get about this is that anyone familiar with the history of China realizes that there has been an enormous amount of internal migration and population mixing over the centuries (and they don’t believe in the systematic selection effects you mention, with urbanites being drawn from the smarter rural population; at least, they don’t think it’s a big effect). Recall that during the cultural revolution many urban intellectuals were sent to the countryside to work alongside peasants. These people, even after close observation, do not tend to believe in any systematic difference in potential for intelligence.

2011-10-19 on infoproc

From your comments I suspect you don’t really understand the subject, but it’s possible I am mistaken.

To explain why I think the argument in a particular paper is wrong takes effort, and before I expend that effort I want to be sure I am not wasting my time. For example, if the author of one of those papers were to contact me for an opinion I would gladly reply to them. Posting a lengthy analysis just for your sake doesn’t seem a particularly good use of my time.

To be clear, I think some of the papers you linked to are actually emphasizing what I refer to as the subjective nature of probability in MW, so we are not in disagreement. You might try to read my paper again.

2011-10-19 on infoproc

I hope you are joking. Responding to your links to a few papers isn’t exactly high on my list of priorities.

If you are someone whose opinion I would care about, please contact me directly and identify yourself. I am happy to discuss MW or the origin of probability with you, as I have with many actual researchers in the field.

2011-10-17 on infoproc

See link for an opinion poll of actual experts (people who work on quantum coherence and decoherence) on whether the dynamics of decoherence is as claimed:…

See Weinberg’s latest paper for his view on the status of qm:…
Aside from the “extravagance” objection, he only mentions the origin of probability as an issue for MW. (So it might be fair to guess he would vote with the majority in the poll above.)

I don’t (personally) know anyone who thinks relativity is an issue for MW.

2011-10-13 on infoproc

Good point. I doubt my friend had ever met any 18 year old like Schwinger.

It’s also possible standards have gone up since 1938 😉…

2011-10-13 on infoproc

Vorrei poter. No, avrò un traduttore.

2011-10-09 on infoproc

Inflationary cosmology could get the prize. The observations already favor it; it’s a question of further observations and when the committee feels the evidence is strong enough. That work was done in the 1980s (Steinhardt was one of the contributors!) and has very significant fundamental implications for the large scale structure and dynamics of the universe. Depending on what we see at LHC, various models of particle physics that extend beyond the standard model could win prizes.

2011-10-07 on infoproc

I’m not against a balance between different metrics. I do think it’s funny that people (like US News) who do rankings have to engage in deliberate tweaking to keep HYP on top.

2011-10-07 on infoproc

Harvard is out for Harvard. Caltech is working for the betterment of Mankind 😉

Of course, idealism doesn’t usually pay off.…

2011-10-07 on infoproc

The standard thing in physics is to post the preprint on arxiv and then possibly submit to a journal, depending on the type of article. I do think we should eventually have a system of peer review or quality signaling that doesn’t depend on journals, but we are not quite there yet.

2011-10-06 on infoproc

IIRC in this case the low AFR women had 4 more offspring on average than the high AFR women. With a heritability like .3 or so (typical for behavioral genetics) you would get significant allele frequency and breeding value changes in 5 generations.

Of course if this is all culturally driven then the genes for AFR might actually be genes for submission to authority, or whatever. But changes in frequencies probably did take place.

2011-10-06 on infoproc

It would be a mistake to underestimate Harvard’s strength in math and natural sciences, where almost every department is among the strongest in the world.

2011-10-06 on infoproc

They must have *some* evidence against cultural effects, like adoptees in the records who resemble their biological moms more than their adopted moms? I still haven’t had a chance to read the paper…

2011-10-05 on infoproc

I’ve thought about anthropic possibilities, but it seems to me that you could have a moderate level of deviation from the Born rule (i.e., “improbability” according to the conventional qm probabilities), without losing the possibility of life (or intelligent life). So, even weighting branches by their “friendliness to life” would still suggest a fair amount of observed Born rule violation for typical branches (like, presumably, ours). It’s possible we’ll see such improbability, but so far I think the Born rule is well supported in, e.g., decays of elementary particles.

2011-10-05 on infoproc

If you look at the books in ref. [2] of the paper you can find some simple examples where a a state is decohered due to, e.g., scattering of air molecules. Perhaps that will help …

2011-10-05 on infoproc

> the number N of ‘decohered branches’ increases with t.

but this only brings us back to my original question, how can N increase with time if the Schroedinger evolution is unitary. < As I mentioned already, coarse grained entropy (counting of number of decohered branches) can increase in time even when unitarity (which is defined at a more microscopic or fundamental level -- i.e., the actual Hilbert space for all dof) holds. Put another way, the dimensionality of the Hilbert space is not increasing due to or during decoherence. The whole point is that decoherence itself is a result of unitary evolution. It might help if you try to keep track of all the degrees of freedom (e.g., electrons in the detector or memory device) to convince yourself that M+ and M- can be radically different states (i.e., decohered), yet both states were present from the beginning as possibilities in the original Hilbert space describing the apparatus. You're not really asking a probability question, you are confused about MW and decoherence at a very basic level.

2011-10-05 on infoproc

The same number. Look at equation 4.

Of course the branches are only decohered after the measurement. Before the measurement it’s just a choice of basis.

2011-10-05 on infoproc

Thanks! Looking at Deutsch’s latest but no comment yet.

2011-10-05 on infoproc

Schrodinger evolution is unitary. Put very crudely, the various versions of the observer were already there if you write the initial state in the suggestive way that I did in the last equation.

However, the so-called “coarse grained” entropy can increase in time even in many worlds. That is the apparent entropy for an observer who can only see one decoherent branch at a time. This observer sees entropy added in each decoherence event: to keep track of his world history he has to add a bit to his memory record to keep track of the + or – outcome.

See my paper “what is the entropy of the universe” for more discussion (there are a couple of related blog entries).

2011-10-01 on infoproc

Spanish villa! Very tempting … 🙂

2011-10-01 on infoproc

If they kick out Greece and a few other countries, the remaining core economies (and hence currency) will actually be stronger.

2011-10-01 on infoproc

Portfolio mgmt advice, please? 🙂

Risk-off, all cash, beared up?

2011-09-29 on infoproc

trust us … we care 🙂

2011-09-15 on infoproc

Yeah, the guy who started two tech companies (including one that made encryption technology which was, among other things, used against the Chinese Great Firewall) and is risking his reputation with the PC powers that be by running an IQ GWAS is unimaginative and obedient. You, on the other hand, show your creative maverick nature by commenting anonymously on other people’s blogs.

2011-09-15 on infoproc

Yes, one could try to measure “value added” by the school. One proposal (see Hanushek at Stanford) is to measure value added for individual K-12 teachers by looking at improvement in scores over the course of a year. In the same way, if you had an objective measure of learning (perhaps the GRE subject test), you could measure how well graduates of a particular school perform relative to SAT score, and perhaps attribute the residual to the school. In the US GPAs are not standardized across universities.

2011-09-15 on infoproc

IIRC the effect is still there when you look major by major.

2011-09-15 on infoproc

I kind of eat paleo (low carb, high protein, not as much fat as some paleo types seem comfortable with, lots of vegetables and fruit). I do crossfit-like training. I don’t really believe the evolutionary justification for the whole thing (“just so” stories for people with low levels of rigor). For example, I view the diet as a simple optimization: what should I eat on a fixed calorie budget to minimize hunger? Different people will end up with different optimizations.

Re: MMA, the guy is a typical paleo thinker, but like you I agree with a couple of his points.

2011-09-15 on infoproc

> …biting, gouging, rabbit punching, below-the-belt shots and other techniques that are not allowed in any dojo but are definitely way more effective. < I hear this all the time but it's a myth (though not the part about rocks or spears). If you modify the rules to allow the techniques listed above it won't change things very much. Takedowns and positional control still work, and the superior grappler will be the one who takes advantage of biting, gouging, etc. There's an old video (I have it somewhere on VHS) that shows one of the Gracies biting off the ear of someone who tried to bite or gouge him during a street fight in Brazil. If I get superior position on you you'll be very sorry that the rules allow these things.

2011-09-15 on infoproc

Haven’t had a real fight since I was a kid. Haven’t trained seriously since having kids, but gave my nephew some coaching recently during a family vacation and really enjoyed it 🙂 I’m pretty sure I could still clean up on anyone who didn’t have a significant background in one of the more practical styles.

I haven’t trained for swimming since college, but when I go to the pool with my kids I get a few laps in and I doubt anyone who didn’t swim in HS or college could beat me in a sprint.

2011-09-14 on infoproc

In our sample I would bet women average higher conscientiousness overall. This might be just an age thing — men mature later.

2011-09-14 on infoproc

BTW, just as BJJ optimizes for the best algorithm for one person to take out another (takedown, position, submission), in our “dojo rumbles” we learned what the optimal technique would be for two guys to take out one. It turns out it’s still BJJ: one guy tackles or clinches, the other guy follows through from above. There are some nasty submissions you can apply to an opponent who is already locked up grappling with your partner. Striking is very overrated. Easy to break your hand, and people can take a lot of punishment before they quit. Just watch the early UFCs.

2011-09-14 on infoproc

Yes, BJJ is really for unarmed dueling. Judo (and to some degree, wrestling) is a bit more realistic for other situations since it can be used to stay on your feet or to quickly drop an opponent. When I trained at a BJJ club in Tokyo (the teacher was a Japanese Brazilian from Gracie Barra), we used to occasionally train for many on many group fights. This is apparently a Brazilian thing and I’ve never seen it at any other school.

2011-09-14 on infoproc

I agree. The *desire* to be dominant is distinct from the context-dependent abilities required to actually dominate. The WSJ article is defining as alpha people who have both the desire and the (business context) abilities.

2011-09-13 on infoproc

You can be alpha in one setting and beta in another. In a business or intellectual setting Gates is alpha, but on the playing field he’s obviously beta.

2011-09-12 on infoproc

I met Gross briefly a few years ago when I gave a talk at PIMCO. No giddiness, but I did invest some money in PIMCO funds which haven’t exactly outperformed since 🙁

The man can do yoga. I saw him at the gym, balanced on his head like a real yogi.

2011-09-11 on infoproc

I hate the sound of Cantonese, and I find the HK culture hyper-materialistic and shallow. It’s also too hot and humid in the summer.

2011-09-11 on infoproc

When I was putting the post up, I was hoping you (specifically) might chime in to tell us whether Gross is correct on the facts!

2011-09-08 on infoproc

Yes. Take it from someone who has worked in information security.

2011-09-07 on infoproc

I don’t think you need the extra degree as a credential. Probably enough just to take whatever CS courses you think you need or are interested in.

2011-09-07 on infoproc

I can’t read Chinese either. Someone explained to me what the sign says — something like “Noodle Room”. It’s not a joke, but rather a place in the cafeteria where you get noodles and dumplings. The contrast with the high tech Gattaca background seemed strange to me. Maybe the noodles are made from genetically engineered wheat 😉

2011-09-06 on infoproc

I often say the same thing: Quit now unless you are very sure you can’t be happy doing something else.

Note the person who originally asked for advice has already had a career in software and is going back in (to biology, not physics) with his eyes open.

2011-09-05 on infoproc

As a kid I preferred LOTR to the Silmarillion, but I read the latter again as an adult and now find it a much deeper and more powerful work — as if Tolkien had single handedly written the Bible or equivalent creation myth for his own universe. I read the Silmarillion again because I was staying at a university guest house in Tokyo for an entire summer and there was very little in their English library. Perhaps now I can stomach the Dune sequels.

2011-09-05 on infoproc

I’m pretty sure these appendices were there in the 1970s edition that I read. The book appeared in 1965.

2011-09-02 on infoproc

The population was both SMPY and SVPY, and the cutoff for the group from which the plot was constructed was 1 in 200 in at least M or V. So the below average individuals in either M or V might be fairly unexceptional (e.g., +1 SD in the general population).

>A lot of science requires no more than arithmetic

This is becoming less and less true over time.

2011-09-01 on infoproc

I wouldn’t be very useful in a physics lab, let alone a chemistry or bio lab!

But that has nothing to do with certain people claiming to “understand” evolution when in fact they may not. This despite the fact that important results are available for them to learn by merely picking up a classic textbook, attending a course offered at their own university or even reading a Wikipedia entry… I have no explanation other than bounded cognition and hubris.

2011-08-30 on infoproc

Yes, one of the main points of the paper I cited is that one can have strong epistasis at the level of individual genes, but if variants are rare, the effect in a population will be linear.

“These two examples, the single locus and A x A model, illustrate what turns out to be the fundamental point in considering the impact of the gene frequency distribution. When an allele (say C) is rare, so most individuals have genotype Cc or cc, the allelic substitution or average effect of C vs. c accounts for essentially all the differences found in genotypic values; or in other words the linear regression of genotypic value on number of C genes accounts for the genotypic differences (see [3], p 117).” [p.5]

2011-08-30 on infoproc

Interesting paper! They are swapping entire chromosomes in mice/rats. This is not exactly what I would consider as the “linear regime” (i.e., where we merely alter a single QTL at a time). In their discussion they note the qualitative difference between their results and what is found in humans and other model organisms.

“Such striking epistasis is rarely detected in humans and model organisms (1, 2, 16–21). One possibility may be that the statistical power to detect pairwise epistasis is typically low both in segregating populations and in crosses with multiple segregating epistatic loci that can obscure pairwise effects (21). Another possibility may be that the genetic architecture varies substantially among traits. In the present study, for example, QTL effects were generally smaller and epistasis weaker for bone traits than for blood and metabolic traits (Table 1). A third possibility is that the study design strongly influences the picture of the genetic architecture obtained. For example, genome-wide association studies of height in the human population, a prototypic quantitative trait (22), revealed many loci with small and additive effects, with little evidence for epistasis (23–25). These human studies necessarily involve large, genetically heterogeneous population samples and are better powered to detect common variants of modest effect than rare variants of larger effects. By contrast, CSS studies measure aggregate effects of whole chromosomes and they involve allelic comparison between 2 defined genetic backgrounds regardless of allele frequencies in the population. Moreover, CSSs enable genome surveys of pheno- typic effects in genetically defined individuals, rather than averaging QTL effects across a heterogeneous background.”

2011-08-29 on infoproc

The male SD is perhaps 10% greater than the female SD. You can correct accordingly.

2011-08-22 on infoproc

The model in the slides is just a toy model to illustrate scaling. In reality there will be distributions in effect sizes and allele frequencies in a particular population. See the height results which are starting to flesh this out for a different quantitative phenotype.

2011-08-18 on infoproc

How many people have super high LSATs but can’t make the SAT/ACT cutoff?

I want to stress again that people can qualify without meeting any of the automatic criteria. The survey form on the site is sufficiently general that you can make your case: e.g., I got 180 on the LSAT; here is a scan of my score report …

2011-08-18 on infoproc

You can’t remain anonymous to us but your identity is protected under the privacy agreement. See web site.

2011-08-18 on infoproc

Those are just automatic criteria. USAMO would probably get someone in as long as their other scores are somewhat close to the cutoffs.

2011-08-18 on infoproc

If you miss the automatic criteria you might still get in based on supplemental information. The form allows you to submit additional information and documentation.

2011-08-18 on infoproc

Not sure. There’s a place in the survey where you can submit your LSAT score (and other additional information), but please submit your other scores or other academic information as well.

2011-08-18 on infoproc

Please do! 🙂

2011-08-18 on infoproc

Thanks, it’s on our to do list 🙂

2011-08-17 on infoproc

1. The first stage of genotyping will involve SNPs only; in the future some or all participants may receive more extensive genotyping such as exome or full genome sequencing.

2. There are no restrictions on ethnicity or citizenship. However, if you reside outside the US or Canada there may be some delay in receiving a saliva dna kit from BGI after qualifying for the study.

2011-08-15 on infoproc

Nice! IIRC X was pretty big in the LA punk scene in the 80s 🙂

2011-08-15 on infoproc

No love for Richard Gere and Jerry Lee Lewis? 🙂

2011-08-14 on infoproc

I actually posted on this — do a search and you’ll find several things on hypomania, entrepreneurism, etc. on the blog.

2011-08-13 on infoproc

It would be a break from tradition, but I can imagine him winning a prize in Physiology or Medicine.

2011-08-13 on infoproc

We’ve known each other since college.

2011-08-12 on infoproc

Court order?

2011-08-11 on infoproc

Plenty of purely informational products have huge economic value. Search results are an obvious example. In addition, I would pay for book or movie or restaurant or gizmo or plumber or physician recommendations that added real value to my decision process.

2011-08-10 on infoproc

The general formalism is basic population genetics (lots of good textbooks, e.g. Lynch and Walsh), probably first worked out by Fisher. I don’t know of any detailed work on IQ per se. The best treatment of fast human evolution (albeit a popular exposition rather than a technical one) is of course Cochran and Harpending’s book The 10,000 Year Explosion 🙂

2011-08-10 on infoproc

Oops, did you leave out the .5 additive heritability? It looks like your formulas assume heritability = 1.

That may be why your answer is a bit more than twice my estimate (lower bound).

2011-08-10 on infoproc

Thanks! I should have checked my estimate (I knew it was a lower bound on the rate of increase) but I was too lazy 🙂

2011-08-10 on infoproc

I don’t agree with everything Clark says. What’s most interesting to me is the demographic history he uncovered. But IIRC the reviewer clearly doesn’t understand population genetics.

2011-08-10 on infoproc

It’s possible this is just some subpopulation, i.e., Manchus, who were undergoing population growth during the period. Some other groups (like subjugated Han) might have shrinking population over the same period. I’m not an expert on Chinese history; IIRC you can find some of the references on Google books if you are interested.

2011-08-10 on infoproc

Sure, but the question is whether there was some sustained period during which, on average, the correlation between X and reproductive rate was positive. It’s not implausible to me that over thousands of years you can get a 1 SD change.

2011-08-10 on infoproc

Sure, some more complicated effect could mimic the signal from heritability. But this study is going to be replicated using different populations from different places and eras.

In this paper the main population was from the UK, born in the early to mid 20th century and the replication population was from Norway, aged 18-79.

If a similar study done on people from very different places and times yields a similar value for the narrow sense heritability, that would strongly suggest that the straightforward interpretation is correct.

This is a complicated topic, so more studies are necessary.

2011-08-09 on infoproc

Not to beat on cosma, but you might have a look at his oft-cited blog post on iq and heritability. You will find further links there.

2011-08-09 on infoproc

You are missing the distinction between a global fit of total variance accounted for and the identification of a specific locus as iq associated. This paper is not claiming any (specific gene) hits so there are none to be replicated. The overall result for total variance accounted for by SNPs (lower bound on narrow sense heritability) needs to be replicated, of course. But it’s a familiar story now because of the height results.

See my earlier post on this and the even earlier post on height. Eventually you will find visscher’s paper which explains the statistical technique. You are not the only one who is confused. The genomics community is still obsessed with “missing heritability” because they don’t understand the distinction above. It’s yet another example of bounded cognition.

2011-08-09 on infoproc

It’s not our work. But I knew about their result ahead of time.

2011-08-08 on infoproc

No disagreement from me.

2011-08-07 on infoproc

You’re talking about McDojos here, not modern MMA.

Take two athletes of roughly equal ability but no MMA (nor wrestling, judo, etc.) training. Give one of them 6 months of training (say, 6 hours per week with a good instructor; the other guy does some basic fitness training 6 hours a week but no combatives). Lock them in a cage. I’d bet 3 (maybe 5 or 10, if he has some feel for grappling!) to 1 on the guy with the MMA training.

2011-08-07 on infoproc

To be a pro fighter you have to be very conscientious and reasonably smart. The sport is technical. You have to understand your weak points, plan and train accordingly. You also have to constantly watch how other fighters train and evaluate whether what they are doing will work for you. Even weight cutting (done properly) is very scientific and requires a lot of self-discipline.

One good technical example is the “arm-in” guillotine, which was thought to be impossible 5 years ago but has now become a standard and fairly effective technique. The sport actually innovates at a surprising pace.

2011-08-07 on infoproc

I didn’t actually do any work 😉

2011-08-06 on infoproc

See “Gracies in Action” video.

2011-08-05 on infoproc

I agree with you about where ideas come from, but not about how easy they are to implement.

If it’s a really new idea other people often won’t believe or support it. Sometimes the new idea is immediately shown to be correct or advantageous, but that is rare. Usually significant effort is involved and this is where people falter.

The personality characteristics of entrepreneurs who are willing to take risks and endure rejection (often for years) to make their idea a reality do show some regularity.

2011-08-04 on infoproc

To me his life is more interesting than his art.

2011-08-03 on infoproc

It’s possible the study I quoted is nuts and the correlations are too high. The second study I mention in the note added gets .7 for single measurements at 7 and 17. I’m still a bit surprised that you can predict adult IQ from a measurement at age 7 so well. I thought the correlation would be even lower.

2011-08-02 on infoproc

Not arguing that’s the whole effect, just that if you have a noisier measure of g early on it will depress the calculated heritability. There might be lots of other stuff going on…

2011-08-02 on infoproc

Apparently IQ stabilizes at age 6 only if you use repeat testing to beat down the error. If you don’t (and I suspect most twin/adoption studies do not), then there is an initially large error term that decreases with age. (Single measurements of IQ are much noisier at an early age.) This would look like increasing heritability but actually it’s just a reduction in the “other” error term (usually ascribed to non-shared environment as it doesn’t correlate with other variables like SES).

If what I just wrote is nuts it’s because I just spent two hours buying a new bicycle, helmet, booster seat, and assorted other stuff for my kids 8-/

2011-08-02 on infoproc

Sorry, I think I was confused about the terminology.

If var(T) = 1 (in units of population SDs) and (let’s assume) var(E) = 1 just for convenience. (Not completely crazy for a little kid.) Then reliability is 1/2. If I test 3 times it decreases var(E) to 1/3, so reliability increases to 3/4 (i.e. 1/(1 + 1/3)). Then the correlation goes up by sqrt(3/4 / 1/2) = sqrt(3/2) which is about 22 percent. (Unless I am still confused.) Not as dramatic, but what is observed is probably a combination of the averaging and maturation of the kid.

If var(E) = 2 for a little kid then reliability is 1/3 and the averaging increases it to 3/5. That increases the correlation by sqrt(9/5) = 1.34.

2011-08-02 on infoproc

Re-test reliability is just over .95, so it seems that var(T) = .05 or a bit less is reasonable. On the other hand for a young kid the error in testing could easily be .5 SD or var(E) = .25 (not sure if I am following your notation properly).

Then, decreasing var(E) by 3 causes a big change in rho, which increases by

sqrt[ (.05 + .25) / (.05 + .25/3) ] = sqrt [ .3 / .13 ] = 1.5

So the averaging of 3 results could plausibly (using my numbers) increase a correlation of .5 to .75 or so. If var(E) were larger (say .7 SD or var(E) = .49) you could get .8 or so.

2011-08-01 on infoproc

What values do you have in mind for sigma(T) and sigma(error)? If the latter term dominates then the increase in correlation is big = sqrt(3).

2011-08-01 on infoproc

Sometimes people get the correct answer by guessing. So, for example, some amount of luck is involved in any multiple choice test.

2011-08-01 on infoproc

Have you ever tried getting a young kid (e.g., 6 years old) to do something you want them to do? Say for a full hour?

2011-08-01 on infoproc

I don’t have the paper so I can’t say.

2011-08-01 on infoproc

The point is that “ability” is even less well defined when you are talking about a 6 year old kid whose mind easily wanders off from the test he/she is taking. This problem goes away as they get older, but repeat administrations might also help.

2011-08-01 on infoproc

It may not all be statistical noise. It could be the kid is learning how to focus on the test and so the score better reflects ability.

IIRC it wasn’t Eysenck’s research. He just quotes the result in the book.

2011-08-01 on infoproc

It could simply be a fluctuation (luck). That’s not the same as faking.

2011-08-01 on infoproc

I might have admin privileges that ordinary users don’t. I seem to be able to edit my comments after posting them. I think I added the sentence you mentioned a day or two later.

2011-08-01 on infoproc

If anyone is doing this please let us know what you think and how it works. I’d consider doing it to improve my Mandarin.

It seems like the best way to do this is web-based. You can make it a browser, FB or G+ app 🙂

2011-08-01 on infoproc

Here’s what I have from fig 4.7 in Eysenck’s Structure and Measurement of Intelligence. This is using data in which the IQ was tested *three times* over the interval listed and the results averaged. A single measurement at age 5 would probably do worse than what is listed below. Unfortunately there are only 61 kids in the study.

age range                 correlation with adult score (approx. value read from graph)

42,48,54 months           .55
5,6,7                            .85
8,9,10                          .87
11,12,13                       .95
14,15,16                       .95

The results do suggest that g is fixed pretty early and the challenge is actually in the measuring of it as opposed to secular changes that occur as the child grows up. That is consistent with the Fagan et al. paper cited above. But it doesn’t remove the uncertainty that a parent has over the eventual IQ of their kid when he/she is only 5 years old.

2011-07-31 on infoproc

Yes you can correct for that if you have a sample. How does that help the parent of a *single* 5 year old who receives a noisy score on an IQ test?

2011-07-31 on infoproc

What is “corrected for unreliability”?

2011-07-31 on infoproc

Nope, I was very precocious. I knew all about g and Terman when I was very young and have been waiting a long time for the genetic technology to catch up with the obvious questions. I took my first class at the university at age 12. But the facts are the facts re: regression and predictive power of very early testing.

2011-07-31 on infoproc

Not very reliable. I don’t have a quick link for you but score at age 5 is not a great predictor of adult score. Since you are interested in this topic you might look for a longitudinal study (book) of Hunter College’s gifted K-12 in NYC which admits using IQ scores at the kindergarten level, and again at the HS level. There is tons of regression among those admitted early. The kids admitted later tend to be stronger because the filter is more reliable. This is a common story with any kind of testing. SMPY showed lots of regression between age <13 and eventual HS SAT score. Put another way, precocity is correlated with adult IQ, but not all people with high adult IQ were exceptionally precocious at age 5.

2011-07-31 on infoproc

The Barry-Kongo fight showed someone can wake up from a flash KO and still win the fight. But I wouldn’t blame the ref for stopping it when such a KO happens. Hendo might have really hurt Fedor if the fight had continued, although there might be a tiny chance that Fedor would have recovered and won.

2011-07-31 on infoproc

Yes, actually one of the big advantages of the HP virtual conferencing rooms (do you have them at GS?) is that the optics are supposedly set up so that you can do eye tracking — you can see which person on your side of the conference a person on the other side (who you see on the HD wall screen facing your half of the conference table) is actually looking at. This apparently makes communication much smoother and it’s a good indication of how complex human interactions are.

2011-07-29 on infoproc

I wish I didn’t have to travel so much. But believe it or not, to get anything done involving organizations or teams of collaborators there is no substitute for two apes meeting face to face. Even for highly abstract subjects like theoretical physics, in-person interactions are far more effective than the alternatives.

2011-07-29 on infoproc

If you remove just a few counties from the analysis — including NYC, Greenwich, Seattle and Silicon Valley — the change in US income inequality over the last decades mostly goes away. That should give you a hint about what is driving it.…

2011-07-29 on infoproc

No, I don’t even know what those are. Could you explain?

2011-07-29 on infoproc

I do a very compressed workout, about 35 min 3x per week. Sort of crossfit/tabata with weights/freebody stuff and then 20min of run/bike. I feel like crap if I go more than a couple of days without working out!

It’s also very helpful that I can do this workout at home — I’ve got dumbbells, pull-up bar, stationary bike, etc.

2011-07-28 on infoproc

No, they mostly live in different worlds as you point out. The guys on the floors way above me are unlikely to return my phone calls! 😉

2011-07-28 on infoproc

Look again at the wiki link. The number of millionaire households in the US is not known with precision. Another estimate has it at 3 M or so. Nobody really collects rigorous numbers on net worth.

2011-07-28 on infoproc

Amazingly, I think I’ve met at least one person from each floor!

2011-07-28 on infoproc

Maybe an hour or two per day if you sum up longer chunks with physical media (books, magazines) and time on the internet. I follow a few sports like college football, track, MMA, Crossfit. I read a lot of blogs and several periodicals: NYT, WSJ, New Yorker, Atlantic, WIRED, Economist, (in order of decreasing consistency). Once I find a first rate or incisive/original brain with a blog I tend to follow them at least intermittently to see what they are thinking 🙂 One way people keep track of things is with a well-organized Twitter feed. I’ve been meaning to do this but haven’t gotten around to it and now might try it in G+ instead. I do have most of the above feeding into Google Reader. Of course most of my time is spent thinking about theoretical physics (or related stuff like genomics, which I am currently working on) or taking care of my kids 🙂

I read pretty fast so I often skim entire books while at the bookstore. I’m bummed about the closing of Borders and B&N stores.

2011-07-28 on infoproc

Sometimes I find them myself, sometimes readers send them to me.

2011-07-27 on infoproc

Yes, the Yale experimenter is measuring it in the lab. QM fluctuations as in Casimir can lead to violation of the null energy condition (e.g., “negative energy” in certain frames), which is necessary to stabilize a wormhole (Raychaudhuri eqn). Sorry if that’s too compressed — it’s explained, e.g., in our paper.

2011-07-27 on infoproc

The sports car (jag) is actually his, whereas the bug was a rental 🙂

2011-07-27 on infoproc

AFAIK only the excerpt I linked to.

2011-07-27 on infoproc

Good. I was worried the CME knew something I didn’t 😉

2011-07-24 on infoproc

On the other hand he has sections in the manifesto about preservation of Nordic genes, classification of mixed race people, etc. Was he trying to convince Stormfronters that they were fundamentally wrong, or that they needed to tweak the (public face of their) ideology to have broader appeal? I have to say he seems pretty open in his manifesto — see the personal comments about his friends and family. I’d guess he is actually stating his real beliefs throughout.

2011-07-24 on infoproc

Thanks for the Lind reference. I wasn’t familiar with his stuff — now it looks like KMac is channeling Lind but with a more anti-semitic perspective. Sometimes the trail of obscure scribblers extends far into the past …

2011-07-24 on infoproc

The immune system is almost certainly under stronger selection pressure and evolves faster than the brain. It is competing against tiny organisms that are themselves evolving very fast.

2011-07-23 on infoproc

In Tiananmen Square the naive student protestors shouted to the soldiers that “Chinese people do not hurt Chinese people” 🙁

2011-07-23 on infoproc

re: dissimulation, he may be saying to his fellow conservatives on Document that they need to adopt the Vienna School “anti-racist/pro-homosexual/pro-Israel line” in order to succeed, but not because it is what they really believe. He says they have to fix the ideology properly to succeed, but that is not the same as saying it is the *correct* ideology. He goes through an analysis pointing out that movements that do not embrace this line cannot win broad support. So I am unsure to what extent the line is simply necessary for pragmatic reasons as opposed to what he really embraces internally.

2011-07-23 on infoproc

You might be right. I don’t think I can peer into the mind of this nutcase with any precision.

Can you point me to non-KMac discussion of the FS that places them (*specifically FS*, not general beliefs about cultural relativism, etc.) at the center of multiculti as a tool against white / euro-cultural interests? I can find plenty of academic stuff which is against FS, but it doesn’t make the broader case that, IIRC, KMac spends a lot of time on in his book(s).

2011-07-23 on infoproc

I could be wrong about this and I’ve edited the post to better reflect the uncertainty. But his portrayal of the Frankfurt School seems very MacDonald-esque (and it’s MacDonald who often cites Salter, IIRC). If you keep searching you’ll find a very telling Breivek comment about tactics which suggests (at least to me) that this “anti-racist/pro-homosexual/pro-Israel line” could be superficial realpolitik and may not reflect his inner views. He also cites a VDARE article at some point. I would be surprised if he had not read MacDonald.

2011-07-23 on infoproc

Exactly. There’s plenty of stuff that can’t be expressed by the mainstream media. I want to hear the “extreme” opinions and the arguments supporting those positions.

The probability that all mainstream views at any moment are completely correct is extremely low. Some ideas which are widely accepted today will turn out to be false.

2011-07-23 on infoproc

“Extremist” just means far from the center or mainstream. It doesn’t mean the ideas are actually wrong. Some people refer to my blog as extremist.

2011-07-23 on infoproc

I find these extremist blogs interesting because they are a window into what some people (admittedly an extreme subset, but I don’t think it’s that small) really think about these controversial issues. RKU has made the same observation in the past. I’d actually like to see the comments that Sailer moderates out.

I think you and Yan Shen should declare a truce. You both have interesting things to say but I don’t want you guys flaming each other in my comments section.

2011-07-23 on infoproc

Breivik was almost certainly a reader of the blogs you mention. For example, in at least one of his comments online he references a VDARE article he particularly liked.

He’s also a tech entrepreneur and (obviously) a socio/psycho-path. In terms of psychological profile (not specific ideology) he reminds me a bit of Julian Assange. (See Assange’s blog from years ago.)

I’m guessing he did this so that he would have a platform from which to expound his political philosophy. The spotlight will certainly be on him when he goes to trial. But his actions are so horrific that very few people will listen to what he has to say. If he has slightly Aspie tendencies he might not realize this.

2011-07-21 on infoproc

The study you linked to addresses a somewhat different question. These kids were initially very deprived and had low measured IQs when adopted at age 4-6 — probably more like the group that Turkheimer studied.

“They had all been (i) neglected and/or abused during infancy, having been definitively removed from their biological family by court order after judicial procedures; (ii) placed in many foster families and/or institutions before adoptive placement; and (iii) assessed by a psychometric test that provided an IQ < 86 and > 60 in the year preceding adoptive placement.”

There may have been specific actions taken by the high SES adoptive families to counteract the previous deprivation that simply weren’t available to the lower SES adoptive families.

Because the initial IQs were tested so early the gain in IQ is a pretty noisy observable. The main result is that the gain was 13 points larger if the kid was adopted into a high SES family vs a low SES family. Sample size here is kind of small, like 20 kids in the low/high SES groups. Taking the study at face value I would say it’s evidence that recovery after early deprivation is possible but that it takes resources.

2011-07-21 on infoproc

I’m not familiar with the French studies, but Turkheimer is sometimes criticized for using IQ scores of fairly young subjects. No one knows whether his reduction of heritability persists into adulthood. In this study the measurements were done in adolescence or later. Also I believe Turkheimer’s sample included cases of more severe deprivation than would be found in this study. The effect may be nonlinear in the extremes: Tiger Moms at the far + tail and severe deprivation at the other extreme might lead to big effects. But in the less extreme regions of shared family environment I would say the evidence supports only weak effects at most.

2011-07-21 on infoproc

Of course I hope my children will never have to fight in a war. Perhaps my wish that my son be tough enough to do so if necessary is an anachronism …

2011-07-21 on infoproc

I was a competitive swimmer from age 7 through college, and my kids will probably also swim. But I think the West Pointers know what they’re talking about … 🙂

2011-07-21 on infoproc

No helmets probably means less brain trauma. In US football your head becomes a weapon once you are wearing a helmet. After a game you should have paint marks on your helmet from the opponents.

2011-07-20 on infoproc

I wouldn’t want my kid in full contact MMA or boxing. Training with headgear, big gloves, and limited head strikes is probably OK. You need to train with some strikes so that your grappling technique is realistic, but real MMA is probably as bad for your brain as football.

2011-07-20 on infoproc

I doubt it’s the Harvard influence. It may have more to do with talking to VCs and business meetings. At one point people used to refer to a “Harvard style” of seminars in theoretical physics, but by now I think it is called the “American style” by scientists from abroad.

2011-07-20 on infoproc

Oops, fixed!

2011-07-18 on infoproc

The field draws from a much larger pool today than it did at the beginning of the 20th century. Not sure that incentives and such are right for producing the best research outcomes, though.…

2011-07-18 on infoproc

There are really many questions: Do we understand what controls recombination at the mechanistic level (e.g., PRDM9)? Why would the rates and characteristics (e.g., size of hotspots) vary between groups? Is that just an accident (founder effect?) or are selection pressures at work?

2011-07-18 on infoproc

There is only a small probability you will learn something really useful. At this point it is mainly entertainment value.

2011-07-16 on infoproc

Tables 7 and 8 of the supplements have details on fixed and nearly fixed variants.

If there is admixture between groups (at least between Europe and the other two; I assume the gene flow was in both directions), why didn’t the introgressing variants spread further in YRI or CHB+JPT? (They seem to have not spread at all.) One possibility is selective pressure, operating differently on the YRI and CHB+JPT populations.

2011-07-15 on infoproc

“shed 20 lbs of muscle”

Aging will eventually do this to me, although by working out I’ve managed to stay the same weight and (roughly) body composition as when I was 20.

2011-07-15 on infoproc

Congrats! Actually your digicam (or even iphone) probably takes decent movies (perhaps even HD) and it’s more convenient than getting a separate video camera. I find convenience trumps sophistication — if you have the device handy you’ll take more photos/video and by sheer statistics end up with good stuff. More shots at slightly lower quality trumps fewer shots with a fancier device.

2011-07-14 on infoproc

Can’t believe what Auburn/Newton got away with. There’s still a chance they’ll end up like Ohio State, but by then it won’t help the Ducks. Our program probably has its own problems, but nothing like the Cam Newton story.

2011-07-14 on infoproc

Wow, that’s quite a compliment! Thanks 🙂

2011-07-13 on infoproc

You can probably order one if you’re willing to pay postage. I got mine for $10. On the back it says “O Physics” with the big Oregon “O” above “Physics”. You’ll be all set for when we win the national title next year 😉

2011-07-10 on infoproc

Turing was definitely in the paradigm-shifter category… the others perhaps not so much. vN was, of course, probably the smartest guy of that era.

2011-07-10 on infoproc

Zuse also appears in Smiley’s book although I left him out of my post because he wasn’t involved in the US patent dispute. I think Zuse’s machine is Turing-complete on a technicality. The original use of the word “computer” was simply something that computes (it originally referred to people who did calculations!), so it didn’t have to be programmable.

BTW, warning to readers: Smiley’s book is good for its portrayal of the people involved. It’s not so great on technical details.

2011-07-09 on infoproc

There’s an incentive for most companies to seed the internet with fake positive reviews, but that doesn’t mean all or most reviews are fake. SA claims 100k+ paying subscribers, IIRC. This could all be a lie, but if it’s even ballpark correct then most of the discussion on their blog is probably for real.

2011-07-09 on infoproc

As always, the theoretical ideas and the necessary technology have to be in synch in order to have real impact.

2011-07-09 on infoproc

What Turing and vN did was creative (theoretical notion of Turing Machine) or difficult (laying out the general theory/architecture of a vN machine in a very elegant way) enough that if they hadn’t done it there would have been some lag. No more than 5-10 years probably, but still substantial as these things go.

With A, M-E it is already clear from how entangled their efforts were that others could have done the same thing at almost the same time. But the other people who might have done it would have had very similar biographies and cognitive profiles — still drawn from a tiny subset of the human population. It’s that *category* of people that I refer to as Creators, or potential Creators.

2011-07-09 on infoproc

Are you sure it’s fake? My friend’s wife read it and is sure the girl is real. She got sufficiently worked up that she posted a comment on the original thread.

Check out the SeekingArrangements blog and you’ll find thousands of comments that are similar to the “letter” I posted. Are those all fake too?

2011-07-09 on infoproc

You might be mixing up forces and potentials 😉

2011-07-09 on infoproc

The most useful thing for me has been reconnecting with high school friends. But many (more than half?) of my “friends” on FB are people I have never met and don’t know/care much about.

2011-07-08 on infoproc

I suppose guilty pleasure is better than none 😉

2011-07-07 on infoproc

Maybe even your own blog! 😉

2011-07-07 on infoproc

I think it’s gone.

2011-07-06 on infoproc

Definitely not — nobody cares except me! (*sobs*)

2011-07-06 on infoproc

I’m following your lead! No more mister nice guy 🙂

2011-07-05 on infoproc

Gopnik seemed not to grasp that Pinker, with his “small quantitative differences” between groups, was leaving the door open for everything Gopnik is sure just *cannot* be true.

2011-07-05 on infoproc

Looking at the comments on the original letter, I’m sort of convinced the author is real. Below is her response to critics.

Author here. I’ll try to explain my situation in a little more detail, and hopefully it’ll answer a few questions/dispel any assumptions that are being made.

1) I’ve been dating The Lawyer for roughly 6 months. He is the only SD I’ve had that has paid my tuition, and he only paid my tuition for this last semester (which, I might add, was the last semester of my Bachelor’s degree). Every semester before then, I paid my tuition myself, through scholarships, loans, bursaries, and busting my ass at crappy office jobs (regardless of whether or not I had an SD). I do not expect him to pay the tuition for my upcoming Master’s degree. If he offers, great, I can keep what I’m saving. If not, I’m fully prepared to pay my own way.

2) I have always paid my rent myself, and I take care of my own bills. I pay for all my necessities.

3) I would have dated each and every one of my SD’s regardless of whether or not they had money. I did not date these men for presents (although presents are a nice byproduct); I dated them because we had excellent chemistry. Also, like I said, these men are successful for a reason. They’re intelligent, charming, kind people, with great personalities. They’re not walking wallets. However, I will not deny that there are SD relationships that are based entirely on presents/money/etc. Although they’re not my cup of tea, I don’t see anything wrong with those relationships, as long as they involve two consenting adults.

4) I don’t spread my legs because I think I deserve money; I spread my legs because I’m horny and would like to have sex. Neither my SD nor I believe we deserve anything. If we give each other gifts, it’s because we want to give each other gifts. If we want to spend time together, in bed or otherwise, it’s because we want to spend time together. We do not expect anything from each other, aside from respect, honesty, and kindness.

5) My SD doesn’t date women his own age because he prefers to date younger women. I prefer to date older men. We all have our preferences. My SD looks for younger women online because he is somewhat high-profile. Unfortunately, most people think men who prefer younger women are creepy. Since he has a business to run, public opinion is kind of important. By meeting women online, relationships can be kept relatively private.

6) A lot of comments are concerned about the potential for a power imbalance. While I understand why most people would be quick to assume one exists, that issue hasn’t come up (at least, as far as I can tell). We have a relatively healthy, normal relationship.

7) I’m sorry about the environment. If it makes you feel any better, I don’t own a car & I always take public transportation, so maybe it balances out.

8) I like bad rom-coms, so I’m okay with my life being one. I have a soft spot for Kate Hudson, so she can play me in my biopic.Hopefully I’ve been able to answer some questions, but please keep in mind I can only speak for myself and my experiences. There are plenty of other types of SD relationships that aren’t like mine. Just like every other kind of relationship, it’s not a one-size-fits-all kind of deal.

2011-07-05 on infoproc

I’m not saying math ability protects completely against nonsense beliefs. For example, there were many good mathematicians and scientists who believed in communism as an economic system. However, at least a math person knows faulty logic when he or she sees it. And a (good) scientist should have a nose for whether a particular model fits the evidence or not …

2011-07-05 on infoproc

I think mathematical reasoning forces you to be logical and apply rules in a consistent way. If you’re only used to the kind of sloppy verbal reasoning used in some areas it is much easier to bend evidence according to your biases. (Gopnik seems able to read whatever he likes into Darwin’s writing!) Even better is actual scientific training which teaches you how to build models and compare them with reality. The worst is probably legal/humanistic training in *how to make arguments* without any emphasis on actually finding the truth or what constitutes a decent epistemology.

2011-07-04 on infoproc

Yes, unfortunately things get out of hand occasionally. I don’t like having to pre-approve comments because it slows down discussion and it forces me to actually read all the comments, which I don’t always do.

Managing comments is the bane of every blogger 🙁

2011-07-04 on infoproc

Yes. On the other hand you can be HVLM and yet *realize* that there is a type of powerful reasoning, which you might not be good at, but which is nevertheless necessary to reach correct conclusions about complex quantitative issues.

2011-07-01 on infoproc

I can’t comment other than to say that our team is working really hard. This is just like a startup — we have some resources and we’re trying to make something happen. If you’re a high net worth individual reading this please consider making a donation to this project.

2011-06-30 on infoproc

Sorry, I couldn’t resist 🙂

Steven Weinberg: “Against Philosophy” (from “Dreams of a Final Theory”).

Physicists get so much help from subjective and often vague aesthetic judgments that it might be expected that we would be helped also by philosophy, out of which after all our science evolved. Can philosophy give us any guidance toward a final theory?

The value today of philosophy to physics seems to me to be something like the value of early nation-states to their peoples. It is only a small exaggeration to say that, until the introduction of the post office, the chief service of nation-states was to protect their peoples from other nation-states. The insights of philosophers have occasionally benefited physicists, but generally in a negative fashion—by protecting them from the preconceptions of other
philosophers. …

Physicists do of course carry around with them a working philosophy. For most of us, it is a rough-and-ready realism, a belief in the objective reality of the ingredients of our scientific theories. But this has been learned through the experience of scientific research and rarely from the teachings of philosophers.

This is not to deny all value to philosophy, much of which has nothing to do with science. I do not even mean to deny all value to the philosophy of science, which at its best seems to me a pleasing gloss on the history and discoveries of science. But we should not expect it to provide today’s scientists with any useful guidance about how to go about their work or about what they are likely to find.

I should acknowledge that this is understood by many of the philosophers themselves. After surveying three decades of professional writings in the philosophy of science, the philosopher George Gale concludes that “these almost arcane discussions, verging on the scholastic, could have interested only the smallest number of practicing scientists.” Wittgenstein remarked that “nothing seems to me less likely than that a scientist or mathematician who reads me should be seriously influenced in the way he works.”…

2011-06-29 on infoproc

Einstein and fellow Nobelist Otto Stern used to frequent bordellos together. You read it here first.…

2011-06-29 on infoproc

What I find amusing about SA is that it forces reflexively moralistic people to think through the whole continuum between traditional marriage and $150/30min. Most of them probably did not understand beforehand that it was a continuum and how hard it is to draw the dividing line.

2011-06-29 on infoproc

Hey man, it’s art.

2011-06-29 on infoproc

Neither does Professor Flory 😉

2011-06-28 on infoproc

I think they would have gone down like Lehman without government intervention. But I do think they are better run than the other shops. This example of knowing their instantaneous firm-wide risk is just one example.

2011-06-26 on infoproc

Clyde Drexler is 6″7 but (reportedly) ran a 4.5 and benched well over 300 when coming out of college. Michael Jordan would probably also have made a good decathlete and he is around 6″6. One problem I have with “World’s Greatest Athlete” is that they are drawing from a small pool in the decathlon, as there is little monetary incentive.

I do think the sweet spot for the decathlon might be just over 6 feet, at least for now. Before Usain Bolt no one would have believed a 6″5 guy could be the sprint king …

BTW, they list Eaton as 6″1 but he seems quite a bit taller to me. He might be 6″2 or 6″3 unless I was on drugs that day at the gym.

2011-06-25 on infoproc

There is already a clear cut example indicating differences in “response style”. E. Asians’ self-ratings on academic (e.g., math) ability are systematically low relative to actual performance, whereas other groups systematically overrate their own abilities.

2011-06-25 on infoproc

The problem with personality measures is that they are very noisy and the self-report inventories are easy to game. I have seen some papers showing consistency between self-report evaluation and third party (co-workers, friends, family) evaluation, but nevertheless the signal to noise ratio will be much weaker for personality than for g. If you try to do things cross-culturally (comparing two different countries, for instance) there are further interpretational and linguistic difficulties (do the subjects interpret the questions the same way?).

It seems clear to me that there are group differences in personality distribution, but even this is hard to establish quantitatively (for example, some of the results from the meta analysis you quote are counterintuitive). Whether group differences are mostly due to culture as opposed to genetics is even less well understood.

2011-06-22 on infoproc

I switched topics slightly with the Brooks talk. He isn’t only concerned with empathy — I just thought the WS/politico observation was funny. WS people *want* to understand what others are thinking (they stare into your eyes) but can’t quite manage it. Politicians are actually good at reading other people.

2011-06-19 on infoproc

PS You can ask RKU if you like, as IIRC he won the Westinghouse his year with a paper about black holes. On our BGI CGU team we have an IMO winner who was a Westinghouse finalist.

2011-06-19 on infoproc

re: IMO performance, I think high g (or, more precisely, innate math ability) is necessary but not sufficient. You need training as well. Nevertheless, anyone who succeeds at a high level in these competitions is very, very smart. On average the IMO guys will be much smarter than the Intel finalists. Trust me, I know a fair number of people from both categories. Usually the smartest Intel/Westinghouse person did a project in number theory or perhaps theoretical physics or CS and also was pretty good (if not at the very top) in IMO competitions.

2011-06-19 on infoproc

Thanks for the link.

Don’t people usually claim a higher g loading for Ravens than what is given in your table for the matrix reasoning subsection (.66)? On a related point, it seems that easy matrix problems may load more on a visual factor (“gestalt”) as opposed to harder problems that require reasoning and logic (“analysis”). I wonder if the WISC matrix questions aren’t mostly of the easy type.

See discussion…

From your table, the block assembly test loads .66 on g (I didn’t realize it was so high) and .47 on Gv (visual subfactor) whereas matrix reasoning loads .66 on g (seems low relative to RPM?) and only .27 on Gv. When people say “spatial ability” I usually think of block design or rotations, not Ravens. The correlation between matrix reasoning and block design is not given, but I would expect (given the loadings on Gv) that it is not that high. (Let me know if I am not reading the table correctly!)

I was recently talking to someone about the standardization of the latest WISC versions in China. When the results are out it would be interesting to see how the various sub-correlations compare to the Western data.

2011-06-19 on infoproc

Ben: I should have said g is approximately the same for these populations. Inevitably there will be small differences in the exact largest eigenvector, but of course these things are only defined modulo statistical noise, etc. You get approximately the same largest eigenvector, at least that is what I am told — I have never looked carefully at this myself.

“PCA categorizes as smartest those who are smart in the most common way.”  This is correct. To me it’s just a convenient way to compress the data.

2011-06-18 on infoproc

I’ve been told by psychometricians that there is no evidence that the principal component deduced from factor analysis over a battery of tests varies between populations. That is, g is the same for all populations for which there is extensive data, including in particular European and E. Asian populations.

Also you are confused in saying that Ravens is spatially loaded. Recognition of how a 2D pattern evolves in “time” (what you call matrix reasoning) is not the same as spatial ability. The former depends mainly on pure logic whereas the latter involves, e.g., visualization of spatial relationships.

2011-06-17 on infoproc

Thank goodness for that. Beats having to check the paternity 🙂

2011-06-17 on infoproc

High V, Low M.

I doubt he really understood (before communicating with people like you) that there are very good arguments against random walks producing functioning eyes and brains.

2011-06-16 on infoproc

You are right that Krugman is not very mathematical as economists go, but he understands enough (the kind of math used to model complex systems like economies or populations) to realize that Gould uses very little math in, e.g., his work on punctuated equilibrium.

Contrast this to what real evolutionary theorists (as opposed to snail collectors) work on. For example:… . I flipped through Gould’s 2002 book The Structure of Evolutionary Theory (“magisterial” say the high V low M admirers 🙂 and don’t recall seeing many, if any, equations! It certainly makes one suspicious…

Why do you not dispute Maynard Smith’s characterization of Gould?…

“… In contrast, the evolutionary biologists with whom I have discussed his work tend to see him as a man whose ideas are so confused as to be hardly worth bothering with …”

Yes, Galton had a nervous breakdown preparing for the Tripos. Performing well on the Tripos (becoming a wrangler) is a much higher hurdle than what Gould needed to do to understand the work of real evolutionary theorists (i.e., population geneticists) like Smith or Fisher or Haldane. Real questions in evolutionary theory (kin selection, for example) simply cannot be addressed without some math.

2011-06-16 on infoproc

This is ancient history, but theorists leaving the field in the 1980s or even 1990s had far fewer opportunities. It was basically finance, defense or fairly boring tech companies. Whole sectors that hire the same type of person today didn’t exist then.

2011-06-16 on infoproc

> … mathematical talent is something new. Many of the high M low V would have been judged not so smart a thousand years ago. < Yes, I agree. Part of my point is that, among the hoi polloi, "smart" is often judged on V, whereas among the cognoscenti the greatest distinctions have to do with M. The Scientific Method is also new. A stronger statement than what I wrote in the post is that the vast majority of new conceptual space humans have been exploring in the last 200 years is off limits to people below some cutoff in math ability. This conceptual space is not only interesting, it is *useful* (See, e.g., Moore's Law, crypto, radar, atomic weapons, bioinformatics, ... 🙂 I would almost assert that you cannot call yourself a serious thinker or intellectual these days without some math chops -- most of the serious and important stuff under discussion would be beyond your ken. PS Galton, although no wrangler, was the first psychometrician. Pearson, who was stronger, was his protege. Both of these men far outdistance Gould in intellectual ability.…

2011-06-15 on infoproc

Congrats on your move to Google! I think you are a good example for the point made in my post, since someone with your background in theoretical physics would have had much different career opportunities 20 years ago.

Re: software vs hardware, you can find point cases where software breakthroughs give you a big speedup, but the hardware advances over the last 30 years can be applied to *any* computational problem. I’d rather live in a world with 80’s software and 2011 hardware than vice versa 🙂 For example, I could still do command line unix stuff, public key encryption and email just fine…

2011-06-14 on infoproc

I interpret that as when we finally “solve” computer vision so that computers can process visual information as well as humans it will turn out that rather than some super algorithm providing the capability it will turn out to be an amalgam of hacks. That’s actually how evolution worked in all likelihood — hacks on hacks on hacks = intelligence. There will not turn out to be a grand unified theory of intelligence.

2011-06-14 on infoproc

Software and internet geeks get most of the attention because of the low barrier to entry and the recent explosion of opportunities. But the real engine powering all of this is applied physicists who keep Moore’s Law going. Software has improved by factors of 2 or (maybe) 10, but we got a factor of a million in processing power, memory and bandwidth thanks to semiconductors over the last 30 years or so.

I think high barrier to entry industries are much less attractive to ambitious young people. Biotech is kind of an exception because although drug discovery is expensive and slow, financing mechanisms exist to support innovation. Biotech companies can IPO on just an idea, which isn’t (anymore) the case in internet and software.

2011-06-13 on infoproc

Both, of course 🙂

2011-06-10 on infoproc

See pictures here or in my data mining paper to get an idea of what correlation of .37 looks like. For social science it’s considered a relatively strong correlation.…

2011-06-09 on infoproc

> I was president of the Prometheus Society (IQ 164+) in the 1980’s < Please tell us more about your experiences as a Promethean! Was Grady Towers one of your members? I've read a few of his essays on the web. Do you agree with his perspective in The Outsiders?

2011-06-07 on infoproc

Actually what is interesting about this experiment is that it allows you to break up the task into sub-ECTs: 1. recognition of number (can be made easier or harder using contrast on the screen) 2. evaluation of whether it is > or < 45 (this turns out to be the most g loaded component) and 3. motor execution. You can also use multitasking versions to test which of these can be done in parallel or require serial processing. It seems that 2. can only be done in serial and, in some sense, requires the "full attention" of some part of your brain/mind. For related stuff, see, e.g., Dehhane's Numerical Cognition.…

PS When I write in the goofy voice it’s not supposed to be me. I am mimicking someone else 🙂

2011-06-06 on infoproc

But, but, I thought BCG only hired the best and brightest? 😉

I seem to recall back in the day they were just a tad pickier (although smaller) than McKinsey……

2011-06-04 on infoproc

It’s now just called BGI (formerly known as Beijing Genomics Institute). The Shenzhen govt lured them here with grant money, land, etc. Now other cities are trying to lure them away from Shenzhen.

2011-06-03 on infoproc

Not bad if it’s a HS kid playing with synthetic biology.…

2011-06-02 on infoproc

The sequencing was done here, as far as I know.

2011-05-30 on infoproc

The casinos I visited were very crowded on Sat/Sun. I think the majority of people were mainlanders — more putonghua than guandonghua.

2011-05-30 on infoproc

I was mainly referring to architecture. For example at the grand lisboa the main gaming area is three huge floors with a giant open space in the middle connecting the levels. Hard to describe — kind of like an ornate, glitzy indoor football arena packed full of punters 🙂 I got in trouble for taking a picture and was forced to delete the file from my phone. It’s understandable that they are very paranoid about people taking pictures of their casino operations, but they wouldn’t even let me shoot the architecture.

I know some guys (via my In-Q-Tel connections) who work on casino security technology and it’s quite interesting what they can do. There are security cameras everywhere in these places.

2011-05-27 on infoproc

So even though it was a banner year for MIT, Caltech still beat them in the team competition and outperformed on a per-capita basis. It looks like among the top 20 scorers 7 were from MIT and 4 were from Caltech. Adjusting for the sizes of the schools we get something roughly consistent with the historical result that Caltech has 2x as many top performers (e.g., Putnam Fellows) per capita as MIT. The school that did worse than their historical trend is Harvard, and this year might be the first year ever for Stanford to have a Putnam Fellow.

2011-05-25 on infoproc

I don’t deny there should be adjustments. There are biology majors at Caltech as well, though 🙂 I think the biggest factor is that MIT’s math department is better than Caltech’s, which really biases the Putnam numbers, but in the opposite direction than the effects you pointed out.

2011-05-23 on infoproc

I talked to the scholarships people at Caltech during my last visit. I was shocked to learn about the Axlines and complained about it! I told them they should use the Hsu Scholarship to recruit super bright applicants… unfortunately it can only fund 1 per year.

The historical outperformance of Caltech over MIT was *not* due to merit scholarships like the Axline. They only created those long after I graduated.

PS Competitions like USAMO/Intel/… aren’t 100% perfect filters, and they don’t really access the full population base. There are a decent number of kids as talented as the winners of these who escape the radar. I would guess >3x at least across the country, mostly from less education-obsessed places. (Back in the old days I would have said 10x because most HS students outside the northeast and a few other places never heard of the USAMO or Westinghouse competition.) Just having a high admission threshold allows Caltech to populate its tail with such people, even if they are losing out on the big name contest winners.

I’ve been following these things for some time now. It’s clear to me that no filter of 17 year olds (or even 22 year olds) is going to be anywhere near 100% efficient. If you read the AMS report on the Putnam, you’ll see that while it identified many top math researchers as college students, there was an equally large group who never took the Putnam, or didn’t do very well on it. All of these tests are just noisy predictors.

2011-05-23 on infoproc

Not sure why you think the top 1% at MIT (10 kids per class?) are better than the top 1% at Caltech (2 kids?). For example, the current Caltech senior class has a 3x Putnam Fellow. Are there 5 3x Putnam Fellows at MIT in the senior class? 😉

Looking historically, I think MIT has 2x the total number of Putnam Fellows as Caltech. But MIT is 5x bigger and so as I pointed out already the population density at Caltech is twice as high (the Nobel count is similar, I think). It is hard to square this with your 1% claim, but the current situation may be different from the historical one.

2011-05-22 on infoproc

This is going to be resolved by the actual studies. If the standard models used successfully in applied population genetics (e.g., to breed better crops, livestock, etc.) are wrong for human height or IQ, we will find out from adequately powered GWAS.

2011-05-22 on infoproc

As you know from my other comments I don’t think g loading has a precise technical meaning (in this way I agree with Schoneman). I meant it in the way you interpreted it — the smartest people, the hardest, etc. Because of the job situation in theoretical physics and math, the *majority* of people trained in these areas are/were forced into other careers, despite being the creme de la creme by any quantitative metric. If you follow their track records they’ve done nontrivial stuff in many other fields, including biology, computer science, defense research, finance, startups, etc. The reverse has almost never happened. So it’s pretty clear which things are easy or hard even for very smart people.

PS On Schoneman, psychometrics has a vocal minority of what I call “methodological scolds” that serve a necessary purpose, but they are like people who would tell engineers that they shouldn’t build bridges or cars until all of solid dynamics can be derived from first principles of subatomic physics. As a physicist I know a crude tool when I see one: I understand the limitations of psychometrics, but am willing to overlook them in favor of practical utility (e.g., predictive power).

You can easily see where Schoneman’s biases lie. For example, he always quotes the lowest validity scores for SAT as a predictor of college grades. More careful modern analyses, which control for course selection and restriction of range, obtain much higher validities. Why doesn’t Schoneman, someone whose major interest is psychometrics, know about this?

2011-05-22 on infoproc

It might, although you’d need a way to quantify research record and evaluation letters, which are a big component of strength of application. Some applicants have already published research whereas others haven’t done much other than take classes.

What is sad is that there is little interest in trying to understand the “validity” of factors (GRE general, GRE physics, UG GPA, research record, letters, etc.) used in the current process.

2011-05-21 on infoproc

Some of each. The descendants of KMT immigrants (1949ers) are probably more Type A than the easygoing longtime tropical islanders.

PS I am ignoring the possibility that that was an attempt at a joke 🙂

2011-05-21 on infoproc

Still, it looks like population density of Putnam Fellows is 2x higher at Caltech, even though (sadly) Caltech’s math department lags some its other programs.

2011-05-20 on infoproc

I meant the E. Asian numbers are realistic. 750V for a theory applicant at a top department is not rare, but in the overall physics applicant pool would be. Did you look at the Roe scores for theory vs expt?

2011-05-20 on infoproc

I should clarify what I meant by “as a whole”, which was not that *every* area was in a bubble, but that many/most places (including ones far from the coasts) were, and that the implied value of total US housing stock was significantly inflated.

But my memory of what exactly Krugman wrote is fuzzy.

2011-05-20 on infoproc

I serve on admissions committees… the numbers I gave you were not made up. BTW, just to be clear by E. Asian I meant foreign nationals applying from abroad.

2011-05-20 on infoproc

I seem to recall at some point circa 2006 or so Krugman claimed that the US as a whole was not in a housing bubble, even if certain cities were. This of course was totally wrong.

2011-05-20 on infoproc

Well, higher Q abilities don’t move the GRE average because the ceiling is so low.

Just imagine taking the equivalent of GRE V but in *Chinese*. Even if you had worked diligently on it for the last 5 years, do you think you could score as high as the *average* college graduate in China? I certainly could not.

When a physics admissions committee looks at a E Asian applicant a GRE V score which is just average (say 500 out of 800) would be considered quite good — they’d look more at the TOEFL score, and even top TOEFL scorers might only score 500-600 on the GRE V. So, a top applicant from E Asia might have Q/V score of 800+500 (the “top” characterization would be due to other things like their course grades, letters, research accomplishments), which certainly lowers to average for the physics department relative to, say, philosophy. Comparable US applicants might be something like 800+750. Now average 1/3 of the former and 2/3 of the latter to get something like 1400 (the number you are looking at) when in fact you are talking about 1550 caliber people (once you adjust for depressed V average).

2011-05-20 on infoproc

You have an investment banker? 🙂

2011-05-20 on infoproc

I agree with your comments although I wonder how Jensen could make that claim about *all* past geniuses. If he thinks von Neumann or Gauss were only 160 I would probably have to disagree…

SMPY showed a lot of regression, but of course they were testing 12 year olds, many of whom had prepped for a widely studied test (SAT). When I was an undergrad I was paid to tutor a little kid who was trying to get into SMPY.

2011-05-20 on infoproc

These are all just rough statements. We cannot really measure these things to much better than 1 SD accuracy.

Note the Fermi estimate guy doesn’t think there is anything like complete overlap between his population of good scientists and the population of +4SD g scorers.

Note also that there are 30x more people at 3 than at 4. So there is a good chance that among the 3s someone will get lucky, or have some non-g type advantage (networking, creativity, lab dexterity, etc.) that gets them ahead of many of the 4s. In the Fermi calculation the correspondent only estimated that a fraction of the people even at top departments are in his “nontrivial brain” category (i.e., 50-100 people total at a top tier university).

Isn’t +3.6SD at the ceiling of the GRE?

2011-05-20 on infoproc

There was an interesting old psychometrics study done by ETS that showed that taking a math intensive major had very little effect in raising GRE M relative to SAT M (taken in high school). GRE M is so easy that taking courses like real analysis doesn’t raise the scores of people who can actually pass real analysis!

BTW, you didn’t address the effect on the V score average from having so many physics grad students from foreign countries (esp. E. Asia).

Too bad Roe didn’t test some analytic philosophers for you 🙁

PS How many philosophy grad students are there? Big physics departments can have large entering classes. In my cohort at Berkeley there were 40 kids, over 2/3 of whom wanted to do theory. But only 5-7 or so were able to.

2011-05-19 on infoproc

Seems implausible to me.

1. Perhaps 30-50% of physics grad students (depending on department) are foreign born. Their V scores are not a good indicator of their actual verbal ability.

2. The GRE M is just way too easy and while it might be a reasonable test for philosophers’ math abilities it is laughable for physics/math grad students. I would guess almost all good physics/math grad students are above the GRE M ceiling, whereas a smaller fraction of philosophy students are.

3. You need to break out the theorists, who are only (roughly) 20% of the physics population. (See Roe study for the difference in psychometric profiles.) Experimentalists often have other abilities (tinkering, building devices, running teams) which neither theorists nor philosophers tend to have. If you want to claim that philosophers have g similar to experimental physicists it might be true (I don’t think g really does justice to the lab abilities I listed above), but theorists are another matter.

Finally, I know (or have known) lots of philosophers as well as physics/math people and your hypothesis doesn’t fit with my experience. Have we already had this conversation before?

PS Re: humanities in general, my wife’s PhD is in Comp Lit (from the top department in that field) which was at the time the glamour area in the humanities (may still be, I don’t know). Let’s just say I wouldn’t compare people in that field to the top brains in math/physics.

2011-05-18 on infoproc

What is your startup? It’s unusual that you get to work with ideas. In my experience it’s mostly execution after a while and not that conceptual.

2011-05-18 on infoproc

They are stressful in different ways. Entrepreneurs deal with a lot of external randomness, so/but if you don’t win it isn’t necessarily a judgement about your intrinsic worth. On the other hand, there are few challenges as pure as sitting in an office with a blank whiteboard and trying to discover/invent something new in theoretical physics. If you fail you can’t blame anyone else.

So, startup environment is more intrinsically stressful, but if you fail as a researcher you can’t blame anyone but yourself.

2011-05-18 on infoproc

I think people who have tenure have a *responsibility* to take on high risk research that others cannot. It’s amazing how risk averse even most tenured researchers are, just trying to keep the grant money flowing.

Sometimes I feel like physics is way too hard and I’m a moron for not staying in the startup world 🙂

2011-05-18 on infoproc

g is not the same as SATV or SATM or MO (math olympiad) ability. They are all distinct, have some dependence on training, etc.

There is probably typically a 1 SD fluctuation in scores for a particular individual on different kinds of “intelligence” tests. This fluctuation may go up a lot at the high end, and if we include tests that load on training (MO).

The main issue is that different tests have different “validities” as predictors for various purposes.

Keep in mind I’m not a real “g man” — I think of all these tests as rough predictors only. It’s not clear exactly what they measure.

I do think that someone who scores +(3-4) on any of these kinds of things has a brain which is qualitatively different from that of an average person. That’s the only crude assumption I need to make in order to be willing to invest time in an intelligence GWAS.

BTW, your friend might be +3 in g and +4 in MO ability, so consistent with my guess.

2011-05-17 on infoproc

You are asking about ancient history!

I was one of the top kids in Iowa but not highly ranked nationally. It wasn’t a big deal back then. The top scorers tended to be from places like Bronx/Stuy and it was clear they were getting training that we weren’t. There was no number theory, tricky geometry, combinatorics, etc. in what we were taught in HS. We had no teachers who could do the problems and no coaching. We just went in and took the test… (In that era there was just one HS exam and that got you invited to the USAMO training camp. In my senior year they added the AIME, which I qualified for, but I didn’t score high enough – top 50? – to be invited to the camp.)

At Caltech I attended some Putnam training sessions but never actually took the test as by that time I was more interested in physics than math. My freshman year I actually planned to take it but I got food poisoning the night before. I suppose if I had practiced I might have gotten an honorable mention (top 40) as some of my friends, who I thought were roughly at my ability level, did so. But I doubt I could have been near the top. It was obvious to me that people who had trained for these things in HS and earlier still had a big advantage later on the Putnam.

At Caltech and in my later physics career I have known people who were at or near the top in these competitions and I generally find them to be quite impressive. I’ve actually written a physics paper with a 3x Putnam Fellow so I think I’ve seen the whole range of capabilities 😉 One person I know who was an IMO gold medalist (US team) told me he found his “problemist” (my terminology) ability only gave him a +1SD advantage on the Caltech curriculum, which is pretty hard stuff (if you search on the blog you can see what econ Nobelist Vernon Smith had to say about it). This is roughly consistent with my experience (although perhaps I would have guessed he would be at +1.5 to 2 SD at Caltech) because I often did as well or better in courses (at about the same level of effort) as kids who had either been at the top in US HS competition or did well on the Putnam.

In retrospect I notice I was much more interested in learning stuff than in solving tricky problems. In HS I had accelerated quite a bit (taking a lot of college math and physics courses) and kept doing so in college. At 19 (Caltech senior) I was taking grad level courses in quantum field theory and general relativity, which didn’t seem like a big deal at the time but looking back I see there aren’t a lot of kids doing that. I was mainly interested in learning the things I really found interesting: quantum mechanics, general relativity, quantum field theory — the stuff I actually work on now. I don’t know of anything as g loaded as this material, with the exception of pure math, which just doesn’t happen to get me as excited.

In Eugene (where I live now) there is a UO math PhD who teaches at a local middle school and coaches kids for these competitions. I am pretty sure the talent pool of kids where I grew up is as strong as the one in Eugene (actually stronger because ISU had a good engineering school whereas UO is a liberal arts school), but Ames never produced a national-level competitor during my time there, whereas Eugene has produced several including an IMO gold medalist in a similar period. So it’s very clear to me training matters.

Nevertheless I think anyone who can get to the national or international level in these competitions is clearly very smart, and I would tend to think these competitions are a better filter at the high end than “real” IQ tests.

2011-05-16 on infoproc

Go to Whole Foods and buy lunch for your kids in the deli section. It will cost you $5-7 bucks per kid easily. Ditto for dinner. So a busy 2 earner family could easily spend $10 per day + occasional real restaurant meals.

There is huge variation in how much can be spent on food, ranging from white bread, PB and cereal (sugar, fat and carbs) to expensive organic food. I’d say $2.5k per capita is very low for most upper middle class families.

2011-05-16 on infoproc

I know many people much smarter than I am  🙁

2011-05-16 on infoproc

Re: math and theor. physics phds I wasn’t referring to any research paper, just to SAT/GRE scores used in the admissions process.

But you are right there is not much on the high end. SMPY and Roe are the two best sources I have been able to find with relatively high ceilings.

A completely open question is whether in the 3-4SD range the old SAT or “real” IQ tests like WAIS or even the IMO/Putnam have more actual validity (e.g., for predicting success as a scientist or professor).

2011-05-16 on infoproc

I think the average dog is at +4SD for happiness relative to people. But I’m not sure I’d get much done in that state 🙂

Given the amount of human suffering in the world +4SD in empathy would be very painful for me. I’d probably go nuts.

2011-05-16 on infoproc

But what about 3 vs 4 SD? Standardized tests like (old) SAT and GRE had ceilings above 3 and I bet a detailed study would show some validity (see also SMPY). The avg for admits to top math and physics (theory) PhD programs is above 3, based on my experience. 3 on old SAT was only a little above Caltech avg.

2011-05-16 on infoproc

HNW yes, UHNW no 🙁

2011-05-15 on infoproc

Sorry if I mistook you for someone else. There is a commenter here who often changes his pseudonym. He went by anon for a long time.

2011-05-15 on infoproc

Can almost anyone learn multivariable calculus if it is presented correctly? I wish someone would show me how to present it properly 🙂

2011-05-15 on infoproc

What do you think a more reasonable estimate would give — that the family is $20-30k in the black rather than in the red at the end of the year? Does that mean a $200k per year household is “scraping by”?

2011-05-15 on infoproc

I am interpreting here, but I think that “capable” means that given the opportunity they could make the breakthrough. They have the necessary brainpower, but of course there’s luck involved to get the opportunity in the first place. In the original calculation many of these people (e.g., 5 people out of a 100-200 ranked physics department) will not have made earth-shaking breakthroughs.

I can definitely think of a dozen people in our class who seem to be as smart as researchers who actually made breakthroughs. Whether they actually will or have done so is another question. Probably half of this dozen are not even in sci/tech anymore, they are, um, allocating resources more efficiently for society’s benefit 😉 Some of them might even be in the 4SD wealth category! Imagine that — a 4×4 double winner 🙂

2011-05-15 on infoproc

Quick summary: Harvard dominates, MIT coming on strong in recent years, Caltech punching far above its size 🙂

“By a wide margin, Harvard has the best record in the Putnam competition. Through 2010, Harvard has won the team competition twenty-seven times, while its closest rival for team titles, Caltech, has won the team title ten times. MIT is in third place with six titles with three of these coming since 2003.” (i.e., before 2003 the little tiny institute completely dominated the big mediocre one 😉…

2011-05-15 on infoproc

Look up my blog post on Roe’s study of US scientists. It addresses many of these issues.

2011-05-15 on infoproc

Caltech avg is higher than MIT’s and if SD is similar then advantage in the tail goes to the former. That means the number of Putnam fellows or Nobel winners will be higher per capita at Caltech, which I believe is the case. Btw, Caltech is 5x smaller than MIT but has won the Putnam team competition more times (2x?) than MIT! IIRC, only Harvard has more team wins than Caltech – pretty amazing for such a tiny school.

2011-05-15 on infoproc

You’re talking about the situation right now, whereas I am kind of averaging over the last 20 years or so. There were times when Harvard or Caltech was better than MIT on the Putnam. Keep in mind training matters a lot for math competitions, and if for some reason a disproportionate fraction of kids who invested energy in AoPS in HS decide to go to MIT (or some other school), that would distort our perception of the real talent distribution.

Maybe +20 years from now the story will be different.

I should mention that the National Merit Scholar numbers I put on the blog some time ago seem consistent with the 4SD numbers I gave in this thread.

2011-05-15 on infoproc

Hello anon. You can stay if you can behave.

The dominant pca vector that is g does not vary significantly between populations, at least according to the limited data available (fairly big studies for western countries and Japan). Once we fix this vector then g is a property of the individual.

You can read the GIANT papers if you want to know more about the population under study, but basically we’re talking about affluent europeans.

2011-05-15 on infoproc

So, are the incentives for producing these public goods optimal for max rate of progress?

2011-05-14 on infoproc

My rough guesses: +4 SD (1 in 30k) is about 100-200 freshmen each year in the US (assume some fatness in the tail; exceptional immigrants from abroad). Let’s say 200. I think that’s about top 5-10 percent at Caltech (avg IQ 140-5, SD around 10) or about 10-15 kids per class. The top 10 or 15 UG schools (private, all with class sizes in the 1-2k range, except for Caltech) account for about 100-150. YPS and MIT probably account for 10-25 each; H has the most, maybe 35. The remaining 50-100 are at public universities, with Berkeley having the most (maybe 5-10 per year), or at other lesser ranked privates.

A >+5 SD (not kidding) correspondent supplies this joke:

The college faculty were in their annual meeting when, suddenly, an angel
appears. Turning to the Dean, the creature said, “I will grant you one
of three boons — infinite wisdom, infinite wealth or infinite health.”
The Dean thought for a minute, then replied “Wisdom.”

“So be it.” and the angel disappeared.

In the silence that followed, the Dean sat thoughtfully, saying nothing and
staring off into the distance. Finally, one of the other faculty members
exclaimed, “Do you have anything to say? What words of wisdom can you
provide us?”.

Quoth the Dean, “I should have taken the money.”

2011-05-14 on infoproc

Thanks for the links! I have to say I think I have more game (or did, back in the day) than this JT guy 🙂

2011-05-13 on infoproc

Yes, but these are non-additive effects and the easiest thing to detect first will be the loci that contribute to the .6 or so narrow sense heritability.

2011-05-13 on infoproc

NSE effects are basically the observation that “random stuff happens” and influences our development. It was shocking to twin/adoption researchers that most of the E effects were random and not correlated with simple observables like SES, parenting style, etc.

Some people (e.g., Bryan Caplan) take this to an extreme and conclude that parenting choices (within a reasonable range) make little or no difference to how your kid will turn out. When applied specifically to adult g I think this is plausible.

Most smart people are surprised at how large the estimates are for regression: why won’t my kid be as smart or smarter than me? I chose my wife carefully… 🙂 But they fail to understand that they probably benefited from luck in both genetic and environmental factors, and they can’t pass on the NSE luck to their kids! (Although some of us are trying … see “fine tuning n-vectors post”:… )

2011-05-12 on infoproc

I recommend Dale Carnegie’s How To Win Friends And Influence People. (Search on blog for an outline.) Public speaking can be improved dramatically through practice (anyone who has taught physics and given lots of seminars will have no problem giving a presentation to VCs or board members). I’m too old to have experienced PUA training as described in the article, but it might work.

2011-05-12 on infoproc

If you search on my blog you can find more detailed discussion of this. Regression is due to the .2 or so of heritability that is non-additive, and also due to the fact that the non-shared environments (which are effectively random) experienced by parents and children are uncorrelated. Note that the shared environments (e.g., SES) might be correlated between parent and child, but not the non-shared environments.

The twin/adoption data suggest that .6 of the .8 heritability is additive. So we assume linearity as a first approximation in our analysis. Working out the nonlinear (gene-gene interaction) details will be very difficult and take a long time, whereas much of the .6 additive stuff can probably be pinned down in the coming decade or two, assuming we can get the phenotype data.

2011-05-12 on infoproc

My experiences are atypical because I don’t really fit the Asian personality type that well.

Physics is a special sub-field because people are generally pretty smart and questions have actual right and wrong answers, so BS doesn’t fly very well. Nevertheless, (non-US raised) Asians are at a disadvantage because of language ability and also the tendency to be less aggressive. Other areas of academia have exponentially more BS with predictable consequences for quiet, straight laced Confucian types 😉

Silicon Valley (the only place I have real private sector experience) is one of the most color-blind places on the planet. There are exceptional and successful Indians, Chinese, etc. and often the smartest guy in the room is a brown guy. Nevertheless, there is a certain type that is considered suitable for a CEO or senior exec. position, and most E. Asians don’t have the right personality. S. Asians are more likely to have the right personality, although even in that population it’s a bit less common than in the euro population. It is interesting to me that the founders/CEOs that succeed in China tend to be quite different.

Note that expectations become reality. If everyone in the US *thinks* that, e.g., a short or quiet person just can’t be an effective leader, then those people will not get the positions, or won’t inspire confidence even if they do. In some other culture the very same people might make great leaders.

2011-05-10 on infoproc

Yes, if you listen to Weinberg he is very critical of the old eugenicists, but he doesn’t seem to realize that you can glean a lot about breeding values (e.g., of plants or livestock) without a complete understanding of DNA or the micro-mechanisms of genetics.

2011-05-10 on infoproc

Hi Henry,

You can certainly use the figure in your class.

The red line is the upper bound at 95% confidence level on the probability of getting an A in a particular quarter of introductory physics, and the blue line is a similar upper bound on the probability of earning a cumulative GPA of at least 3.5 or so. You can see the blue line changes nonlinearly (seems to have a threshold) at around 90th percentile ability on the math SAT. Note these are *upper bounds* on probabilities, obtained using a simple model described in the paper. If you look at raw data, the prob. of earning a GPA > 3.5 in physics is only about 50% even for kids with SATM around 800 (if I recall correctly). GPA of 3.5 (i.e., equal number of A’s and B’s) is kind of the minimal threshold for “mastery”, or, more practically, to be admitted to a decent graduate program.

It is easy to see that over the 10 years or so that we studied, if one had excluded all the kids who scored below 90th or 95th percentile on SATM from the physics program, one would have only lost a couple of kids capable of a PhD. Of course these results imply severe “disparate impact”!

The whole paper is here:…

2011-05-10 on infoproc

“Rather than rely on the scientifically unsupported claim that we are all equal, it would be better to emphasize that we all have inalienable human rights regardless of our abilities or genetic makeup.”…

2011-05-10 on infoproc

No, it’s just a weird consequence of latex on OS X. Something to do with the way PDF figures are handled, I think. Try using another PDF viewer like Google Docs.

2011-05-10 on infoproc

I never get that kind of reaction from physicists or mathematicians. I wonder why? 😉

2011-05-07 on infoproc

I saw the earlier comment you deleted, but I’m not the one who ‘liked’ it 🙂

2011-05-06 on infoproc

I laid out the up side in the talk. Humans should take control of their own evolution eventually, and this is a crucial step.

Cognitive genomics can hardly be compared to nuclear weapons technology which easily could have (and possibly still will) lead to the extinction of the human race. Yet many (most) physicists were OK doing that kind of work.

2011-05-05 on infoproc

Sometimes I think buddhism is the deepest of belief systems, but other times I think detachment will keep me from accomplishing things in this life.

2011-05-05 on infoproc

There are plenty of ex-military guys who can think analytically and plan ahead. They might not be quite as willing to think “out of the box”, but among special forces guys I bet it’s more common.

2011-05-05 on infoproc

I am not particularly good at either 🙂 I was more of a grappler than a striker, and as a chess player I am very prone to blunders (maybe in life as well).

2011-05-05 on infoproc

When I was doing MMA I would still do a very quick HIT workout (20-30 min), usually before training MMA. I liked to be tired already in the dojo so I would rely more on technique than on physical abilities. During training we would split between technique drills (usually to warm up) and sparring. I didn’t have to do much aerobic stuff since sparring really takes it out of you, but occasionally I would still run so my body wouldn’t forget how to do that. Nowadays my workouts are closer to the warmups of the past 🙁

2011-05-05 on infoproc

Do I have to plead guilty to this?

Actually SEALs aren’t that large on average. Big beefy guys might not have the stamina it takes. If you think about it, a small durable guy is better for modern combat operations than a typical football player. I am 100% sure I could smoke Dave in a 40 yd dash but on a 5k run he would probably beat me.

2011-05-05 on infoproc

“… we assume that each player is rational from some perspective”

If that includes crashing a 747 into a skyscraper because of something you believe about the afterlife, then I suggest it’s not much of a constraint on decisions / actions in this life.

2011-05-05 on infoproc

@Dave and Vic

Of course, the whole thing hinges on whether one actually has a good definition of rationality (or utility). The early researchers in decision and game theory thought they might actually solve this problem but it’s obvious they cannot. Is the goal (normative) to define what a “rational” person “should do” in a particular situation? Or is it to predict what actual people do? For the latter, psychology is as or more important than economics and all the pretty math will turn out to be of limited use.

2011-05-05 on infoproc

Great link! I think I had read that before. I especially liked:

“… That quitting conversation will show up a hundred different ways. You can say, “This thing is bullshit. This thing is crazy. I can’t believe I’m doing this.” Or you can create things like, “Well, that person’s just trying to hold me back,” or “I’m just not good enough.” Whatever… I’m telling you, quitting sounds very reasonable. Literally, it’s a very simple process of getting on target and moving forward vs. the conversation that leads to quitting. And those conversations show up every day for people. And really, that’s the battle. That’s the war. If there’s a war going on, really the war is within yourself and one you have to confront day after day.

The majority of people who get selected to SEAL training will quit, drop out or simply go away. You’ve got to be able to generate within yourself the mind frame that you must always be going forward. And you’ve got to want the thing bad enough to be willing to do anything to get there, regardless of and in spite of all the obstacles, in spite of all the hurdles, in spite of all the doubts that get in…the stress and the pain, you’ve got to keep going forward.”

2011-05-04 on infoproc

I haven’t had time to train much since my kids were born, but I used to be pretty serious about it. I started back in the day, around the time of the first UFCs. I had done judo as a kid.

I think that, like everything else, mental toughness is partially heritable and partially developed. I am sure Dave got through SEAL training because he’s (always been) a tough guy, but being a SEAL probably made him even tougher. Sometimes when we were grappling I could see him mentally focus and raise his game to escape a position or a submission. Perhaps I was imagining it, but I always thought he might be using his SEAL training or experience to do that.

Is the picture of you from Kamakura, Japan? I had a similar picture on my web page for many years.

2011-05-03 on infoproc

No problem for our design unless you can think of an environmental mechanism that boosts genetically ordinary kids up to > +3 SD. Decreased h2 in China would be due to kids growing up in crappy environments, so it doesn’t really affect our high group.

For a big brute force GWAS there is a problem because the measured g is more loosely related to the genetic component.

2011-05-03 on infoproc

Data on h/v/r are sparse for China, although there is probably some literature in Mandarin I do not have access to.

We are going to do the usual things to control pop strat (correct for largest PCA vectors). Things are a bit easier in a Han sample than in a multi-ethnic sample.

2011-05-02 on infoproc

It was prepared using latex on OS X so there might be issues. I notice occasional non-portability (usually due to figures) but it’s rare. You might try google docs.

2011-04-30 on infoproc

It’s a question of statistical power. He didn’t have many outliers. You can get those either by selection or having a huge sample size. In any case, the percentage of variance in genes that are explicitly identified is likely to be small for some time to come. However, by using a different kind of analysis one can estimate total variance accounted for by, e.g., the entire SNP set and that is significant in the case of, e.g., height.

2011-04-30 on infoproc

Full sequencing of 100k+ people is expensive and it’s not easy to get the phenotype measurement here. If you have a ready supply of extreme cases you can save money on sequencing and on administering IQ tests. You have to do some math to figure out the relative power of the two designs.

2011-04-28 on infoproc

15 vs 16 SD is unclear, but she uses 15 in the figure and in her discussion.

Some people doubt whether (say) a +4 SD person has an advantage over a +3 SD person in science and I think Roe’s study (+ SMPY) are the best for addressing this question with a large data set. I like Towers’ essays but I think we might disagree slightly on a few things.

Unfortunately the post 1995 SAT is less g-loaded than the old one.

IIRC, the cohort of SMPY that was above the 1 in 10k level at age 13 tended to score above +3 SD when they took the SAT as seniors in HS. There was some regression as the average was not as high as 1 in 10k, but that is to be expected.

2011-04-27 on infoproc

SMPY has data on how their subjects scored on the SAT as seniors in HS. There is some regression, but the scores are pretty high (> +3 SD as I recall).

Roe did not use a conversion to SB at all. How could that make sense for M, V and Spatial scores reported? What they did (as far as I remember from reading the book) was test a bunch of students, get their raw score mean and SD on each test (M,V,S), then convert the scientist’s raw scores to a deviation score (using SD=15, at least from the figure in the book) relative to students. So, if the students’ raw score average on test M was 20 with SD 5, and a scientist scored 35, then I think Roe gave the scientist an M score of 145. (There was probably some correction in mean since the students were above average compared to the general population. They could probably guess this correction since the students took a battery of ordinary tests that had been normed on the general population.) Note this method does not tell you how rare a particular score is in the general population because the raw score to rarity relation could be nonlinear (the population distribution of scores does not have to be exactly normal).

All that is important (for my interests at least) is that the eminent scientists scored much higher than the average PhD. This suggests some causality: that high psychometric scores predict success in science. I do not know how rare (in a population sense) the scientists’ abilities are: I agree with you that the score of 190 on the math section doesn’t mean 1 in a billion+ ability. It just means that that particular scientist had a *raw score* which was 90/15 = 6 SDs better than the average among the students. However, since normality applies usually out to at least 3 SD, we would have to guess that the 190 = +6 SD score is much rarer than, say, 1 in 30k (which is +4 SD for a normal distribution).

Actually the description in the book is a bit ambiguous, so it’s also possible that they did something more sophisticated than what I just described using (ordinary, low ceiling) V,M,S tests that had been normed on the general population and also given to the students. They could have tried to fit the mean and SD on the hard V,M,S tests by knowing the student scores on both the ordinary and hard V,M,S tests. In this case there would an error in the fitting and that should have been quoted but was not. As I said, it’s obvious Roe is not a psychometrician and she received help from a colleague (I think I even remember his name: Irving Lorge). Sorry if I went on too long… 🙂

2011-04-27 on infoproc

Roe did not use Stanford Binet and it was not a 16 pt SD. That is an internet rumor. If you read my comments on the Roe thread I summarize exactly how she normed her tests. I had a copy of the book out from the UO library when I wrote the post so I was not speculating.

2011-04-26 on infoproc

I’ve switched to using Steve because it is shorter and there is no chance of confusion in the spelling 🙂

2011-04-26 on infoproc

Percentiles are for financial success. In the other examples we assume no additional info on the individual, so we’re just talking about probability of reaching a particular financial milestone.

2011-04-26 on infoproc

Here’s a payoff table for you — tell me if you agree 🙂

Individual is a smart, conscientious, driven kid with Tier 1 credentials. But no superpowers 😉

startup / tech

mid: ~$200k/yr, no hit, stuck working for big company as VP, no early retirement

90th: small hit, makes a few million, still working as above

99th: big hit, makes >$10 million, retires early, maybe rinse and repeat

99.9th: big, big hit, $100M payoff, livin’ large 🙂


mid: ~$400k/yr, lives in NJ, can’t retire

90th: ~$2M/yr for 10 years (mid-career), retires at 45-50 with $7-10M put away

99th: same as above but with several times higher net worth and earlier exit.

A professor comes out (financially) behind both of these tracks, but perhaps happier, more relaxed, more job satisfaction.

If we further restrict the group of individuals to Tier 1 PhDs in physics, it seems to me that the 90th percentile outcome in finance described above is more like 60-80th percentile. What do you think?

2011-04-26 on infoproc

I thought your review was an excellent piece of writing on its own — a very balanced treatment of the issue.

I would like to read Caplan’s response — perhaps you should ping him and give him the chance to comment.

PS “a very large fraction of twin and adoption studies was done in the Nordic countries” — should be “were done”.

2011-04-26 on infoproc

When you say median pay is $400k for someone who could have been a tenured prof, do you mean that this subset of people make less money in general than the people described in the article? Perhaps because they are on the nerdy side?

2011-04-26 on infoproc

Who is the sandwich guy?

2011-04-24 on infoproc

I think you are arguing against a straw man. No one is saying SAT and GPA are everything. The more serious claim is that even after extracurriculars, legacy status, etc. are controlled for there is an Asian penalty assigned (de facto) on the basis of race. As far as I can tell, Jewish applicants are not assessed any such penalty these days, although they certainly were in the past. Please read the comments on the previous post to get a more nuanced idea of what people here are talking about.……

You write as if Asians are one monolithic group, whereas what is really happening is that individual students (together with their families) are each deciding, for their own reasons, that they really want to attend an elite university. Just because a large fraction of Asians are highly motivated to attend an elite, and a higher percentage (compared to other groups) are able to put together a strong HS record, is hardly reason to accuse the group as a whole of “piggishness”!

2011-04-23 on infoproc

What you wrote above is just terribly confused…

2011-04-23 on infoproc

Have you tried the RAPM? It clearly loads on reasoning and memory because you have to reason about what is happening to the objects in the picture, holding several things in memory at one time. (You seem to be confusing spatial pattern recognition with a pattern in the *evolution* of a picture according to an algorithm or in “time”.) There might be loading on verbal ability depending on how you try to solve the problem (some people kind of talk to themselves when conceptualizing what is happening in the pictures — they make a list of things in the picture, how they change, etc.). In any case, as I hinted in my previous comment, the separation between these factors is not as clean as sometimes implied.

RAPM is clearly not the same as what is usually called spatial ability, which for example tests your ability to visualize shapes, rotate them in your head, view them from various perspectives, etc. RAPM is mainly 2 dimensional and its pictorial quality is kind of an accident of trying to make the test culture neutral.

Your 2nd and 3rd paragraphs are pretty confused. Sometimes people don’t understand that they don’t understand something. I can’t give you an entire course on psychometry, but keep in mind

1. the g loading of RAPM can be (and usually is) tested on a group who are all from the same culture and background. So the claim that RAPM is most g-loaded is not a consequence of trying to do cross-cultural testing.

2. my impression (perhaps incorrect) from your previous comments is that you believe that the factor with the highest g-loading gives the best estimate of g. This is only true if the alternative is to use only a single other factor. Once a test like RAPM loads on multiple factors it could easily be more g loaded than the single factor with highest g correlation (e.g., verbal or reasoning). I am sorry but you need to learn a little math to understand this point. Ideally you measure ability scores on all factors and then *sum them up* with appropriate weightings.

I keep writing RAPM because I am more familiar with the “Advanced” version which has a higher ceiling, but my comments also apply to RPM.

2011-04-23 on infoproc

Aaron, it’s kind of obvious from what you’ve written on this point that you do not understand factor analysis (or even linear algebra).

Just because the verbal or reasoning factors have the largest g loading out of the 7 factors you listed (which are themselves a bit controversial among experts), doesn’t mean that a test like RAPM can’t be more heavily g loaded than any of the single 7 factors. It’s obvious that RAPM has a visual aspect, but it obviously also loads on reasoning, memory, perhaps even the verbal factor depending on how people solve the problems. All of these factors overlap in a complicated way. g isn’t reducible to any single factor, in fact it is the certain special combination of all those factors that best predicts general performance on a battery of tests.

I’m not even a big believer in the “strong” interpretation of g. But it’s just mathematically obvious that the argument you give is not correct. Look up the empirical results on g loading of various tests.

Also, look at GRE scores or SMPY data and you’ll find that subjects like philosophy are definitely more verbally loaded than, say CS or engineering. SMPY shows very different career trajectories for verbally gifted (law, humanities), mathematically gifted (engineering, math, CS) and both V and M gifted people (theoretical physics?) that they’ve tracked for 50 years now.

You seem very interested (obsessed) with this topic, but you don’t seem to have done any research other than misinterpreting some things you’ve read on the Web.

2011-04-23 on infoproc

Very strange that you would denigrate Russian science and mathematics. (Do you know what you are talking about?) The whole point is that these things change over time, and over timescales too short for there to be anything but a cultural or institutional explanation (gene frequencies don’t change much in 100 years). Please have a look at Needham’s book. You will find many examples of individual creativity or even genius that are entirely unknown in the West. “History” (as constructed by any particular group) is a very imperfect record of what happened in the past.

2011-04-23 on infoproc

I have 10+ years of SAT data from the University of Oregon. It shows that only very rarely do SAT scores move by as much as 1 SD on retest. So most methods of preparation do not work, or only yield modest improvement. Obviously many people spend money on things that don’t work. Astrologers are still in business, sometimes even in the White House.

2011-04-23 on infoproc

Expensive equipment is not the only important factor. 150 years ago most parts of the world had no organizations dedicated to scientific investigation. You may as well attribute the failure of Spanish or Russian scientists to anticipate Newton to lack of creativity!

This topic is quite complicated and has been studied from a variety of perspectives. I suggest you go to Google books and read a bit from Needham’s Science and Civilization in China. While Needham probably exaggerated some of his claims, you will find all sorts of things (ranging from pure math to very practical devices) that may have been invented in China hundreds or even thousands of years before they were developed, e.g., by NW Europeans. Progress in science and technology depends on lots of things — institutions, resources, culture and (of course) individual abilities.

2011-04-22 on infoproc

I agree. If you praise ability kids can get insecure about it. Praising hard work and effort is always good.

2011-04-22 on infoproc

He might know the odds are against him, but is still willing to try 🙂

2011-04-21 on infoproc

If you can find the reference I would be quite interested. If you’ve never looked at The Chosen by Jerome Karabel, I highly recommend it.

Another guy who understood this stuff very well was Bourdieu, although he didn’t believe in g.

2011-04-20 on infoproc

Any psychometric test is only a rough estimate of true underlying ability. A reasonable error bar (if you are familiar with the theory of error estimation) would probably be .5 SD or so, where here SD refers to the population standard deviation. That means that occasionally a person’s true ability is severely (e.g., by more than 1 SD) over- or under-estimated by the test. However, when comparing groups of individuals (say, a group that all scored 1300 and a group that all scored 1500, so more than 1 SD apart) you can be pretty confident that most of the people in the higher group are actually “smarter” than those in the lower group. This is of course assuming no other systematic differences between the groups (e.g., differences in test prep). From data I’ve seen, SAT test prep *on average* has no more than a .5 SD or so effect.

I do agree that success in life is influenced by many factors beyond g. It’s just that those other factors are even harder to measure in a quantitative way.

Re: Aaron and Asian scientific achievement, just imagine what the Greeks or Chinese would have said about Germanics in, say, 500 AD: “those guys are still savages”, “they’ve accomplished nothing and we have already invented X, Y, Z …”, etc. So it’s very tricky to predict these sorts of things over long timescales. Anyone who is actually familiar with the details of E. Asian development in the last 20-30 years is likely to expect a lot more innovation and breakthroughs from the east in coming decades. There were almost no real research universities in Asia ex-Japan until just 10-20 years ago (everyone came to the US to get a PhD), science funding levels in Asia (*including Japan*, at least for pure science as opposed to technology or corporate R&D) are lower than in the US and Europe, etc., etc. But these things are all changing and the trends are pretty clear.

2011-04-19 on infoproc

An extended Turing Test, conducted over, say, a week or month, is relevant because the tester could easily evaluate the *learning ability* of the counterpart.

2011-04-18 on infoproc

Espenshade’s analysis already accounted for (4): the 3x white admit rate for applicants with similar grades, scores, legacy status and athletic status does not include legacies. Espenshade cannot address (2) very well, but the internal Stanford study tried to include it in their analysis and found anti-Asian bias even after controlling for “leadership qualities”. (See earlier posts if you don’t know what I am referring to.)

There is no perfect academic study of admissions rates, but anecdotal evidence plus the studies suggests that the 3x effect is not simply due to race-blind evaluators responding to population differences in non-academic characteristics such as extracurricular activities. See also the Harvard data from around the time (1990s) they were under investigation for anti-Asian bias. There is very suspicious variation in admit rates for Asians. How could that be if everything were fair and race-blind?

You obviously see a practical advantage to having your kids claim to be white. Why is that, if you think Asians are actually admitted at “appropriate levels”?

2011-04-18 on infoproc

Nice! The reviews are pretty negative, though.

2011-04-18 on infoproc

I particularly liked this paragraph from the second link:

Meanwhile, many Hispanics are solely or primarily of European
heritage, and about 50% identify themselves as “white” on census
bureau forms. UT’s approach (supported by the Obama Administration),
is that a state university can and should favor white descendants of
Spanish conquistadors or Italian immigrants to Argentina or Jewish
Mexicans of Eastern European descent–I know people in all the latter
categories–over a dark-skinned child of Vietnamese boat people, solely
because the former have Spanish-speaking ancestors.

2011-04-16 on infoproc

Hmm… I’ll have to listen to Levy again; perhaps I wasn’t paying close enough attention. What you say about Goto is true, of course.

2011-04-14 on infoproc

Gross is an amazing guy. When it comes to ideas for startups he may be the single most prolific entrepreneur in the recent era. He shares a lot of good insights in the interview I linked to.

When I was a student there was very little in old town except a couple of dive bars and obscure restaurants. We used to go to a place called Barney’s to eat and watch the Lakers-Celtics finals. Last time I checked it was still in business. Don’t know which bars Feynman went to but I remember him carrying a sketchpad around 🙂

2011-04-11 on infoproc

You might be right. I guess I was favoring the FT for being a bit more sophisticated in analysis. But for actual reporting the WSJ might do a better job.

2011-04-10 on infoproc

The Financial Times went the other way — this trick doesn’t work, and outside links still hit the paywall. I think the FT has the best coverage of finance and economics of any daily, including the WSJ.

2011-04-09 on infoproc

Hmm.. very interesting. Is this the same design Myhrvold is backing? (TerraPower?)

2011-04-09 on infoproc

He was a senior fellow when I was a junior fellow so I saw him almost weekly at the Monday dinner. He was a very open guy, great to talk to. I distinctly remember him describing how anti-semitic Princeton was when he was there as a student. During a conversation with an administrator (a dean or something) he was told “You people are just lucky to be here”.

I knew he had thought a lot about death (“the problem of death”). So when he came down with cancer while still relatively young I really felt for him.

2011-04-08 on infoproc

Yup. But I had already taken the “equivalent” of Ph1 (Halliday and Resnick) and a QM course (junior level, in their sequence) at Iowa State. In my year there were also 2 kids who took Ph12 as freshmen, IIRC. The other guy also graduated in 3 years and he started his PhD under Wilczek at UCSB, but then switched into mathematical logic or something. He’s now in Silicon Valley doing chip design stuff. The CEO of his startup (also a techer, and a famous hacker) refers to him as their “secret weapon” 🙂

What happened to your roommate? DE Shaw? 😉

2011-04-08 on infoproc

The guy on the far left of the group photo (also a Caltecher, the guy on the far right is an MIT man; both were physicists then but are now financiers) and I met the girls in 1994 while traveling in Portugal after a conference (see pictures above). We returned to spend a few weeks with them in Conil in the summer — they had leased an apartment just off the beach. Every night was discotecas until dawn and then sleep until afternoon and head to the playa — the Spanish way 🙂 My friend is now married to one of the girls — the one dressed in black — and they live in Manhattan.

I have a picture of us sitting at a thatchy concession stand at sunset — I wonder if it’s the same one?

2011-04-08 on infoproc

Dos Sevillanas 🙂

2011-04-07 on infoproc

He’s sufficiently sociopathic that he is willing to use a proxy, fake email address, and fake disqus account in order to post each batch of juvenile comments. I already have his home IP address blocked. The work factor I’ve described would be enough to deter most normal people.

2011-04-07 on infoproc

If it holds up it’s a big deal. But it’s only a 3 sigma result, so there’s a good chance it will just go away.

2011-04-06 on infoproc

You may see comments from the nut I’ve banned from this blog (he posts under lots of pseudonyms). Please don’t respond to the comments as I’ll delete them as I find them.

It just takes me one click to delete lots of his comments, which forces him to re-register. It just shows how unbalanced he is that he can’t take a hint and go away.

2011-04-04 on infoproc

Times change.

“instantiation” is used in software development.

“modernity” is used by, e.g., the NYTimes — are they lit crit or French?

2011-04-04 on infoproc

BTW, the Terman study started in 1921, so the participants were still kids or children in 1922. I don’t know how much of the personality data came from the different evaluations over different years. Terman et al. might comment on the stability of personality over the period of the study. You can consult the many volumes they wrote on their results.

2011-04-04 on infoproc

If you don’t like g, then you probably won’t like Big 5 or personality factors. All the results for correlations, predictivity, stability, heritability, etc. are weaker for personality than for g. (There are also systematic group differences, but let’s not go there…)

It is interesting that if you do a purely statistical (factor) analysis of words used to describe people, most of the variation is captured in just 5 or so variables. (If you think about it, we usually only need a few variables or primitives to describe someone else’s personality, at least in a crude way.) Whether you can measure those variables through self-report (the data shows third party descriptions correlate reasonably with self-report), whether the results are stable (the claim is they are), etc. are detailed questions you have to look at more carefully.

Below is a link to the web page of one of my colleagues (a UO psych prof) who works in this area.…

It’s strange that you, anon, are so negative about this study, since it confirms your repeated claim that people in the US get ahead by being “pushy” (Conscientious and Extraverted?) rather than “smart” (as measured by tests).

2011-04-04 on infoproc

Please read more carefully. I wrote “could maintain A modern level of development” not the SAME level of development. I agree with your point, but it doesn’t contradict what I wrote. Of course no country has world class capability in every single area (well, maybe the US does, almost). But there is a qualitative difference between a country that has strong capability in most categories and one which is only capable in a few areas.

I repudiate the nutty Chinese nationalists who sometimes comment on the internet. Does that make you happier? The Confucians who under-promise and over-deliver are, as I wrote, the traditionalists, which not every Chinese person is.

2011-04-02 on infoproc

Yes, it will be interesting to see how it evolves. Considering how easy it appears to avoid the paywall they might be disappointed in the number of subscribers.

Do subscribers still see ads? Not so long ago people would have complained about paying a fee and having to see ads.

2011-04-02 on infoproc

They spent $40 million on that? 🙂

2011-04-02 on infoproc

I didn’t mean the post to be about HBD or cultural limitations. I agree with you that institutions and historical path-dependence are very important for understanding what is going on.

The title “Silicon Valley + Africa” was not meant to be about IQ (or culture), but rather about the current development pattern.

However, there is an India-optimist tinge to a lot of media coverage (as remarked in the comment by RKU below; paging Tom Friedman!) that suggests India is somehow going to make it to modernity through software and services alone (well, some Indians are, but the majority are not). This doesn’t make sense for any country, whether it is China or India or Germany. The dirty energy-intensive path of industrialization is, unfortunately, the only way. It may also be true that draconian population controls (like the one child policy) can’t really be avoided.

2011-04-01 on infoproc

Yes, I really liked it as well.

I wonder whether physicists who take Math 55 really get much benefit from it, other than a sense of accomplishment. Seth Lloyd described his Math 55 experience to me and his recollection of the score distribution was similar to yours.

His favorite story went like this: famous Japanese mathematician (the professor) is lecturing to class: “It is obvious from lemmas 1 and 2 that …” Pauses, frowns, turns around and scribbles some calculations on the board that the class can’t see … after a while he turns around and begins again “Yes, it is obvious …” 🙂

I placed out of the usual math sequence at Caltech and took Ma 108 my freshman year, which was taught by Tom Wolff, a Ma 55 alum and contemporary of Bill Gates (see link; I didn’t realize it then but I was taking analysis from a world-class (Bocher prize) analyst). At the time I was still considering math as a possible direction, but Ma 108 helped me realize my real interests were more in theoretical physics than math. At the same time Wolff was grinding us down with problem sets I didn’t find that interesting, physics 12 was introducing the path integral and other nice stuff.…

2011-04-01 on infoproc

A couple of comments, based on personal experience only.

1. Aggressive or glib salesy types are more common among the S. Asians. There is even a stereotype among E. Asians (from Asia, even in my father’s generation) that S. Asians can be big talkers who don’t back it up. In Confucian culture, this is considered very bad. Traditional E. Asians are more “under promise — over deliver” types.

2. It’s harder to operate in the aggressive or salesy mode if your English is not good. There are plenty of those types in, e.g., China, but they operate in Mandarin. When you make your generalizations I believe you are including recent immigrants (as opposed to strictly S. and E. Asians raised in the US) in the discussion. That heavily biases things toward Indians as their elites have been using English all through school (legacy of British colonial experience). There are very few recent immigrants from E. Asia who can operate in English at a high level. Ever look at the number of cognates between Hindi and English (two Indo-European languages)? On the other hand it is extremely hard for, say, a native Korean speaker to learn English late in life, or to ever become fully fluent.

3. Almost all of the E. Asians raised in the US come from immigrant families where the parents’ English is not good (even, I would bet, Amy Chua’s dad the Berkeley EE prof. spoke accented and occasionally grammatically incorrect English). This definitely affects V scores, although I would bet more of the differences between (US raised) E. and S. Asians are due to culture or personality.

4. As anon (our favorite commenter with so many pseudonyms) likes to point out, what works in one place (loud superficial guy is seen as “natural leader” or CEO material, or even superior Ivy material) may not play in other cultures or contexts. Would our society work better overall if individual personalities (which are partially genetically determined and may, on average, vary between groups) or culture were such that quiet consensus builders were seen as leaders as opposed to loud braggarts? (Note I am not characterizing S. Asians as such — I have nothing but respect for their great accomplishments.)

2011-03-31 on infoproc

Will a double dip in housing cause a double dip recession?

2011-03-31 on infoproc

That’s a nice bio. I had read it before but forgotten about the mention of Math 55.


“With physics and math, I could never fi gure out a way to contribute,” says Stallman, recalling his struggles prior to the knee injury. I would have been proud to advance either one of those fi elds, but I could never see a way to do that. I didn’t know where to start. With software, I saw right away how to write things that would run and be useful. The pleasure of that knowledge led me to want to do it more.”

I’ve heard this kind of feeling expressed by a lot of people who became software developers.

2011-03-31 on infoproc

The wild card is of course the possibility of social upheaval in China, but this seems relatively unlikely to me. The articles I quote from in the post make it seem like social unrest is actually a possibility in India.

2011-03-31 on infoproc

It could be lots of things. Drive, ambition, willingness to be a maverick, etc.

Finding the correct solution when you know one exists is not the same thing as actual research, in which sometimes posing the right question is more important.

Hmm… was your professor named Lenny?

2011-03-31 on infoproc

He claimed that he had lost points on one problem, but that it was a mistake in grading and in fact his solution was perfect. You can take it up with him… 🙂…

2011-03-31 on infoproc

This was analyzed in an AMS article I linked to (search under putnam on the blog). Many IMO kids from other countries come to the US or to Canada for higher ed, so that (these days) the Putnam ends up pitting a lot of gold medalists against each other every year. When you see a kid with a PRC name at an obscure university, who nevertheless scored high on the Putnam, you will often find he is from China where he probably did well in their national competitions (maybe even was on the national or provincial team) before coming here. I bet a lot of the Korean and Eastern European names on the list are similar stories. Many schools now recruit top IMO, IPhO, etc. performers from abroad with scholarships. I even made that a positive criteria for the scholarship I endowed at Caltech in my father’s name.

It’s also true that some kids who were big HS competitors don’t want to invest the same level of energy in the Putnam while in college. One IMO gold medalist I know who went to Caltech told me that his “problemist ability” (as I refer to it) only gave him a 1 SD advantage over the other kids in (the relatively hard) classes they had to take. So even these guys have to work in college (at least at certain places in certain majors) and may not have as much time to prepare. At Caltech prep for the Putnam was very casual. There was usually just one meeting, typically led by a superstar problemist. In my day it was Barry Simon, who claimed to have earned (at that time) the only perfect score.

2011-03-30 on infoproc

I think you misunderstood my point — there is no place where more than a small fraction of the population can write code. That’s why you need stuff like manufacturing.

2011-03-30 on infoproc

Anon, are you off your meds again? Why the bipolar behavior — contrite one moment, then suddenly so aggressive? If you want to be blocked from commenting, just say so.

2011-03-30 on infoproc

China’s development is broad based and real. They have capabilities across a range of technologies and industries, and have first rate infrastructure in many places. If the rest of the world vanished (with the exception of oil producing states!) China could maintain a modern level of development using just its domestic industry and technology. There’s not much they can’t do at this point, from chips to software to airplanes to cars to nuclear power to satellites and space launch. I think India is on the other end of the spectrum – very unbalanced.

There are perhaps more prominent Indians in the West than Chinese, but two key drivers of this are the use of English by the elite in India and the fact that the Chinese economy can utilize so many more talented people at home than India can. 100+ billionaires in China, all created just in the last 20 years or so (albeit many made their fortunes through corrupt practices).

2011-03-30 on infoproc

I was hoping Noam might enlighten us 😉

I suspect Allen’s recollection of the story is just incorrect.

2011-03-30 on infoproc

I think they are experimenting. Not with technology, but maybe looking at behavior.

2011-03-30 on infoproc

Yes, he’s said that in a number of interviews. Coincidentally, I know people from his eating club and vintage at Princeton, so I think I know the actual individuals who scared Jeff out of physics 🙂

2011-03-30 on infoproc

If I read the criticism correctly if appears even the 20% cardiovascular mortality reduction might be an artifact of their stopping the study early.

What do you think of the Ray et al. meta-analysis, which covered high risk but healthy (no cardiovascular disease) subjects?

2011-03-30 on infoproc

Response to this response is linked to in the post above.

2011-03-30 on infoproc

I’m also not seeing the paywall yet, despite having read lots of articles.

2011-03-30 on infoproc

“majority of patients currently receiving treatment”

Does that mean statins?

2011-03-30 on infoproc

I’ve been waiting for over 10 years for a good micropayments system. I even spent some time thinking about it as a startup idea long ago. But you need scale to make it work. I think Google could roll it out and be successful — I seem to recall related announcements — but where is the product?

2011-03-30 on infoproc

They claim that post 2005 the ethics and disclosure requirements for medical studies have become much stricter, and not coincidentally it is the later studies that show no benefit from statins. But maybe they are nuts…

2011-03-30 on infoproc

The paleo logic is hilarious: e.g., primitive man had no access to clean water, so clean water must be bad for you! 🙂

2011-03-30 on infoproc

Sam, thanks!


Do have a link to an article that summarizes recent results on relation between cholesterol levels and morbidity/heart disease or on clinical trials of cholesterol lowering drugs?

I found this article interesting but want a broader picture (in case these authors are biased):…

2011-03-29 on infoproc

Thanks for the nice comment. Can you provide any links on statins (particularly your 2nd and 3rd points)?

2011-03-29 on infoproc

>Thank goodness that’s the last time I’ll try.< If only that were true!

2011-03-29 on infoproc

>Thank goodness that’s the last time I’ll try.< If only that were true!

2011-03-29 on infoproc

I strongly suspect that people vary in terms of which diets work for them. I find it funny that the subset of paleo people who are also HBDers often haven’t figured this out.

2011-03-29 on infoproc

Anon, we’ve discussed this before. I acknowledge that heritability is defined wrt a specific range of environments. In the case of IQ I am specifically interested the effect of genes in the “prevailing environment”, which I expect to prevail, insofar as producing scientists, for some time.

2011-03-29 on infoproc

Thanks for the references! I have friends (some are physicists) who endorse the ideas you mention, so I’ve heard them before. But even they would admit the confidence level in most of this science is not that high. One such friend characterizes nutritional theories as analogous to religion 😉

2011-03-29 on infoproc

A lot of that difference is probably early mortality due to disease. Primitive people (even more primitive than Afghans) who live to adulthood can often have very long lives. So the statistic you quoted doesn’t necessarily imply their diet is worse.

2011-03-29 on infoproc

I try to cut down on carbs and eat a lot of fruits and vegetables and protein. But I don’t necessarily buy the whole ideology. I just find it’s easier to restrict calories by restricting carbs.

2011-03-29 on infoproc

She’s very smart so can overcome much of the brainwashing, probably like your girlfriend. She also has a good tutor 😉

2011-03-29 on infoproc

I’ve read bad things about statins and I’m only borderline high so I’m resisting for now.

2011-03-27 on infoproc

The Chinese seem to be taking a slow and cautious route with this. Let’s hope that if they decide to proceed with rolling these out it will be because they’ve solved the safety issues.

2011-03-25 on infoproc

The Case-Shiller data (see second figure in post) seems to show that housing is a good inflation hedge — holds value in real dollars. The cost of the leverage will of course fall with inflation.

2011-03-22 on infoproc

This is getting into fine details, but fight night weights are probably:

GSP 185
AS 195
Jones 220

So, despite what most fans think, AS is closer in weight to a big WW like GSP than he is to a big LHW like Jones.

Having said all that, I would love to see AS fight either of those other guys.

2011-03-21 on infoproc

I just read the summary at Wiki and it sounds interesting. Amazing how much Dick stuff I’ve never heard of.

I highly recommend TMITHC. Somewhere on my blog I link to The Sound of His Horn, which influenced TMITHC, and which I also recommend.

2011-03-21 on infoproc

I agree about Dick’s appeal to physicists. The recurring question “what is real?” resonates with us 🙂

2011-03-21 on infoproc

I’m not the only one… Try this query: “schizophrenia philip dick”

2011-03-20 on infoproc

I suspect he is ambivalent about taking the fight, which he would (in my opinion) likely lose.

GSP on fight night is probably around 185 and Anderson is maybe 195. So it’s not a big difference. If GSP puts on 5 lbs of muscle and they have it at a catch weight of 180 or so then it seems pretty fair.

But back to the main point, Jones is a lot bigger than AS. Jones cuts water from about 217 and a 217 lbs AS is fat.

2011-03-20 on infoproc

Not sure Shogun was on his game after the long layoff and knee surgery. He may be in long term decline due to all the wear and tear. He also should be fighting at 185.

Nevertheless, Jones is amazing. It’s interesting that he wasn’t any good at basketball or football. I think he has two brothers (both bigger than him) who play in the NFL.

I would take Jones over Silva at 205. But I think Silva is actually not that big a middleweight — I think he’s only 195 or so before his water cut. GSP has no size excuse for not stepping up against AS.

2011-03-17 on infoproc

TEPCO? What’s their liability exposure under Japanese law? Couldn’t the stock go to zero? Is it already almost there?

2011-03-17 on infoproc

I predict that they will.

2011-03-17 on infoproc

Are you buying the Nikkei?

2011-03-08 on infoproc

Very plausible that there is a Flynn effect for attractiveness for the reasons you mention.

You can use Google images to find other pictures of BP. Definitely above average but not a supermodel by any means 🙂

2011-03-07 on infoproc

I don’t know enough about Australia but usually when houses appreciate faster than rent or other fundamentals, it’s a bubble.

2011-03-07 on infoproc

>I didn’t know there were von Neumann’s of geology, chemistry, meteorology.< See Rutherford on stamp collecting. >For scientists who get results rather than spin theories genius is useless without hours in the lab.< Luis Alvarez was arguably the greatest experimenter of the 20th century and made many practical contributions (even to geology). I think his opinion should be weighted more heavily than yours.

2011-03-06 on infoproc

It means vN is a lot smarter than the rest of us.…

Lev Landau, a Nobelist and one of the fathers of the great Soviet physics system, had a logarithmic scale for ranking theorists, from 1 to 5. A physicist in the first class had ten times the impact of someone in the second class, and so on. He modestly ranked himself as 2.5 until late in life, when he became a 2. In the first class were Heisenberg, Bohr, Dirac and some others, and Einstein was a .5! (For reminiscences of great physicists in that generation, see From a Life in Physics.)

My friends in the humanities, or other areas of science like biology, are astonished and disturbed that we physicists think in this essentially hierarchical way. Apparently, differences in ability are not manifested so clearly in those fields. Personally, I find Landau’s scheme appropriate. ……

Luis Alvarez: “The world of mathematics and theoretical physics is hierarchical. That was my first exposure to it. There’s a limit beyond which one cannot progress. The differences between the limiting abilities of those on s