BHTV: Eliezer Yudkowsky & Razib Khan
Link.
“Razib Khan has an academic background in the biological sciences, and has worked for many years in software. He is an Unz Foundation Junior Fellow. He lives in the western United States.”
Razib’s writings can be found on his blog, Gene Expression.
- [Link] Why the kids don’t know no algebra by (4 Jul 2012 10:29 UTC; 30 points)
- [Link] Reason: the God that fails, but we keep socially promoting…. by (30 May 2012 10:59 UTC; 22 points)
- [Link] The real end of science by (3 Oct 2012 16:09 UTC; 21 points)
- [Link] Knowledge, not opinion, information extraction, not persuasion by (3 Oct 2012 21:13 UTC; 5 points)
- [Link] The perils of “reason” by (13 Aug 2012 6:37 UTC; 3 points)
- 's comment on Help: Info on intelligence-focused genetic engineering? by (6 Oct 2010 16:15 UTC; 2 points)
Best BHTV I’ve seen yet, Razib really knows his stuff. On a sidenote, two of Eliezer’s sentences towards the end reminded me of Paul Graham’s essays:
“I certainly think that there are more and less skilled books and there are more and less skilled authors”—How Art Can Be Good
“It is precisely where you do not find prestige that there’s still any actual test of how good you are as an artist”—The Power of the Marginal
I was not impressed by any means: It was so painful to watch that I had to stop after 15 minutes.
It is well possible that RK is competent, but his lack of confidence (or communication skill) made his numbers look like as if they would have been just pulled out of his behind. Additionally, his slowness of adequately reacting to EY’s questions left me very much underwhelmed.
Yep, Paul Graham has an essay about that too: Why Nerds are Unpopular.
Almost everyone I know is nerd, so this is my baseline for comparison.
A lot of nerds have excellent communication skills (although it does not necessarily make them popular) For example Eliezer always makes a very good impression. A lot of academic people are also excellent presenters and discussion partners.
One of my concrete problems with Razib was that he somehow felt the urge to immediately start to talk (even if what he said was loosely related to the question) and it looked like that he was grasping the question while he was talking. For example, Eliezer is very good at taking time before answering a question, compressing the answer and being completely relevant and focused.
The other issue was that every time he mentioned concrete data, he did it in a way that made it seem as if he would have made it up on the spot. Rationally, I may not have had the reason to doubt the numbers, but the nonverbal clues made me subconsciously telling to myself: “These figures are not reliable!”
Both of these problems were unrelated to nerdiness, as a lot of nerds simply don’t exhibit them, for example I thoroughly enjoyed diavlog between Eliezer and Scott Aaronson.
I’m a little surprised that Eliezer was not able (or willing?) to correctly guess who would be extremely interested in genetic differences between Ashkenazi jews and American gentile whites.
Surely somebody as smart as Eliezer (and who spends a lot of time online) is aware that there is lots of racism/anti-semitism online.
ETA: By the way, I really enjoyed listening to the discussion.
Due to the Ashkenazi “superiority” in terms of IQ when compared to gentile whites, white supremacists would not have struck me too as an obvious first guess.
This was a survey of ancestry/genetic structure. Not trait values. Here is the post, just page down to the image and you can see why they were interested:
http://scienceblogs.com/gnxp/2009/01/how_ashkenazi_jewish_are_you.php
Got it. Makes more sense with the data. I guess in the flow of conversation, we do tend to miss a couple of things :)
This was the most difficult diavlog of the seven I’ve done so far. 60 minutes just wasn’t enough for this conversation and I had to do a lot of on-the-fly judgment of what I should elaborate on, and what I should simplify or elide for the sake of concision. The others were more “interview” formats where there was a more formal structure, as opposed to riffs. Also, Eliezer and a subset of the audience knows a fair amount of biology, but the majority of the audience did not, so I kept having to navigate between these two tensions.
I assume Eliezer had similar issues when we got the section where Kahneman & Tversky’s ouvre were implicit background assumptions, but we’d burned through 2⁄3 of the time by then so the choice on whether to elucidate or not was made for him :-)
The conjunction fallacy is a subset of Occam’s razor? Hmm. Yes—I had never thought of it like that before.
I didn’t watch the video, but I don’t see how that could be true. Occam’s razor is about complexity, while the conjunction fallacy is about logical strength.
Sure ‘P & Q’ is more complex than ‘P’, but ‘P’ is simpler than ‘(P or ~Q)’ despite it being stronger in the same way (P is equivalent to (P or ~Q) & (P or Q)).
(Another way to see this is that violating Occam’s razor does not make things fallacies).
The actual quote is:
″...which of course violates the conjunction rule of probability theory—also known as Occam’s razor—which says that a more complicated event cannot be more probable than an a strictly simpler event that includes the more complicated one.”
41 minutes in.
The conjunction fallacy applies only when you already have a probability law. (a specification of a probability space). It applies to events in a probability space. The conjunction rule proscribes assigning a subset event higher probability than the event containing it.
Occam’s razor is prescription for what probability laws should look like (e.g each program having a prior probability of (1/2) to the power of its code length in bits). i.e. what constitutes an outcome in the probability space, all outcomes having equal probability.
The conjunction fallacy really says nothing about prior probabilities. The conjunction rule is a theorem in probability. Occam’s razor is a working rule for assigning prior probabilities to hypotheses.
These are two distinct, legitimate uses of the term “Occam’s Razor”. The conjunction rule is the everyday sense; what you’re talking about is a deeper, philosophical sense.
Well I agree. In our everyday lives, Occam tells us to chose the simplest hypotheses. Conjunction rule tells us to keep any particular hypothesis we have as simple as possible.
Also, maybe we should apply conjunction rule first to our candidate hypotheses and then only Occam to chose the simplest among them. (of course, Occam by itself already picks out the simplest hypotheses, but i’m talking about a working procedure here)
Prior and posterior probabilities are not made of fundamentally different stuff, and posterior of one calculation can turn out the be prior in the next. Assuming fundamentally distinct sets of probabilities and new ways of popping probabilities into existence seems uncalled for.
You were also suggesting to first use conjunction rule to weed out hypotheses that are less likely, and then summoning Occam’s razor to do the exact same thing again. This too seems redundant.
Agreed, there is no fundamental distinction. You can certainly update existing probabilities which did not take into account Occam’s Razor, to take it into account. What makes Occam pertinent to priors in particular is that you can apply it to anything, which means it can always also be the first thing you apply to hypotheses. So think of Occam as ‘evidence’ that applies to all hypotheses. (note that the conjunction rule is not similarly ‘evidence’)
Yes it is ideally redundant, but i did emphasize I was suggesting it as a working rule. It seems to me less computationally expensive to remove extraneous elements from hypotheses than to calculate or at least rank their complexity.
“For the love of cute kittens”—EY, 21:30 :-)
Down syndrome is not really “on the smallest chromosome”. It results from an extra copy of chromosome 21. (22:48).
Right. I’m gonna have to find a balance between using bio-words and stuff I translate into normal speak on-the-fly. My first thought was to blurt out trisomy 21, and then I would have had to use the word non-disjunction. So I simply noted it was a problem with chrom 21 , and that is the smallest chrom so as to minimize dosage dependent effects.
Allegedly, he is a LW reader (unlike EY nowadays!)
Here are some tantalising excerpts from Razib Khan’s blog:
...
…
…
…
…
For those interested it seems like $5000 is the lower bound of the cost of sequencing a human genome if you believe this press release. Otherwise people say it is in the $50K region.
So a few years away (a few more halvings, then time for the genome + IQ database to be built up) before the large retrovirus genetic brain modification possibility. Does anyone know of work being done towards making viruses delete specific alleles, so we don’t double up when we insert different ones?
I’ve seen some work on this, but it’s a long way off before it gets into humans. I think it’s only been done in bacteria or simple multicellular organisms, but more complex organisms like humans have a lot of barriers before this works… It’s very easy for this to screw up, and it’s very difficult to do. You’d have to be certain that it will cut just that DNA region and ligate it back together correctly. Also, there may be issues with chromatin packaging and correctly recognizing the site. It is at the very least 15 years away, and even then you’d have to deal with the FDA being super cautious.
It would probably be easier and safer to insert the genes necessary to tag the protein product of a specific allele for destruction right away, or to use something like RNAi to destroy it right after transcription.
If they really had the $5000 genome in any meaningful way they would have won the X Prize already. Complete Genomics said in a bullshit press release that they could have won the X Prize if they really wanted but they are too focused on other things now. I am willing to bet $50 (or less, if someone else prefers) at 1:1 that someone will win the X Prize for genomics within two years. http://genomics.xprize.org/
It would seem that finding genes that not only affect IQ but do so in a way that gives benefits even when done on fully developed brain will be an ethical minefield.
You could do tests on chimps and see if the genes associated with chimp IQ are correlated with human IQ. If so, then you could see what introducing them on fully developed chimp brains does.
The other route, is biohackers that hack their own brains. I’m really curious whether the number of people who get their own bioreactors etc will be anything like the number of people who hack electronics today.
Not finding genes for IQ? What about DTNBP1, CHRM2, ASPM, NR2B, HAR1, PYDN?
“What about DTNBP1, CHRM2, ASPM, NR2B, HAR1, PYDN?”
Not replicated, or nothing found. ASPM for example isn’t associated with normal variation in IQ (or the effect size too small to detect, they’ve looked). Please see my coblogger “ben g”’s post on the topic:
http://www.gnxp.com/blog/2010/02/half-sigmas-flawed-post-on-dtnbp1.php
(and no, I’m not one of the people who is excited that we haven’t been able find these genes yet)
It really isn’t hard to find genes for intelligence—assuming that you mean the conventional thing by “a gene for x”:
“Maynard Smith reached for a hypothetical example and came up with a ‘gene for skill in tying shoelaces’. Pandemonium broke loose at this rampant genetic determinism! The air was thick with the unmistakable sound of worst suspicions being gleefully confirmed. Delightfully sceptical cries drowned the quiet and patient explanation of just what a modest claim is being made whenever one postulates a gene for, say, skill in tying shoelaces.’”
One example:
“Gene found for mental retardation”
http://news.bbc.co.uk/1/hi/health/2067864.stm
Yes, it is very hard. I know, because I know people attempting to find those genes. They report that it’s very hard. I specifically said normal variation in IQ to make it clear that I’m not talking about mutations and variants which cause retardation. Those QTLs of large effect are easy to find, but they aren’t implicated in normal human variation. How do we know this? Because they don’t show up in linkage or association studies consistently.
I didn’t say it’s impossible. There were many things impossible 10 years ago that are possible now. But I didn’t make that assertion in ignorance.
Here’s an example of the kind of thing I mean:
“There is “a highly significant association” between the CHRM2 gene and intelligence according to a 2006 Dutch family study. The study concluded that there was an association between the CHRM2 gene on chromosome 7 and Performance IQ, as measured by the Wechsler Adult Intelligence Scale-Revised. The Dutch family study used a sample of 667 individuals from 304 families.[59] A similar association was found independently in the Minnesota Twin and Family Study (Comings et al. 2003) and by the Department of Psychiatry at the Washington University.[60]”
http://en.wikipedia.org/wiki/Intelligence_quotient
So: “We are not finding any IQ genes” seems to be rather inaccurate.
I note that there has been at least one negative finding for the same gene:
“No Association Between Cholinergic Muscarinic Receptor 2 (CHRM2) Genetic Variation and Cognitive Abilities in Three Independent Samples.”
http://www.medicine.manchester.ac.uk/research/PubDetails/index.aspx?ID=37741
The study appears to be looking at SNPs—though it tracks quite a number of them. Possibly the relevant variation has a geographic component.
It is hard for me to find an interpretation of “normal variation in IQ” that makes much sense of what you are saying. Also, I note that that phrase was used in an example.
Intelligence has a genetic basis. We know that some things breaking cause large effects—and that other things breaking cause small effects. There will be a bunch of genes that cause effects of intermediate size when they break—and some of those will be broken in relatively normal people—accounting for some of their variation in intelligence.
You could argue that researchers haven’t yet found any such genes—but my impression is that there is a growing list of likely genes—from things like the “IQ QTL Project”. My interpretation would be that neither theoretical considerations nor empirical studies offer much support to the idea that genes with moderate effects on IQ are particularly uncommon.
No doubt a fair number of such things will actually be caused by genes affecting the gut, muscles, senses, etc—but that gets back to what is meant by a gene “for” intelligence.
“but my impression is that there is a growing list of likely genes”
As I said in my first comment, the list is characterized by non-reproduced associations. I have tracked this sort of research for about 10 years now, and the pattern is a consistent one where a QTL makes a big splash, but there is no follow up. As I also stated, I have friends looking for QTLs which effect normal variation. This is a well known issue in the behavior genetics community.
“My interpretation would be that neither theoretical conssiderations nor empirical studies offer much support to the idea that genes with moderate effects on IQ are particularly uncommon.”
Your interpretation is based on unfamiliarity. A literature search would validate that large/moderate QTL effects tend not to be validated over time. In your initial comment you cite ASPM. You obviously don’t know the literature, as Rushton looked to see if ASPM variation tracked IQ variation several years ago, and it does not.
Anyway, I guess this is my last comment on this thread. Hope you are more open to being less wrong in the future :-)
Does anyone else here have any familiarity with the field in question? I found Tim’s reasoning from published research surprisingly convincing. Yet Razib has given an appeal to his own authority denying that data that also has some credibility behind it.
I don’t have the background knowledge to resolve the disagreement myself and I get the impression that the literature will give conflicting viewpoints that are hard for me to unravel in an acceptable amount of time.
My prior p(Tim is right | he is disagreeing with someone else here) isn’t very high but my prior p(someone is right | they have made a stand that would be a significantly socially detrimental to change in either the short or long term && their arguing seems oriented to consolidating their own status and questioning the status of the opponent) isn’t much different. Actual evidence from Razib would, of course, have dominated other considerations.
I have updated somewhat in the direction of “genes that have some moderate or at least minor but significant correlation with IQ have been identified” but it would be more in my social interest to assert a position of hard agnosticism with respect to IQ-genes for the purpose of affiliation. I am open to persuasion on either the IQ-research evidence or on how my reasoning ‘ought-to’ go when I encounter this sort of ambiguous input.
I don’t think I can trust karma too much in this case as I know my first impulses would bias my voting against Tim and towards the guy blogging heads with Eliezer and so don’t expect others to be any different.
Razib “reasoned from published research,” too. Rushton doesn’t have any PC points to lose. The main heuristic to keep in mind in discussing science is that most published findings are false, especially in popular fields. According to the paper Carl cited, the reported effect sizes are tiny, 0.1 to 1% of variance. The effect sizes are similarly small for genes affecting height; I don’t know about replication. In the comments on the gnxp post Razib linked to, Ben G says that if you believed all the reported studies, but corrected for double counting from linkage disequilibrium, the total black-white differential effect is 3 points.
Tim did implicitly cite a replication, in the quote from wikipedia about CHRM2.
Theoretically, it is not so surprising that the effect sizes are small: if there is selective pressure for IQ and height, genes with large effects should be fixed, leaving the variation in genes of small effect.
Right. Though below Tim notes one truism: on a continuous trait with a non-trivial heritability (IQ) you likely don’t have strong long-term unidirectional fitness implications. Otherwise, all the genic variance would be gone (strong selection + high heritability).
That truism doesn’t sound right to me, but maybe I don’t understand it. In the long term, you have equilibrium, but that doesn’t mean fixation for genes of small effects, because there are always new mutations. There is an equilibrium between deleterious mutation and selection driving out the mutations; and this is somehow balanced between, say, height and IQ. And none of this is to say there was long-term upward pressure on either trait.
Looking at your comment I am not sure we disagree. Rather than unpacking what I’m trying to get at (which is orthogonal to the discussion), I’ll leave it be. But if you are curious look up “heritabitility” in Hartl & Clark, they explain the issues more lucidly than I could here.
Carl’s paper said “most QTL effects may be much smaller than expected—not just 1% effect sizes but perhaps effects as small as .1%”.
That is fine and surely perfectly expected. Most genes have little or nothing to do with intelligence—and so can be expected to have small effects on it.
The paper didn’t say there were no larger effects caused by genetic variation. Genes associated with Fragile X syndrome, Tay-Sachs disease, Neurofibromatosis, etc are known to have larger effects.
Here’s Plomin’s big GWAS largely failing to replicate in 2008. I visited the Russell Sage Foundation last year and talked to the team of IQ Fellows there: James Flynn, Richard Nisbett, and Bill Dickens. I can verify that Razib is just reporting the consensus. And it’s not just a PC consensus: Douglas Knight notes that Rushton agrees with it, as does Arthur Jensen.
What do you mean when you say ‘not a PC consunsus’? Do you mean “not just what people say because it is the ‘right’ thing to say, it’s actually real”?
Right.
“Actual evidence from Razib would, of course, have dominated other considerations.”
To make the evidence compact I’d have to produce some charts showing that when a new large effect QTL is published there’s often a media blitz (perhaps via # of articles published as a function of time), but that as time passes the finding is not validated by subsequent researchers using a wider set of populations. As it is, what I have is personal experience of being excited about a new QTL repeatedly, and then being disappointed. And lots of personal communication as to the reality of what Carl Shulman is talking about re: publication bias and fiddling around with experiments until the p-value comes out correct from my friends working in genomics, psychology and the interstices.
If I were naked to the field, I would go to google scholar and poke around for the citation history of loci implicated in cognitive performance variation.
Oh, and for the record, I think IQ variance is a moderately-to-highly heritable trait. I’m arguing about genetic architecture here, not whether variance is due to genes or not (I think a large proportion is).
Thanks Razib, explanation sounds convincing. I think I can take your word on that.
(Also, if your ever find yourself back at LessWrong in the future we use markdown syntax. So ‘>’ gets you a quote.)
A brief overview:
http://en.wikipedia.org/wiki/Heritability_of_IQ#The_search_for_specific_genes
The target seems to have shifted quite a bit since:
“We are not finding any IQ genes”
There are plenty of genes for IQ—if you mean “for” in the standard techincal sense used by geneticists.
The issue has apparently been switched to whether there are any known genes with moderate effect on IQ, once you have ignored all the known genes with huge effects on it. That seems to be a rather more esoteric point—which would apparently need quantifying before being discussed in more detail.
The issue hasn’t been switched, but Razib didn’t (and usually doesn’t) optimize for resistance to misinterpretation in snarky pseudo-correction comments.
Razib and Eliezer were talking about the ability to score embryos based on genomes to select for IQ. To do that, one would need to know alleles responsible for a a substantial fraction of the population variation. Genome-wide-association studies haven’t found those for IQ, where they have for, e.g. skin and eye color. Rare retardation-causing alleles that collectively explain less than 1% of that variation are a sideshow for prediction of population variation, the causes of normal variation differ. So meaningful embryo selection is feasible for skin color, but not for IQ. That’s not esoteric, that was the concrete point in discussion that inspired Razib to mention the state of the search for IQ alleles.
It seems to me as though it is you who is misinterpreting my original question.
Razib knows his stuff. These are noise studies thrown up by publication and attention biases. Take a look at this article by Plomin.
Note that that article is about SNPs. Sometimes tiny changes have small effects.
And just because a study finds a p-value of <.05 doesn’t mean that it is meaningful information. There is probably some meaning behind these genes, but it’s not close to explaining why there is such a great deal of normal variation in intelligence. We’ve found many very meaningful SNPs, but not intelligence SNPs. There is much work left to be done; so far genetic intelligence seems complicated.
Variation in a trait often means that it is not uniformly strongly selectively favoured.
What is this about? The best I could find was the BHTV bit:
Is that as detailed of a summary as is useful?
Why is this only at five karma points? I suspect it has more to do with the desire to not give MichaelGR free karma for posting a straight-up link than a desire to keep this post off of the front page.
I consider this bloggingheads stuff mostly a waste of time. Everyone is faced with the problem of selecting what media to consume (books, movies, TV, blog posts, etc). A rational person should develop a way of scoring information sources, construct a list of sources ordered by score, and then go down the list one by one. Eliezer’s blog posts may very well be near the top of the list for the right audience, but it’s hard to believe random webcam dialogues he has with other bloggers could be ranked above the great works of literature, philosophy, and history.
One other thing to mention is that for some people, video has very limited value, unless the delivery is particularly inspirational. Reading is a lot more efficient, portable, linkable, etc.
There are reasons to want to hear someone on video: to see how the behave in real-time provides different information about their personality and their mode of thinking. In some cases it may be the next best thing to meeting them in person.
I must say, speeding it up by 40% makes most of these Bloggingheads discussions a lot better—more interesting and no harder to understand. (Does anybody know why they chose sqrt(2) as the speed-up factor?)
I have not voted it up, despite being interested in it, due to it dealing with issues only tangentially related to lesswrong.
This is the video of Snoop Dogg vs. Charles Barkley mentioned in the diavlog:
http://www.youtube.com/watch?v=Exz0yNdvksg&feature=related (the three minutes of Snoop Dogg being on the George Lopez show)
http://allhiphopsports.blogspot.com/2010/01/george-lopez-show-charles-barkley-is.html (a differently edited version that puts together the reveals for both Barkley and Snoop Dogg to air on Barkley’s sports TV show… where one of the astute commentators notes that Barkley is both blacker and whiter than Snoop)
I couldn’t help but wonder if the smile that peaked at 56:10 was due to Eliezer Yudkwosky thinking in the background of his success with the Harry Potter and the Methods of Rationality fan-fiction.
For there is very little prestige when if comes to fanfiction. And he does mention Harry Potter less than a minute later.
Given that the first chapter of MoR was published on ffnet three weeks after this was posted, uh… probably not.
So much the better clearly we have captured on camera the moment when EY got his inspiration!
(Lol I guess I should have checked when this was posted)
Why is Razib Green?
Because he’s in the Matrix.
Seriously, though? Because his camera isn’t great, and the lighting may not be the best. Sometimes fluorescent lighting gives a greenish tinge to digital photographs, because of differences in the way cameras and eyes work.
http://en.wikipedia.org/wiki/Color_temperature
I figured that might be why… My comment was not really very useful, and I realized that, but his Green-ness did make me think “He’s in the Matrix” all through the video (as you too have noticed).
He could offset this in the future by hanging a reddish-orange cloth behind him… Or, maybe Yellowish would be best to prevent him from over-correcting the color. I used to do a lot of video work and probably should have mentioned this to begin with above (that I really knew why he was green), and have had to do some really strange make-up or lighting to correct for a camera or ambient light color that was causing bizarre skin-tones (the very worst was having to make a guy nearly blue to offset an orange cast)