Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Interview with gwern

dwarkeshpatel.com

Post reply on HN
Print view View original

241–250 of 262 posts

Re: Interview with gwern

#241

Post by A_D_E_P_T »

tptacek wrote: post #222
A_D_E_P_T wrote: post #220

Earlier quoted context omitted.

Sure, heritability is orthogonal to the question of whether a trait is determined by genetics. But traits like IQ, height, and eye color are both (A) highly heritable and (B) substantially shaped by genetic factors. In casual online discourse, I believe that (B) is usually taken for granted, so it's glossed over, and when people say that any given trait is "heritable" they're also assuming that (B) is true for the tr…

When you say "substantially shaped by genetic factors", you should present evidence. It's easy to provide evidence for the heritability of intelligence (again, stipulating IQ), but as we've established, that begs the question of whether the genetic connection is correlation or causation. Environments are inherited, too. There is growing evidence that group IQ heritability isn't evidence of genetic causation.

There's this: https://pmc.ncbi.nlm.nih.gov/articles/PMC5985927/

First sentence of the conclusion: "Genetic association studies have confirmed a century of quantitative genetic research showing that inherited DNA differences are responsible for substantial individual differences in intelligence test scores."

Related: https://www.pnas.org/doi/full/10.1073/pnas.1408777111

The trouble is that, like height, IQ is governed by a vast network of "genes of small effect," so a comprehensive view has proven difficult to nail down. Progress is apparently being made, though slowly.

> There is growing evidence that group IQ heritability isn't evidence of genetic causation.

What evidence do you speak of?

Re: Interview with gwern

#242

Post by tptacek »

A_D_E_P_T wrote: post #241
tptacek wrote: post #222

Earlier quoted context omitted.

When you say "substantially shaped by genetic factors", you should present evidence. It's easy to provide evidence for the heritability of intelligence (again, stipulating IQ), but as we've established, that begs the question of whether the genetic connection is correlation or causation. Environments are inherited, too. There is growing evidence that group IQ heritability isn't evidence of genetic causation.

There's this: https://pmc.ncbi.nlm.nih.gov/articles/PMC5985927/ First sentence of the conclusion: "Genetic association studies have confirmed a century of quantitative genetic research showing that inherited DNA differences are responsible for substantial individual differences in intelligence test scores." Related: https://www.pnas.org/doi/full/10.1073/pnas.1408777111 The trouble is that, like height, IQ is governed…

The Plomin and von Stumm study --- easy to find critiques of it! --- is about heritability. It cited unpublished GWAS data to claim polygenic scores would predict more than 10% of educational attainment, and thus intelligence. The educational attainment GWAS study that was eventually published found less than 5%. And remember, the twin study heritability data that most people cite in threads like this claimed over 50% (twin study heritability numbers for all sorts of traits are apparently overstated by factors of 2-4x, which is wild).†

An example of evidence against the reliability of educational attainment and intelligence heritability statistics: comparing intra-family heritability (across large numbers of families) to population-wide studies: for educational attainment, it turns out there's little correlation between the two; for simpler phenotypical traits, there's almost 100% correlation.

To sum this up:

1. The 2018 Plomin study gives sharply lower genetic/EA numbers than were floating around previously (say, from the Jensen-ist era)

2. Plomin's own numbers were preliminary and overstated

3. Researchers in the field criticized that study nonetheless

4. Subsequent studies on direct heritability and molecular heritability put even lower ceilings on it (basically, all credible behavioral trait heritability work has been done after 2018 --- and in fact this is broadly true of a lot of genetics work, not just trying to statistically mine behavioral traits out of genome scoring)

5. Even those results have flunked basic sanity tests (for instance, getting wildly different results in intra-family vs population-wide studies).

It's not looking good for people fixated on this idea.

† I'm being very loosy-goosey with the numbers and units here

Re: Interview with gwern

#243

Post by pontsprit »

mlsu wrote: post #240
motohagiography wrote: post #148

Earlier quoted context omitted.

I've been writing for decades with the belief I was training a future AI and used to say that the Turing test wasn't mysterious at all because it was a solved problem in economics in the form of an indifference curve that showed where peoeple cared whether or not they were dealing with a person or a machine. the argument against AI taking over is we organize around symbols and narratives and are hypersensitive to wan…

This is an interesting comment. Let's extend it further. The pop industry is a: - machine - which takes authentic human meaning - and produces essentially a stochastic echo of it ("slop") - in an optimization algorithm - to predict the next most profitable song (the song that is most "likely") So, this sounds an awful lot like something else that's very in vogue right now. Only it was invented in 1950 or 1960, not in…

yes but the pop industry is just one facet of music as a transcendental cultural/spiritual/secular meaning-making expression activity as a whole. its a place where both bob dylan ans taylor swift can exist on the same plane, the implications of such cannot and would never be able to be reduced to data. "if the only tool you have..." not to undermine the implications of complex automation as symptomatic of the current epoch, but I would argue that art offers a materially grounded experience irreducible to data. it doesnt mean that the affects of art wont be harnessed by corporate interests in ever accentuated ways, it seems like an obvious and natural development; on one level its simply the cyclical production of crappification occurring in tighter spirals of output/dynamism, weeee!

Re: Interview with gwern

#244

Post by maujim »

nutanc wrote: post #230
maujim wrote: post #198

Earlier quoted context omitted.

What did you use to create the chunks?

It's a new approach I am experimenting with. https://gpt3experiments.substack.com/p/a-new-chunking-approa...

Interesting approach. Have you tried using existing clustering algorithms to determine the chunks instead of using distance or slope as a measure?

Re: Interview with gwern

#245

Post by CamperBob2 »

keiferski wrote: post #232
CamperBob2 wrote: post #221

Earlier quoted context omitted.

Having an obviously AI-generated header image on your blog post was cool two years ago, but now it is passé and marks you as behind the trends. This is true only because publicly-accessible models have been severely nerfed (out of sheer panic, one assumes), making their output immediately recognizable and instantly clichéd. Dall-E 2, for instance, was much better when it first came out, compared to the current incarn…

My point is that having work that is obviously AI-generated is now a negative thing. If, in the future, there is a way to validate humanity (as I mentioned in my comment), then any real writers will likely use it. Anyone that doesn't validate their humanity will be assumed to be an AI. The reaction to this may or may not be negative, but the broader point is that in this scenario, the AI won't be eating all human cre…

And my point is that if you can't tell the difference -- and you won't be able to -- the question of validating one's humanity will be moot.

Any technology capable of distinguishing between machine- and human-created text or artwork will simply be placed in a feedback loop and used for training the next generation of models.

If this indeed turns out to be a fight, humans won't win it. So let's not bother with the fighting part.

Re: Interview with gwern

#246

Post by mlsu »

pontsprit wrote: post #243
mlsu wrote: post #240

Earlier quoted context omitted.

This is an interesting comment. Let's extend it further. The pop industry is a: - machine - which takes authentic human meaning - and produces essentially a stochastic echo of it ("slop") - in an optimization algorithm - to predict the next most profitable song (the song that is most "likely") So, this sounds an awful lot like something else that's very in vogue right now. Only it was invented in 1950 or 1960, not in…

yes but the pop industry is just one facet of music as a transcendental cultural/spiritual/secular meaning-making expression activity as a whole. its a place where both bob dylan ans taylor swift can exist on the same plane, the implications of such cannot and would never be able to be reduced to data. "if the only tool you have..." not to undermine the implications of complex automation as symptomatic of the current…

Yeah, I think we agree here. I thought your comment -- what is "slop" vs what is "not slop" got to what is "ai" and what is "not ai" in an interesting way.

Re: Interview with gwern

#247

Post by A_D_E_P_T »

tptacek wrote: post #242
A_D_E_P_T wrote: post #241

Earlier quoted context omitted.

There's this: https://pmc.ncbi.nlm.nih.gov/articles/PMC5985927/ First sentence of the conclusion: "Genetic association studies have confirmed a century of quantitative genetic research showing that inherited DNA differences are responsible for substantial individual differences in intelligence test scores." Related: https://www.pnas.org/doi/full/10.1073/pnas.1408777111 The trouble is that, like height, IQ is governed…

The Plomin and von Stumm study --- easy to find critiques of it! --- is about heritability. It cited unpublished GWAS data to claim polygenic scores would predict more than 10% of educational attainment, and thus intelligence. The educational attainment GWAS study that was eventually published found less than 5%. And remember, the twin study heritability data that most people cite in threads like this claimed over 50…

I don't doubt that there are methodological and empirical problems all around. The scientific literature on these points is a mess.

We might want to look at the fundamentals: How is IQ qualitatively different from height, eye color, schizophrenia -- or any other highly complex, heritable polygenic trait? (One could also extend this to the many traits that animal breeders keep an eye on.) None of them have been fully pinned-down yet, but I don't believe that they can't be fully described in principle.

It's true that GWAS for intelligence explains <5% of variance today, but GWAS for height was in the same position a decade ago. Today polygenic scores for height predict over 40% of variance.

Re: Interview with gwern

#248

Post by tptacek »

A_D_E_P_T wrote: post #247
tptacek wrote: post #242

Earlier quoted context omitted.

The Plomin and von Stumm study --- easy to find critiques of it! --- is about heritability. It cited unpublished GWAS data to claim polygenic scores would predict more than 10% of educational attainment, and thus intelligence. The educational attainment GWAS study that was eventually published found less than 5%. And remember, the twin study heritability data that most people cite in threads like this claimed over 50…

I don't doubt that there are methodological and empirical problems all around. The scientific literature on these points is a mess. We might want to look at the fundamentals: How is IQ qualitatively different from height, eye color, schizophrenia -- or any other highly complex, heritable polygenic trait? (One could also extend this to the many traits that animal breeders keep an eye on.) None of them have been fully…

One obvious difference between what is known about the heritability of height and what is known about the heritability of intelligence or educational achievement is that when you put height heritability to the test, for instance by checking to see if the intra-family studies agree with the population-level studies, the height heritability stuff holds up, and the EA/IQ stuff does not. Another might be that successive rounds of study of simple phenotypical traits like height have not demolished previous estimates of heritability, while that has in fact happened in the EA/IQ case.

All this happens before we even reach questions about test-test reliability of IQ, or of whether gene-environmental interactions are uniform between Europe and other population cohorts (they do not appear to be!). It defines away SES confounding (which appears to be a significant issue). It has thus far largely ignored epigenetics. And, of course, for it to mean anything, the hypothesis also has to defend the idea that IQ/EA, at least in its genetic component, is immutable.

All that aside, I'm mostly just here to say that simple heritability statistics don't say what people on HN seem to think they say.

Re: Interview with gwern

#249

Post by throw05678931 »

Vecr wrote: post #54
throw05678931 wrote: post #22

Lots of the claims about Gwern in the intro are exaggerated. Gwern is an effective altruist and his influence is largely limited to that community. It would be an exaggeration to claim that he influenced the mainstream of AI and ML researchers -- certainly Hinton, LeCun, Ng, Bengio didn't need him to do their work. He influences the AI safety crowd, who have ironically been trying to build AGI to test their AI safety…

"Wordcel" is roon terminology, right? I highly doubt Aschenbrenner is an EA, and if he's a "wordcel" he somehow managed to do mathematical economics without too much problem. Gwern's probably not a "wordcel" either, he can program, right? I've never seen any of his publications though. It's called Situational Awareness too, not "Situation Report", and Yudkowsky said he didn't like it. Not that Yudkowsky is an EA eith…

I didn't claim Gwern was a wordcel. He's clearly not. Dwarkesh, yes.

I didn't claim Yudkowsky was an EA. He's not. He's a rationalist who originated many thoughts about AI doom, which have influenced EAs.

Look, Gwern and Dwarkesh and Yud are all AI doomers, which is where the rationalists and the EAs overlap. Gwern is all over LessWrong, which kind of springs out of the rationalist community and has all sorts of EAs there now. But of course, the first rule of EA is the first rule of Fight Club. That's why Dario consistently denies he's an EA, even though his chief of staff was formerly running Sam Bankman-Fried's philanthropic grants.

The weird thing in these conversations between guys like Dwarkesh and Gwern is how much they wallow in the masochism of imagining a world taken over by AGI. Yud has a similar attraction to pain, but he channels it into his silly BDSM novels.

https://www.amazon.com/Dark-Lords-Answer-Eliezer-Yudkowsky-e...

If that's how they wan't to spend their time, fine. Let's all just recognize it's an acquired taste, a self-indulgence, and the furthest thing from a rational act.

Re: Interview with gwern

#250

Post by throw05678931 »

wyager wrote: post #129
throw05678931 wrote: post #22

Lots of the claims about Gwern in the intro are exaggerated. Gwern is an effective altruist and his influence is largely limited to that community. It would be an exaggeration to claim that he influenced the mainstream of AI and ML researchers -- certainly Hinton, LeCun, Ng, Bengio didn't need him to do their work. He influences the AI safety crowd, who have ironically been trying to build AGI to test their AI safety…

> Anthropic has moved so much slower than its competitors In what sense? Claude is way better than any other LLM available, in the sense of providing useful/correct/intelligent output.

o1 does everything I need.
Post reply on HN