Good recent update (and original article, very comprehensive), though I'd argue you haven't quite weighted the recent evidence for AI R&D acceleration/AI uplift hard enough. But I'd be very happy giving this to someone to read for a comprehensive introduction to the state of AI, relevant arguments, and forward projections. Perhaps with recommendations to read a couple different articles than the ones you linked if they want solid bear cases, assuming I could find any high quality ones that are up to date of course!
There were also some minor things that got lost like footnotes de-syncing at some point so that e.g. inline 31 corresponds to footnote 32, and inline 40 corresponds to footnote 42.
Agree, my original aim was to cover RSI in a separate article (my feedback loops one), but at this point RSI is such an important reason for very short timelines that it would be ideal to make it a section within this article.
> This paradigm also largely solves the data bottleneck that might have halted the scale-up of pretraining. You can ask an AI to solve 100,000 math problems, then take only the cases where it got the right answer, and use them to train the next model. Because the solutions can be quickly verified, this isn’t circular. Rather, you’ve generated more examples of genuinely good reasoning.
Doesn’t this just guarantee that the answer is right, but not the whole reasoning chain? The models could get the correct answer with faulty reasoning.
I’d guess strictly speaking yes, though often it’s possible to verify steps in the reasoning chain too. And if the answer is right, that’s also evidence that the reasoning is right.
Thanks for that writeup, that's really a lot of useful information.
Incredible work Ben! Supervaluable! 🚀🚀
Good recent update (and original article, very comprehensive), though I'd argue you haven't quite weighted the recent evidence for AI R&D acceleration/AI uplift hard enough. But I'd be very happy giving this to someone to read for a comprehensive introduction to the state of AI, relevant arguments, and forward projections. Perhaps with recommendations to read a couple different articles than the ones you linked if they want solid bear cases, assuming I could find any high quality ones that are up to date of course!
There were also some minor things that got lost like footnotes de-syncing at some point so that e.g. inline 31 corresponds to footnote 32, and inline 40 corresponds to footnote 42.
Agree, my original aim was to cover RSI in a separate article (my feedback loops one), but at this point RSI is such an important reason for very short timelines that it would be ideal to make it a section within this article.
> This paradigm also largely solves the data bottleneck that might have halted the scale-up of pretraining. You can ask an AI to solve 100,000 math problems, then take only the cases where it got the right answer, and use them to train the next model. Because the solutions can be quickly verified, this isn’t circular. Rather, you’ve generated more examples of genuinely good reasoning.
Doesn’t this just guarantee that the answer is right, but not the whole reasoning chain? The models could get the correct answer with faulty reasoning.
I’d guess strictly speaking yes, though often it’s possible to verify steps in the reasoning chain too. And if the answer is right, that’s also evidence that the reasoning is right.
Given the recent math proofs, can we say we have AGI now?
Impressive
The example question for quantum physics itself is incorrect.
Can you explain?
In the Kraus operator question, It asks to pick the correct representation.
The correct answer should be $$\left( 1- \frac{3}{4}p \right)\rho+p\frac{X\rho X}{4}+p\frac{Y\rho Y}{4}+p\frac{Z\rho Z}{4}$$
That's not one of the options, it seems option C had a typo.
You can ask Claude to confirm this.
What the question is asking you to do is essentially take the list of Kraus Operators, A_1, A_2 , A_3 and A_4
and write their action on a density matrix rho. Their action is given by
$A_1 \rho A_1 ^\dagger+A_2 \rho A_2 ^\dagger+A_3 \rho A_3 ^\dagger+A_4 \rho A_4 ^\dagger$
Terrifying stuff
with an article title like that…
now that’s how you bait the EAs and the e/accs into reading. Describe and prescribe. Well played, sir. Well played.