I think it’s weird that the Riemann hypothesis proof is 250,000 pages. Partly just because this is far outside the human norm, but also because—judging just from the length—it sure sounds like the sort of output brute force-ish computations tend to generate? Of course the proof was not literally brute forced, not found via literally exhaustive search, but I would sure guess given its length that the proof is far from minimal.[1]
I am not sure what to make of this; it feels strange to me, in the rabbit/duck sort of way many AI results have felt strange to me lately. On the one hand, it is of course terrifying that AI has become capable of proving the Riemann hypothesis at all, in any way. And even if the proof was found in a “brute force-ish” sort of way, I am not sure how much comfort to take in that, since I am not sure how far “brute force-ish” knobs can be twisted. I know of no particular reason to think they can be twisted far enough to prove the Riemann hypothesis, but not far enough to foom.
But it does still seem weird, though? Historically, many humans have said things like “Occam simplicity is a core desiderata of scientific hypotheses,” and “intelligence ≈ compression,”[2] and “intelligence probably has something to do with inference efficiency,” and similar such things. And sure, perhaps humans have been somewhat overrating this virtue, relative to the amount AI should, since our minds cannot search as exhaustively. But personally, it seems unlikely that we have been radically overrating it. And so to the extent AI currently has less of this virtue than one might expect a human capable of generating similar results to have, then I do think it makes sense to wonder (among other things) whether AI insights and abilities might tend to generalize less far than one’s priors about human minds would predict.
That is, I would guess there is e.g. large algorithmic information distance between this proof and the shortest possible proof. This might be wrong—e.g., if the final proof is a more normal length than (as Alpöge put it) the “massive unreadable travesty of ‘slop'” Claude generated—in which case I'd consider my argument here wrong/irrelevant.
I grant I am somewhat conflating the Occam simplicity of the output vs. of the generating computation here. These can diverge, but they are related; all else equal, one should expect less Occam simple outputs to have been generated by less Occam simple computations, since all else equal it tends to take more search to find less simple models.
Well, which is more likely to happen first?
We should predict the latter for most tasks at the frontier. AI has the ability to search very broadly and deeply, and this ability is amplified with access to significant computational resources. To get the clean version first, AI would have to not attempt the problem until it is far past it the degree of capability at which ... (read more)