logoalt Hacker News

alethic • today at 8:44 PM • 3 replies • view on HN

Unfortunately, it is not. The same level of slop nuke occurred with Anthropic's auto-formalization of Fermat's Last Theorem (~13 million lines of code). What Anthropic's FLT did do was undercut the human-led efforts to get idiomatic formalization of the requisite mathematics surrounding Fermat's Last Theorem into Mathlib.

This result is not quite as harmful as Anthropic's FLT was because it is indeed telling us something new -- that the Riemann Hypothesis is true, with full confidence -- but it is not reusable in any sort of way, beyond proof mining to figure out what in the world the AI model did. It is still in fact harmful to the greater project of mathematics formalization (irrelevant to foundations), because it undercuts the ability to fund such efforts for reusable proofs.


Replies

emtel • today at 8:56 PM

You cannot harm someone or a community by publishing a proof. If the math community stops doing work they were previously doing, that’s their choice.

alethic • today at 8:56 PM

One indeed might wonder whether the "human-led" efforts of Mathlib are worthwhile. First, it must be noted that Mathlib serves two purposes -- to be a reusable and robust collection of formally-verified mathematical statements, but also to be a learning resource for students of mathematics without access to the rigour that an academic education provides. It is for this reason that the human nature of mathematical exposition is deeply important. In mathematics, a proof is more than a blackbox -- it is (or was) the way in which you explain mathematics.

(I consider "proof" somewhat unfortunate terminology for these slop nukes, as such. I think the term "certificate" is better. 17 million lines of Lean does not pragmatically convey anything other than whether a statement is true or false. It may be easier communication, but it is deeply, deeply, inefficient communication.)

The first goal of Mathlib is indeed not in conflict with slop nukes, however. One of the projects of the Lean FRO (or just a Lean FRO member? unclear to me how endorsed it is) is a version of Mathlib that fully embraces slop: https://taucetiproject.org/. However, there's really no bottleneck there. If there is any non-novel mathematics in Anthropic's proof of the Riemann Hypothesis that is not already in Tau Ceti, it is simply because they have not bothered yet. Auto-formalization works, at scale, very well, as harmful to the human as it might be.

baq • today at 8:57 PM

Everything changed this fall and what is being funded must change alongside it. Explaining, simplifying and understanding these slop roofs may be very well more valuable than trying to prove the conjectures.