9 Comments
User's avatar
Matt Peterson's avatar

You don’t touch on the concurrent announcement by Alessandro Masullo that he had also solved (or come close to solving) the RH in a much more elegant non-AI aided proof. If true, it might lay bare the difference between the absurd brute force power of AI vs. the broader knowledge-unlocking approach of human mathematicians.

Tonecolor's avatar

Feels like this is not good. The mathematics we could’ve invented to solve it may yet show new and more beautiful, useful problems about reality. But maybe that will still continue.

Peter van Beek's avatar

Mathematicians might become more like priests to a higher power. Only they will be able to decipher the messages that came from above. It might not be a bad job at all.

Kenny Easwaran's avatar

You say Hilbert would be rolling in his grave - but Hilbert was actually specifically looking for this! He hoped that math would actually be completely mechanically decidable. He said “in math there is no ignorabimus” - no “we will never know”, because he thought the axioms should solve everything. His tenth problem in his list was to find the algorithm that would solve every polynomial equation in integers.

It took Gödel to show that Hilbert was wrong here - there are undecidable questions in math, and (after a few decades of work by Davis, Putnam, Robinson, and Matiyasevitch) it turns out there is no algorithm that determines which polynomial equations have integer solutions.

In retrospect, mathematicians breathed a sigh of relief that Hilbert was wrong. What’s disturbing about AI in math is that it gets closer to Hilbert’s vision than anyone thought possible in the wake of Gödel and Turing’s impossibility results.

Billy Youngblood's avatar

But has it? It’s 250,000 page proof that no one has checked. Let’s not get ahead of ourselves here

Alberto Romero's avatar

Yes it has. The proof is not 250k pages. It's in Lean and it has been independently verified already. What you say is Alpoge's attempt at having Claude write an English document. Different things. The proof is correct but we don't understand it.

Billy Youngblood's avatar

If we don’t understand it is it useful? Riemann’s hypothesis is widely suspected to be correct, so what does this add to human knowledge?

Billy Youngblood's avatar

I guess my follow on question here is whether Lean be spoofed or finessed? Given LLMs have a history of gaming outcomes how can we be sure they haven’t figured out how to get a spurious verification out of Lean if humans can’t understand what it’s doing?

Kenny Easwaran's avatar

Something like this did happen a few months ago, with an apparent disproof of the Collatz conjecture that turned out to be a discovery of a bug in Lean: https://leodemoura.github.io/blog/2026-8-1-postmortem-for-kernel-soundness-bug-14576/

I don’t know what the status is yet of any provably correct theorem provers. (I believe some exist, but they may be harder to work with than Lean.)