Post

Colin Hoad
@colinhoad.com

I think the wider problem is that even if some of these proofs are correct, nobody has any understanding of how they've been arrived at (OpenAI aren't sharing the model they used) so the *utility* of any of them is next to none. I'm no mathematician, but failing to show your working is never good.

2 reposts30 likes

Colin Hoad
@colinhoad.com

It's a bit like Fermat's famous margin scribble, that he had a proof for his conjecture but no space to share it in. It took several hundred years before Andrew Wiles was able to provide one!

muxauy
@lanzz.org

the proof itself is the mathematical apparatus used to arrive at the final result (confirming or refuting a conjecture). it should be mathematically sound and reviewable, otherwise nobody has proven anything. the LLM model itself would be irrelevant to the proof's validity or utility

merman1974
@merman1974.bsky.social

Mathematically, you should always show your working. This is the issue with the AI “solutions” - either we cannot see how they worked out key steps, or the AI is using an unproven or unknown method at certain points that the mathematicians say may not work

muxauy
@lanzz.org

well if they're skipping steps or using other unproven mechanisms, then these aren't proofs. no amount of information known about the models themselves would change that