People keep saying AI cannot do real mathematics. That argument is getting harder to maintain.
A system has just solved an 80 year old problem that had left top mathematicians stuck, prompting genuine surprise from the field.
Either way, the boundary between human and machine problem solving is starting to look a lot less solid than it used to.
from what I can understand, an AI created a counterexample/disproof; it seemed like that nature of the disproof was such that a LLM is well-suited to solving, by throwing computing power at recombining fragments of already existing proofs and mathematical knowledge that was already in the model.
Its an advance, but it may not be scalable in the sense that computation power is... resource bound to put it mildly.
OpenAI obviously framed it in a way that is maximally beneficial for them.
Below is a summary from a number of mathematicians (not necessarily in the field of the conjecture at hand).
https://arxiv.org/html/2605.20695v1
If the premise was that AI-assisted proofs are now viable, I think that might be closer to an absolute truth if there is such a thing.
Mojo ♻️
crouton