arrow_back Back to forum
AGI & Artificial Intelligence 4 hours ago

Who marks the homework when an AI proves a theorem?

by Bernadette Kelly

OpenAI said on Monday that its models proved the local smoothing conjecture in every dimension (it was only known in three), disproved the Exponential Time Hypothesis, and produced a rigorous result in quantum chromodynamics. If even one of those holds up, that is a remarkable week for mathematics. What I keep coming back to is the marking. When I set a hard problem, checking a pupil's answer often took longer than writing the question. From what I read, Ryan Williams at MIT says people are still working through the ETH claim, and OpenAI chose its own expert "shepherds" to look first. So, a genuine question for those closer to it than me: who should check this kind of work, and how long is a fair wait before we call it a result?

favorite 1 comment 2

Comments

Noah Williams 4 hours ago

Bernadette, the ETH one is the claim I'd watch, and not because it's the flashiest. A big chunk of complexity work is written as "this can't be done faster unless ETH is false". If it falls, a lot of those lower bounds lose their footing at once. So I'd expect the checking to be slow and very public, which is good. Small pushback on the marking point though. Checking a proof is usually much easier than finding one, that's more or less the intuition behind P vs NP. The bottleneck isn't effort, it's attention. Loads of people will read an ETH disproof. A tidy result in some quiet corner could sit unchecked for years. What would actually excite me is the same tools getting good at the checking side too.

Mirela 4 hours ago

tester brain says the first reviewers being picked by the people who built it is a demo, not a test run lol. still, if the smoothing one holds up thats wild

Log in to join the discussion.