OpenAI's 372 math results explained: what is claimed so far

A well-known computer scientist says OpenAI's latest batch of maths results includes a proof of a famous open problem. The claim is so far unchecked by outside experts, so treat it as a report, not a settled result.

Scott Aaronson, a computer scientist who writes the blog Shtetl-Optimized, wrote on 8 October that OpenAI's release of 372 maths results includes a proof of the Unique Games Conjecture. A conjecture is a claim that mathematicians believe is true but have not yet proved.

Aaronson says the proof is "Lean-verified". Lean is a proof-checking program: it tests a proof step by step, so a person does not have to trust that every step is right.

He also notes a personal link. His wife, the complexity theorist Dana Moshkovitz, worked toward this result for her career.

A separate report from AI Weekly, published on 7 October, says OpenAI states that most of the 372 results came from one prompt. A prompt is the instruction you type to an AI system. The notes do not say what that prompt was.

Four experts on our tracking list shared Aaronson's post within hours of it going up.

We worked from summaries of both pages, not the full texts, and we have no independent expert verification of the proof.

Nothing here changes what you can do with an AI tool today. The reports do not mention a new product, a price change or a new feature.

What the story does show is a claim about how AI is being used in research. OpenAI says one prompt produced most of a large set of results. If the proof holds up, it would be a notable result in complexity theory, the study of how hard problems are for computers to solve. That "if" matters, because the checking is still to come.

Outside mathematicians are expected to check both the proof and the Lean formalisation, which is the version written so that Lean can verify it. Until they report back, the claim rests on Aaronson's account and OpenAI's own statements.