openai.com
|
ksl
|
|
OpenAI submitted AI-generated proofs to the First Proof challenge, a set of ten unsolved lemmas written by eleven mathematicians across subfields from algebraic topology to symplectic geometry. The company’s chief scientist Jakub Pachocki said they believed six of their ten solutions had a high chance of being correct, but independent review found only two actually held up – problems nine and ten. Each problem had taken its author weeks to months to solve by hand. The gap between confident output and verified correctness is the recurring theme every time AI models meet research-grade math, and it showed up here cleanly. OpenAI published all ten attempts along with prompt patterns and a new appendix detailing how human mathematicians guided the models through back-and-forth refinement.
