OpenAI has published 377 new mathematics results produced by an unreleased AI model. The release has rattled the maths world.

The company posted 722 manuscripts, grouped into 372 families, on GitHub this week. It says an internal model analysed about 4,000 problems. Each result took computing power equal to roughly three hours of ChatGPT Pro thinking, on average.

One result claims a proof of the full Birch-Swinnerton-Dyer leading term formula. Mathematician Alex Kontorovich called one finding “insane,” saying a human achieving it would win a Fields Medal at once.

OpenAI shared reasoning summaries for 10 results. Some proofs are written in Lean, a language that lets computers check proofs. But the company has not released the prompts or the full problem list.

Sceptics want receipts. An advisory group at the Institute for Advanced Study asked AI labs on Sept. 29 to stop testing advanced maths on secret models and to publish their methods. OpenAI says it is taking the advice seriously but is not bound by it.

The results now face the slow work of peer review. That will decide whether they hold up.