OpenAI publishes 722 AI-generated math preprints on GitHub

OpenAI posted 722 mathematical manuscripts to GitHub on October 6, organized into 372 families and produced by an unreleased internal frontier model that was pointed at roughly 4,000 open problems. The preprints directory sits alongside a Lean proof library in a repository released under Apache-2.0.

"The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model," the project states. Each accepted result averaged about three hours of ChatGPT Pro thinking compute, and ten manuscripts ship with abridged reasoning summaries covering areas like the Mahler conjectures, NP-hardness, spin glasses, and the Vlasov-Maxwell system.

Verification is explicitly uneven. The repository warns that "Not all have accompanying Lean formalizations" and that "Some of the unformalized results could have issues." Lean, the proof-assistant language used here, lets a computer check a proof step by step; the unformalized papers are PDFs that still need human referees.

Two researchers we follow were circulating the link within a day, which fits a release OpenAI appears to have staged with mathematicians rather than at them. According to Unite.ai, OpenAI consulted the Institute for Advanced Study's Advisory Group on Mathematics and Artificial Intelligence, which published responsible-release recommendations on September 29 after collecting more than 600 responses from the mathematical community.

The model itself remains unreleased. Interesting Engineering reported the collection spans pure mathematics, theoretical computer science, and mathematical physics, with specific projects on topics such as the irrationality exponent of pi.