OpenAI Releases 722 Math Manuscripts From Unreleased Model

OpenAI released 722 mathematics manuscripts produced by an unreleased internal frontier model, according to The Verge. The company says the collection covers 372 result families and contains solutions to hundreds of open questions.

OpenAI published the manuscripts in a public GitHub repository with source files, citation instructions and 10 abridged summaries of the model's reasoning. The model attempted about 4,000 problems, and OpenAI estimates that an average published result used compute equivalent to roughly three hours of ChatGPT Pro thinking.

Many manuscripts include proofs formalized in Lean, which allows computers to check mathematical proofs, but some results remain unformalized. OpenAI warns that those results may contain errors and says it will record corrections while preserving earlier versions.

The Advisory Group on Mathematics and Artificial Intelligence recommends releasing results through independent scholarly repositories, disclosing model and compute details, and formalizing proofs where possible. OpenAI says it is considering community hosted alternatives and plans to improve citations and mathematical explanations in future releases.