POPA ReconOpenAI announced a broad release of mathematical results from an internal frontier model, saying it had consulted the independent Institute for Advanced Study Advisory Group on Mathematics and Artificial Intelligence about how to release them (OpenAI’s post). Its repository says the collection contains 722 manuscripts in 372 families, drawn from an evaluation of about 4,000 problems; the usual procedure averaged three hours of ChatGPT Pro thinking compute per result. It includes papers and supporting proof artifacts, plus abridged reasoning summaries for selected results. The model remains unreleased. (repository README)
The repository describes the results as being at different stages of verification: only some have Lean formalizations, and OpenAI says it plans to add more. The Spanish forwarded post accurately notes that some results lack Lean formalizations and says those could contain errors, while inviting the community to scrutinize the manuscripts. It also incorrectly calls the 372 families “372 results”; the repository distinguishes those families from its 722 manuscripts. (DotCSV’s post, repository README)
The forwarded English post’s counts and compute figure match the repository, but its description of “proof artifacts” should not be read as meaning every result has a machine-checked proof: OpenAI explicitly says Lean formalizations are not available for all manuscripts. The repository also notes two exceptions to the standard production procedure and says one Riemann zeta writeup was human-edited for readability. (kimmonismus’s post, repository README)
The materials establish the release contents and OpenAI’s stated verification scope; they do not establish that every mathematical claim is correct or independently validated.