OpenAI Publishes 722 Manuscripts From Unreleased AI Model

OpenAI has published 722 mathematical manuscripts produced by an internal artificial intelligence model that has not been released publicly, describing the collection as 372 families of results drawn from work on about 4,000 problems. The papers are available in a public GitHub repository, but their accuracy and novelty have yet to be independently established. The company says each result required about three hours of ChatGPT Pro reasoning compute on average. That estimate concerns computing time and does not establish how much human input contributed to individual results. OpenAI has also posted ten summaries of the model’s reasoning and information about the work’s preparation and citation practices. Many of the proofs have formal versions written in Lean, a programming language that allows computers to check proofs against formal rules. OpenAI acknowledges that some results do not have formal versions and could contain errors. The company says the repository will preserve earlier paper versions and provide citation information for each manuscript. The release does not make the model available for researchers to run.