OpenAI Releases 722 Mathematical Manuscripts and Details Its Method

OpenAI has published 722 manuscripts produced by an AI that is not yet available and is framing their dissemination with safeguards for transparency and verification. The company claims to rely on an independent committee, verifiable evidence in Lean, and data on the computational resources used. At the same time, a professor from Rutgers considers one of the results worthy of a Fields Medal if it had been produced by a human.

Verifiable Evidence and an Advisory Committee Announced

OpenAI has released 722 mathematical manuscripts on a platform that allows for citation management and revisions. Some proofs can be verified using the Lean programming language. The company also provides additional information in the repository about how the results were obtained, including 10 summaries of the model's reasoning, estimates of the computational power used in terms of ChatGPT Pro usage, and statistics on the number of problems addressed. OpenAI specifies that it relies on an independent advisory committee to define best practices.

Evaluation by Open Problems and Average Time Announced

OpenAI generated 722 manuscripts, divided into 372 groups, from an AI model that has not yet been published. According to OpenAI, solving each problem required an average of three hours of "ChatGPT Pro thinking." The company explains that it has adopted this new evaluation method based on open mathematical problems because its previous approaches had become outdated.

An Unreleased Model and Reactions Up to the Fields Medal

The AI model used to produce these results is not yet available. The scientific community is examining the documents posted on GitHub, and reactions are emerging on X. Alex Kontorovich, a professor at Rutgers University, believes that a reported result would be worthy of a Fields Medal if it had been obtained by a human.