OpenAI publishes 722 mathematical works from its new model on GitHub

OpenAI has published a catalog of 722 mathematical works and proofs generated by its new, unannounced artificial intelligence model on GitHub. The materials are combined into 372 thematic families. The research was conducted based on about 4,000 open problems after the previous benchmarks were exhausted. Generating most of the solutions required an average of about three hours of computation in deep reasoning mode. The catalog includes advancements on the Hodge conjecture for abelian varieties, the zero-free region of the Riemann zeta function, and the Hott conjecture. Part of the proofs are formalized in the Lean verification language, and for ten tasks, the company revealed the model's abbreviated line of reasoning.

OpenAI warned that informal works may contain errors and promised to publish corrections while preserving versions, emphasizing that the publication cannot be considered automatic confirmation of the solution to all problems. The release was prepared with the participation of an advisory group at the Institute for Advanced Study (IAS).

The publication has caused controversy in the scientific community. As 3DNews reports, citing Wired, in August, at a meeting with OpenAI representatives, about 40 mathematicians recommended that the results be formalized in the form of full-fledged academic articles with a description of methods, rather than being limited to repositories and social networks. The company stated its intention to fund a series of conferences and seminars to study the data and also provide access to the model itself to researchers after completing security checks.