OpenAI has published 372 mathematical results generated by an internal frontier AI model on GitHub. The collection includes solutions to open problems, improvements to computer algorithms, and progress on the Riemann hypothesis. The company opted for a repository over traditional academic journals, citing the need for faster dissemination of findings.
The generation process for these results typically involved a single prompt to one AI agent, with an average compute cost equivalent to three hours of ChatGPT Pro Thinking. This differs significantly from a previous Navier-Stokes solution by the same model, which required 10,000 agents and millions of dollars in compute resources.
To address the challenge of manual verification, OpenAI provided formalizations in the Lean programming language. The company also released methodology details, including reasoning summaries and compute statistics. OpenAI plans to fund future workshops and conferences to facilitate the integration of AI-generated results into the scientific community.
OpenAI consulted with the Advisory Group on Mathematics and Artificial Intelligence, which includes Fields Medal winner Timothy Gowers. However, the company restricted the group from influencing the speed or methods of production. The release has sparked debate, with 25 Fields Medal winners expressing concerns about the focus on mass-producing statements over conceptual understanding.
