OpenAI drops 722 manuscripts of math results

OpenAI's latest release follows a September statement that its model had 'resolved more than 100 long-standing open problems across most areas of mathematics.' The new batch of 722 manuscripts, covering 372 result families, arrives after weeks of expectation, during which OpenAI had not provided details about which problems were solved or when they would be published. The released materials also include summaries of the model's reasoning, estimates of compute used, and stats about problems attempted.

The timing intersects with new guidance from AGMAI, or the Advisory Group on Mathematics and Artificial Intelligence, which published its first recommendations in late September. AGMAI urged AI labs to release mathematical results promptly and through established academic channels where possible, while disclosing details such as the model name, prompts, and compute costs. It also implored companies to 'refrain from treating the release of mathematical results as marketing vehicles to promote their models,' a practice it said inflicts significant harm on the mathematical community. OpenAI said it is publishing through a GitHub repository with protocols for paper revisions and citations.

How the field receives this batch may hinge on whether mathematicians find the exposition and citations sufficient for verification, and on whether OpenAI's approach is seen as consistent with AGMAI's guidelines. The results add to a growing body of mathematical output from OpenAI and rival labs like Anthropic, including results concerning a Millennium Prize problem, that researchers are still processing.