OpenAI Releases Hundreds of Math Manuscripts from Unreleased AI Model
- OpenAI published a total of 722 math manuscripts on GitHub, generated by an internal AI model that remains unreleased.
- Only 162 of these manuscripts have a Lean-formalized main result, indicating potential issues with unformalized results.
- MIT’s Andrew Sutherland expressed skepticism about the one-prompt, single-agent claim, emphasizing the need for verification through model release.
- The papers are organized into 372 families, with each family potentially containing multiple related results or proofs.
- OpenAI released reasoning summaries for only ten results and acknowledged that not all manuscripts have complete Lean formalizations.
OpenAI’s decision to publish these manuscripts has sparked debate in the mathematical community due to the lack of full verification and transparency regarding the AI model used. The publication includes significant claims that remain unverified until further disclosure occurs.
Despite releasing average compute figures and some reasoning summaries, OpenAI has not disclosed specific prompts or fully formalized all results, leaving many mathematicians calling for more evidence and transparency. (Source)
