AI-rewritten: This is a summary of an article from New Scientist, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

OpenAI recently released 722 mathematics papers on GitHub, addressing over 4,000 open problems using an internal model. Each result reportedly required the equivalent of three hours of "ChatGPT Pro thinking compute," though OpenAI has not provided a true cost figure for these outputs. While one specific problem previously solved by the company took around 88 hours and cost approximately $15 million, the average effort here is likely lower. The release aims to test the limits of AI models rather than create a dedicated mathematics machine, as part of a broader goal toward artificial general intelligence.

A key advantage of mathematics for AI development is verifiability, unlike fields such as creative writing where objective measures are harder to establish. By formalizing proofs into computer code called Lean, which can be mechanically checked, companies can verify accuracy and improve their models through a feedback loop. However, OpenAI’s current release is incomplete; only 162 of the 722 papers have had their main results formally verified in Lean. Furthermore, just 235 "families" of results include any accompanying code out of 372 total families across all papers.

Terence Tao at UCLA compared the unverified papers to "dumping carcasses of raw meat onto our communal village table." OpenAI cites guidelines from the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) for releasing results quickly, though AGMAI has stated its advisory role is not an endorsement. The company plans to update its repository with more formalizations as it obtains them, but Kevin Buzzard at Imperial College London noted that formalization requires filling in vast amounts of existing machinery, which can be time-consuming.