OpenAI publishes mathematical manuscripts whose significance remains to be assessed

OpenAI announces the publication of new results on open problems in mathematics, produced by an advanced internal model, along with the sharing of proofs formalized in Lean and research details on GitHub. The Verge describes a collection of 722 manuscripts, grouped into 372 families of results, produced by an unreleased model. According to AGMAI, the independent advisory group formed to support communication about the results, the collection includes solutions to hundreds of open questions.

This assessment remains attributed to the group. The announced volume is insufficient to establish the validity of each proof or measure the importance of each result. The Verge emphasizes that the mathematical community still needs to examine and absorb the work.

AGMAI published its first recommendations in late September, calling for rapid dissemination and for established academic channels to be used where possible. The group also requests information about the model, the instructions used and computing costs. It advises against turning these publications into promotional materials.

OpenAI says it is using a GitHub repository, with procedures for revising papers and citing them. According to The Verge, the documents include summaries of the model's reasoning, computing estimates and statistics on the problems attempted. The challenge now is to distinguish the claimed solutions, their verification and their scientific contribution. The information released can facilitate this examination without replacing its conclusions.