OpenAI posted 722 mathematical manuscripts to GitHub on Tuesday, grouped into 372 families of results and all produced by an internal model it has not released.
The company says it posed about 4,000 open problems to the system and kept what it judged significant. The average result took the equivalent of roughly three hours of ChatGPT Pro reasoning. The subjects run from number theory and combinatorics to theoretical computer science.
Its boldest claim concerns method. OpenAI said almost everything came from a single prompt handed to a single AI agent.
The check on that is thinner than the headline. Some manuscripts come with proofs written in Lean, a language that lets a computer verify each step. Decrypt counted 162 of them, about 22 percent. OpenAI itself warned that the unformalised results could have issues.
Mathematicians were divided. Abhishek Saha called it a very big day for the field, while noting most results are incremental advances. Andrew Sutherland of MIT said claims of solving problems in one shot should be treated as unverified, and that the community should ask for receipts.
The stakes extend past mathematics. A model that produces new proofs at this rate changes what research labs, and cryptographers, can assume about the pace of discovery. Within a day an Ethereum researcher cited the release as grounds to rethink wallet security.
What did not arrive is the means to check the central claim. OpenAI released neither the prompts nor the model, which falls short of recommendations an advisory group at the Institute for Advanced Study published on 29 September. Roughly four in five papers now wait on human referees. Nobody has said how long reading 560 of them will take, or who will do it.
