OpenAI Says 372 Math Results Came Mostly From One Prompt
TL;DR
- Only 162 of 722 manuscripts carry Lean-verified proofs; 560 results rest on community trust, not machine-checked correctness.
- The model attempted roughly 4,000 problems to produce 722 publishable manuscripts, an ~18% acceptance rate at around three ChatGPT Pro compute hours per accepted result.
- OpenAI disabled GitHub Issues on the repository and withheld the prompts the IAS advisory group requested, blocking two standard community verification channels.
OpenAI released 372 mathematical results from an unreleased internal model, and a company spokesperson told Scientific American the new model "produced almost every one of the results in response to a single prompt handed to a single AI agent," though some needed multiple attempts.
The haul includes a claimed resolution of the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis, with proofs posted in a public GitHub repository. Per-result compute averaged roughly three hours of ChatGPT Pro inference, a sharp drop from the "10,000-strong agentic swarm that cost millions of dollars" OpenAI used on an earlier Navier-Stokes result.
Mathematicians want receipts. "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," MIT's Andrew Sutherland said. An independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study had recommended OpenAI publish the model, the exact prompt, and compute time per problem; the company released only averages, and the spokesperson said OpenAI "is not bound by these recommendations."
Not every mathematician sides with secrecy. "If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret from us," said Toronto mathematician Daniel Litt. Terence Tao has called the pace of AI-generated results from frontier labs "insane."
The post lands inside a steady drumbeat of OpenAI coverage on our beat, and the field now has 372 claims to work through, plus a repository to pick apart.
What others are reporting
-
Decrypt Read →
Draws direct comparison to Anthropic's Fermat release as a transparency benchmark; reports GitHub Issues disabled and prompts withheld despite IAS recommendations.
Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified.
-
The Decoder Read →
OpenAI prohibited its advisory group from advising on release pace; 25 Fields Medal winners objected that the mass approach deprioritizes conceptual understanding.
Nearly every result came from a single prompt to a single AI agent, though some took multiple attempts.
-
Interesting Engineering Read →
Grounds the story in computational economics: ~4,000 problems attempted, ~18% acceptance, ~3 hours of ChatGPT Pro compute per result, the cost structure that makes volume releases repeatable.
An accepted result consumed about three hours of equivalent ChatGPT Pro thinking compute on average.
-
Unite.AI Read →
Leads with the IAS Advisory Group's specific responsible-release recommendations including model name, prompts, compute costs, and formalization standards, documenting what OpenAI declined to provide.
The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model.
Shared on Bluesky by 1 AI expert
Originally reported by scientificamerican.com
Read the original article →Original headline: OpenAI Says One AI Agent on a Single Prompt Produced 372 New Mathematical Results