OpenAI Releases Math Results from an Internal AI Model, Complete with Verifiable Proofs
The company posted machine-checkable proofs and research process materials on GitHub. It says the model behind the results is being prepared for release, but gave no date or price.
On October 6 local time, OpenAI published new mathematics research results obtained with an AI model it is developing internally. The accompanying materials were posted to GitHub, the site developers use to share code and documentation. Several of the proofs were released in a form that computers can verify.
OpenAI said the internal model that produced these results is also being prepared for release. The announcement page did not specify who would get access, when it would launch, or how much it would cost.
Can the proofs be checked by computer?
The released materials include Lean files that translate the proofs into a form a computer can check. Lean is a tool used to verify the logical steps of a mathematical proof by machine. Readers can view the write-ups explaining the research results alongside the proofs intended for verification. OpenAI said it would post more material as it becomes available.
The company also uploaded materials documenting how the model arrived at its answers, including 10 summaries of the model’s reasoning process. Statistics on the number of problems attempted and estimates of the compute used were included as well. Rather than showing only the final results, the release lays out how many problems were attempted and how much computing power went into them.
OpenAI described the average amount of compute required to produce a single result by comparing it to roughly three hours of using ChatGPT Pro’s thinking feature. The announcement page did not include separate benchmark scores.
Who is this for?
Mathematics researchers and anyone tracking AI’s problem solving ability can begin reviewing the material right away. The repository also includes the procedures to follow when correcting or citing the papers. The idea is not just for people to read the results, but to carry them forward into follow-up research.
OpenAI said it settled on this release approach after incorporating recommendations from an independent advisory group of mathematics and AI experts. It is also considering an additional, community-run publication channel that would align with the advisory group’s guidance. The company plans to fund workshops, conferences, and special programs that help people understand major AI-generated results.
The core of this announcement is that OpenAI published AI-produced mathematical results together with the materials needed to scrutinize them. What to watch next: how much the body of Lean-verifiable proofs grows, and when the model being prepared for release opens up, and to whom.