OpenAI Publishes 722 AI-Generated Math Manuscripts From An Unreleased Model: Here's Why It's A Big Deal

OpenAI has released 722 mathematical papers generated by an undisclosed internal model, published on GitHub. The papers are arranged into 372 groups of related findings. The papers originated from a test where the model tackled approximately 4,000 problems, according to the repository. In addition, the release contains Lean formal versions of numerous proofs and abbreviated overviews of the model's thought process for ten groups of results.

OpenAI asserts that most results followed a single procedure, with an average time of about three hours of ChatGPT Pro computing. However, some work, such as the zeta function region and Hodge Conjecture proof for CM abelian varieties, required more time. The release is significant due to its scale, cost, and verification capabilities.

Firstly, the release showcases an unprecedented quantity of AI-generated mathematical work, covering topics such as number theory, operator algebras, statistical physics, and fluid-related equations. Secondly, the cost of producing each result is minimal, as each took about three hours of ChatGPT Pro compute, suggesting a different economic landscape for research output. Thirdly, the use of Lean formalizations allows computers to verify proofs, a crucial aspect when the volume could overwhelm human referees.

OpenAI initiated these evaluations after its existing math benchmarks reached a plateau in performance. They have made some papers available with Lean formalizations while acknowledging that some unformalized results may contain errors, promising swift corrections. OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study to guide the release of the work.

The company emphasizes its commitment to understanding AI-generated results and responsible release, though independent verification of these claims is pending.