OpenAI Releases 372 AI-Generated Mathematical Demonstrations

OpenAI publishes 372 mathematical results produced by an internal model and places them on GitHub with automatic verification elements. While the company promises more formalizations and details its methodology, leading mathematicians are warning about the limits of large-scale production and its effects on the discipline.

Warnings from Fields Medalists and the Limits of Lean

Formalizations in Lean allow for the certification of the logical coherence of a proof, but do not address its relevance or originality. Twenty-five Fields Medal laureates have signed an open letter highlighting a profound disconnect between the goals of the AI industry and those of mathematics. They argue that problem-solving is merely a means, far from the goal of conceptual understanding, and that the mass production of true statements could impoverish the landscape of ideas. Timothy Gowers warns that within one to two decades, the literature could swell without a human community truly understanding it. Terence Tao advocates for human-centered training for the young and strict limitations on AI tools to preserve authentic learning. Reactions range from enthusiasm to irritation, while OpenAI, on its part, bets on advancements at the frontiers of knowledge.

372 Results Published on GitHub with Lean Verifications

OpenAI is releasing 372 mathematical results from an internal model and has chosen GitHub as the medium rather than journals. A significant portion of the proofs is accompanied by Lean formalizations, with more expected to follow. The collection includes improvements to major algorithms and advancements related to the Riemann hypothesis. The documents come with revision logs and citations, and each result is presented as solving an open problem or making substantial progress. The company describes the generative model as a state-of-the-art system.

Production: One Agent per Query and an Average of Three Hours of Computation

OpenAI indicates that most results come from a single query to a single agent, sometimes with multiple attempts. On average, each result requires about three hours of computation with ChatGPT Pro. In contrast, a solution to a Navier-Stokes problem, attributed to the same model, mobilized a swarm of 10,000 agents and millions of dollars in computation and has been undergoing formal review for several weeks. The company publishes reasoning summaries, statistics on attempted problems, and cost estimates, asserting that the scale of production could exceed manual proofreading capabilities, which motivates the use of formal verification.

Process and Governance: IAS Consulted and Workshops Announced

OpenAI states that it has consulted the Advisory Group on Mathematics and AI at the Institute for Advanced Study, where Timothy Gowers is a member, and claims to have loosely followed its public recommendations by not publishing queries and only communicating average costs. The company plans to fund workshops and conferences dedicated to the appropriation of these results, aims to enhance the quality of citations and presentation, and is working towards a responsible publication of the model. The choice of GitHub over peer-reviewed journals structures the dissemination of this corpus.