OpenAI says its AI systems produced proofs for three major open problems in mathematics, computer science and physics - and it's bringing in outside experts to check them.
The ChatGPT maker is also testing new ways to share AI-generated research with the scientific community.
The company said it is experimenting with three approaches to disclosure: publishing AI-generated work immediately, releasing it alongside an outside group of collaborators, or allowing an external team to decide when materials become public. OpenAI is seeking feedback as it works out which model best supports researchers.
The collaborations follow four principles: no nondisclosure agreements, no claims to academic credit for the underlying discoveries, recognition for researchers who develop explanations and supporting materials, and voluntary participation without compensation. Collaborators will speak only for themselves.
In mathematics, OpenAI said its models proved a decades-old conjecture about how waves spread and focus. Known as the local smoothing conjecture, it predicts that waves averaged over time behave less intensely than they do at their most extreme moments.
The result would also resolve two related longstanding problems: the Kakeya conjecture, which concerns how small a shape can be while still containing a line segment in every possible direction, and Stein's restriction conjecture, which examines how the mathematical frequencies of functions behave along curved surfaces.
The company said it generated the output on Oct. 18, formalized it in Lean, a proof-verification language, the following day and published both a natural-language explanation and a formal proof on GitHub.
In computer science, OpenAI's system produced a disproof of the Exponential Time Hypothesis, a widely used assumption underpinning conditional hardness results for satisfiability problems and related questions in computational complexity. The natural-language disproof was generated on Oct. 16 and formalized in Lean on Oct. 17, according to the company.
A team led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams is now examining the result. OpenAI also pledged to fund a conference to explore its potential implications.
In physics, an external group led by David Gross, alongside Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, will assess what OpenAI describes as a rigorous existence proof for interacting quantum field theories in four dimensions, including quantum chromodynamics (QCD) and a grand unified extension of the Standard Model.
Unlike the mathematics and computer science projects, the physics group has opted not to release the AI-generated materials publicly for now. Instead, its members will focus on preparing their own written account of the work.
The claims could have far-reaching implications if they hold up to independent scrutiny, particularly the claimed disproof of the Exponential Time Hypothesis, which underpins much of modern computational complexity theory.
But OpenAI's announcements are only the starting point: outside researchers must still determine whether the proofs are valid and what they mean for their respective fields.
The effort also highlights a broader question for AI research - how to make potentially groundbreaking machine-generated discoveries transparent, reproducible and credible to the scientific community.
