An internal OpenAI model has produced a claimed disproof of the Exponential Time Hypothesis, a conjecture on the difficulty of Boolean satisfiability that underpins hundreds of computer science papers. The company said the result was found on 16 October and checked in the machine-verification language Lean the next day. The write-up and Lean proof were published on GitHub. If confirmed, the claim matters because the hypothesis has been used to assess hardness in graph theory, computational biology and databases.
How OpenAI presented the three results
OpenAI announced three results on the same day, covering computer science, mathematics and physics. In its blog post, the company described the Exponential Time Hypothesis as widely believed to be true and linked it to P versus NP. The computer science claim concerns Boolean satisfiability, a central problem for complexity analysis. The other two results concern the local smoothing conjecture proposed by Christopher Sogge in 1991 and the existence of certain four-dimensional quantum field theories. All three were attributed to internal models rather than to outside collaborators.
The computer science verification path relies on Lean and public release of materials. OpenAI said it verified the claimed disproof by machine on 17 October, one day after discovery. It published both the model write-up and the Lean proof on GitHub for independent study. For mathematics, it followed the same pattern, releasing the write-up and Lean proof, consistent with the 722 mathematics papers it published on 6 October. In physics, the approach differs: the review group decided not to release the AI-produced material yet and will prepare its own account.
The review structure places the original proposers at the center of scrutiny. Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT will check and explain the complexity claim. Impagliazzo and Paturi proposed the hypothesis with Francis Zane, which gives the examination direct continuity with its origin. In an open letter, the three researchers said their involvement should not be taken as certification. Williams said the group signed no non-disclosure agreements, will not be paid, requested all raw data and a funded community conference, and OpenAI agreed to both requests.
What confirmation would mean for business use of AI
For companies that rely on optimization, search and data systems, the practical effect depends on independent confirmation. The hypothesis has supported claims that many problems in graph theory, databases and computational biology cannot be solved quickly. A verified refutation would force a reassessment of those hardness arguments and of the assumptions behind solver selection and performance expectations. Small teams using off-the-shelf solvers would feel it through updated guidance, while large organizations with custom optimization stacks would need deeper review of models built around the old lower bounds.
The mathematics and physics claims show a separate operational question around verification and trust. Local smoothing in every dimension would also settle the Kakeya conjecture and Stein restriction conjecture in all dimensions, after local smoothing in two dimensions in 2019 and Kakeya in three dimensions in 2025. The physics work includes pure Yang-Mills existence, part of a Clay Mathematics Institute Millennium Prize problem, alongside quantum chromodynamics of quarks and gluons. The physics group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, will write up the work itself, so business users cannot yet evaluate methods or limits.
The marker to watch is the outcome of the independent examinations and the promised conference. The computer science group must report whether the Lean proof holds and what weaker hardness statements survive if the original conjecture fails. The physics write-up will show whether the existence proofs meet community standards without reliance on unreleased model output. OpenAI said it does not seek academic credit, assigns recognition to researchers who explain the work, and gave $5m to the Simons Foundation for workshops. How credit and checking are handled there will signal whether AI-produced results can move from announcement to usable knowledge.
