Three New Claims in Fundamental Science
Less than a month after publishing hundreds of manuscripts claiming solutions to long-standing mathematical problems, OpenAI has announced three additional major scientific claims spanning mathematics, theoretical computer science, and physics. The latest disclosures mark an attempt by the artificial intelligence laboratory to adjust its disclosure methodology by involving outside researchers directly in the evaluation process.
Unlike its previous mass release on October 6, where model outputs were published directly, OpenAI has introduced external academic experts termed “shepherds” to help oversee and explain the findings in theoretical computer science and physics. For the mathematical discovery, the company stated that advisers recommended publishing the raw model outputs immediately so that the broader academic community can work through explanations independently.
Refuting the Exponential Time Hypothesis
In theoretical computer science, OpenAI claims to have demonstrated that the widely accepted Exponential Time Hypothesis (ETH) is false. The research presents a deterministic algorithm designed to calculate exact satisfying assignments of fixed-width conjunctive normal form (CNF) formulas within subexponential time and polynomial space. Under this framework, the algorithm’s application to 3-SAT refutes the hypothesis.
Verification of this claim is being directed by an independent working group of prominent computer scientists: MIT professor Ryan Williams alongside UC San Diego researchers Russell Impagliazzo and Ramamohan Paturi, both of whom originally formulated the Exponential Time Hypothesis in 1999. At the request of the working group, OpenAI released the write-up and a formalized proof in Lean directly to GitHub to allow open examination by the wider theoretical computer science community.
Alongside the technical materials, the working group published an open letter addressing the broader implications of evaluating machine-generated scientific claims. The letter invites community input around three foundational questions: establishing appropriate expectations for early access, conflict of interest disclosures, and the handling of unpublished material; determining how credit should be apportioned between automated discovery and the subsequent human effort required to check, correct, simplify, and explain the work; and finding methods to support emerging researchers as workflows evolve while preserving rigorous analytical judgment.
Wave Equations and Conjectures in Mathematics
OpenAI’s mathematical finding addresses local smoothing for the Euclidean wave equation. According to the company, the generated proof utilizes an auxiliary partial differential equation to establish Christopher Sogge’s local smoothing for waves across all spatial dimensions equal to or greater than two.
The company stated that the result additionally resolves the Kakeya maximal and dimension conjectures as well as Elias Stein’s spherical restriction estimates across the interior of the conjectured range. The Kakeya problem has drawn significant attention across modern mathematics; just months prior, New York University professor Hong Wang received the Fields Medal from the International Mathematical Union for resolving the Kakeya conjecture specifically in three dimensions. OpenAI’s model output claims to resolve the local smoothing, Kakeya, and Stein restriction conjectures across all dimensions.
Quantum Field Theory and Academic Concerns
The third result, rooted in theoretical physics, has not yet been released in full. OpenAI characterizes the manuscript as a mathematically rigorous existence proof addressing specific interacting quantum field theories containing gauge fields and fermions in four dimensions. The scope includes Quantum Chromodynamics (QCD) and a unified completion of the Standard Model of particle physics.
Guidance for this physics work is being led by David Gross alongside a group of physicists and mathematicians including Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov, who will shepherd the findings through review.
Despite the implementation of expert shepherds, OpenAI acknowledged that its collaboration model presents structural challenges. Just a day prior to the announcement, the AGMAI advisory panel of mathematicians convened by OpenAI warned against selecting privileged reviewers. The panel stated, “we recommend that AI companies do not adopt” procedures that grant early access to designated individuals, highlighting problems including unequal access across the research community, borrowed institutional reputation, and biased reviewer selection.
The AGMAI panel advised that artificial intelligence laboratories should instead distribute their findings openly across public scholarly repositories, supply Lean-formalized proofs, publish details explaining how research problems were chosen, and provide financial resources to support independent researchers who dedicate time to validating and understanding the results.
