OpenAI Claims Three Scientific Breakthroughs, Introduces 'Shepherd' Review Model

OpenAI announced on October 26 that its internal models have produced research-level results in three fields simultaneously: a proof of the local smoothing conjecture in all dimensions (which also resolves the Kakeya and Stein restriction conjectures), a disproof of the Exponential Time Hypothesis in computational complexity, and a mathematically rigorous existence proof for certain four-dimensional quantum field theories including Quantum Chromodynamics and a unified completion of the Standard Model. The company characterized the manuscripts as "a form of raw research data" that must be converted into scientific understanding through external expert review [1].

Mathematics: Local smoothing and Kakeya in all dimensions

The mathematics result concerns Sogge's local smoothing conjecture from 1991, which describes how wave focusing behaves when averaged over time. Guth, Wang, and Zhang proved the two-dimensional case in 2019; Wang and Zahl proved the three-dimensional Kakeya conjecture in 2025. OpenAI says its models produced a proof for every spatial dimension of at least two, using what it calls a novel auxiliary partial differential equation [1][2]. The company obtained the AI output on October 18, formalized it in Lean the next day, and released both the natural-language writeup and Lean proof publicly on GitHub, following the same strategy as its October 6 release of hundreds of mathematical manuscripts [1][2].

University of Rochester professor Alex Iosevich, whose PhD adviser was Chris Sogge, called the claim a "historic development" but acknowledged "a twinge of sadness" about how AI may change theoretical mathematics [10].

Computer science: Exponential Time Hypothesis claimed false

The Exponential Time Hypothesis, proposed by Impagliazzo, Paturi, and Zane, asserts that 3-SAT cannot be solved in subexponential time. It has been used in hundreds of papers to derive conditional hardness results across graph theory, computational biology, databases, and computational geometry [1][2][5]. OpenAI says its system produced a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space, with the 3-SAT case refuting ETH. The natural-language disproof was generated October 16 and formalized in Lean on October 17 [1][2].

The result will be examined by a collaboration led by Russell Impagliazzo and Ramamohan Paturi (UC San Diego), the original proposers of ETH, and Ryan Williams (MIT), who has previously refuted strong versions of ETH. At the group's request, OpenAI made all raw data public and committed to funding a conference on the implications [1][2]. The Lean formalization documentation states the algorithm runs in subexponential time and polynomial space [2].

Physics: QCD and Yang-Mills existence proofs

OpenAI says its models produced a natural-language proof of pure Yang-Mills existence on September 9 and a proof for QCD on September 18. The Lean formalization of Yang-Mills existence completed on October 16 [1][2]. The claimed existence proof covers interacting quantum field theories with gauge fields and fermions in four dimensions, including QCD (describing quarks and gluons) and a unified completion of the Standard Model. Pure Yang-Mills existence is a subproblem of the Clay Mathematics Institute's Quantum Yang-Mills Millennium Prize problem, though the announcement does not claim the full mass-gap problem is solved [1][2].

A collaboration led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov, will shepherd this work. Unlike the other two results, this group decided to withhold the AI-produced artifacts while preparing its own writeup [1][2][6].

The 'shepherd' model: Three release strategies

OpenAI asked external experts to act as "shepherds" organizing groups to understand each result, and the three groups advised different release approaches [1][2]. For mathematics: release AI output directly. For computer science: form an external collaboration while making AI output public. For physics: form a collaboration and defer to it on release timing [1][2][9].

OpenAI published four principles for these collaborations: transparency (no NDAs, participants free to discuss and publish), no academic credit for OpenAI (the company says it does not seek academic credit for discoveries produced by its internal models), recognition for explanatory work, and independence (participation voluntary and unpaid, conclusions are the groups' own) [1][2][6]. The company acknowledged that "choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information" [1][2].

This follows criticism of OpenAI's October 6 release. An OpenAI-convened advisory panel (AGMAI) recommended the day before this announcement that AI companies "do not adopt" the practice of designating certain people to receive output first, citing unequal access, borrowed reputation, and biased selection [3].

Verification status: Claims, not settled science

None of the three results have been independently verified by the broader scientific community. OpenAI itself calls the manuscripts "raw research data" [1][2]. The computer science result has Lean verification, but as the shattered.io analysis notes, "a proof assistant confirms that a formal argument is valid. It cannot confirm that the formal statement matches the informal claim in the paper's title" [5]. The physics artifacts remain unpublished. The mathematics proof was released directly without a designated review group.

OpenAI's $5 million contribution to the Simons Foundation, announced previously, will fund workshops and conferences to help researchers examine these and other AI-generated results [1][2].