If you followed the publishing calendar of OpenAI's scientific outputs over the past months, this October 26, 2026 announcement was not a surprise — only the next chapter in a narrative that had been unfolding since the massive release of 722 manuscripts on October 6.
This time, however, OpenAI did not simply publish raw results. The company announced a new validation methodology involving external experts — dubbed "shepherds" — to help interpret, contextualize, and explain the outputs of its internal models to the relevant academic communities. And what was discovered is substantial enough to justify the effort.
Mathematics: the local smoothing conjecture in all dimensions
The first result lives in the area of harmonic analysis. In 1991, mathematician Christopher Sogge proposed the local smoothing conjecture, which describes the behavior of waves as they become focused. The core idea is that extremes of focusing must be short-lived: at typical random times, the amount of focusing is always much smaller than at worst case.
This is a familiar pattern throughout mathematics — what can happen at worst case is more extreme than what is observed after averaging. And because this conjectured phenomenon is so fundamental, it has far-reaching implications across all of mathematical analysis. The Sogge conjecture implies, for instance, the Kakeya conjecture and Stein's restriction conjecture.
Since 2019, when Guth, Wang, and Zhang proved local smoothing in two spatial dimensions, and 2025, when Wang and Zahl resolved the Kakeya conjecture in three dimensions, the mystery remained: could it be generalized to all dimensions?
OpenAI's answer was yes, and the proof is based on a novel auxiliary partial differential equation. The company's models produced a complete proof of the local smoothing conjecture in every spatial dimension greater than or equal to two, simultaneously resolving both the Kakeya conjecture and Stein's restriction conjecture in all dimensions. The output was obtained on October 18, 2026, and formalized in Lean the following day — the same proof verification system OpenAI had used in its October 6 publication.
The decision to release the output directly for mathematical community assessment, without intermediaries, reflects the strategy tested earlier and the company's confidence in the work. Not coincidentally: OpenAI's own mathematical advisory team (AGMAI) recommended on October 25 that AI results be published in open academic repositories, with Lean-formalized proofs, information about how problems were selected, and funding for researchers dedicated to understanding them.
Computer science: the Exponential Time Hypothesis is false
The second result is arguably the most radical. The Exponential Time Hypothesis (ETH) is a conjecture about the computational complexity of boolean satisfiability — a problem deeply linked to the famous P vs. NP question. Since its proposal by Impagliazzo, Paturi, and Zane in 1999, ETH has been widely considered true and used in hundreds of papers to derive conditional computational hardness results.
If OpenAI's proof holds, ETH is false. The model produced a deterministic algorithm that counts exactly the satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space. The 3-SAT case of this result would directly refute ETH.
The most remarkable aspect is not just the result itself, but the community response. OpenAI acknowledged the sensitivity: hundreds of papers in graph theory, computational biology, databases, and computational geometry depend on ETH as a premise. For this reason, the company adopted a different approach: rather than simply publishing the result, it asked three experts to lead an independent verification.
Russell Impagliazzo and Ramamohan Paturi — the same researchers who originally proposed ETH in 1999 — together with Ryan Williams (MIT), will form a working group to examine and explain the proof. Impagliazzo and Paturi were among the first to recognize that if the proof holds, fundamental assumptions in complexity theory will need to be reassessed.
The group already requested that OpenAI make the materials publicly available immediately, even before completing the analysis. The result and Lean formalization will be released on GitHub the following Monday, allowing the complexity theory community to comment in record time.
What ETH's refutation does NOT mean — and this is important — is that we will automatically deliver faster algorithms for existing computational problems. The theoretical proof opens the door, but translating theoretical advances into practical improvements will require significant research.
Physics: the existence of quantum chromodynamics
The third result, and perhaps the most ambitious, resides in theoretical physics. OpenAI claims to have produced a mathematically rigorous existence proof of certain interacting quantum field theories containing gauge fields and fermions in four dimensions — which includes Quantum Chromodynamics (QCD) and a unified completion of the Standard Model of particle physics.
The work builds on Balaban's approach in constructive quantum field theory and includes, as a subproblem, a proof of the existence of pure Yang-Mills theory — which is part of the Clay Mathematics Institute's Millennium Prize problem. The Standard Model describes the interaction of fundamental subatomic particles for three of the four fundamental forces, except gravity.
The Yang-Mills proof was produced by the model on September 9, 2026, and QCD on September 18. OpenAI delayed the release until the Lean formalization was completed, which occurred on October 16, 2026.
David Gross — who shares the 2004 Nobel Prize in Physics for his work on asymptotic freedom in Yang-Mills theories, the key property underlying the existence of these theories as mathematically rigorous objects — will lead a group of shepherds including Michael Douglas, Arthur Jaffe (who co-formulated the Millennium Prize problem with Witten), Igor Klebanov, and Nikita Nekrasov.
OpenAI has not released the full physics material — something the company acknowledges could be potentially problematic, as it prevents independent verification by the scientific body. Instead, Gross's group will prepare its own account and explanation, under the four principles of the collaboration: full transparency, no academic credit, recognition of contributors, and a fourth principle the company has not yet disclosed publicly.
What changes: the scientific publishing model
The most interesting aspect may not be the results themselves, but the structure OpenAI proposed to validate them. The company openly acknowledged that "choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information" — a trade-off that no prior peer review structure had articulated so clearly.
The collaboration principles include full transparency (no confidentiality agreements, freedom to discuss and publish conclusions), no academic credit (OpenAI does not seek credit for the results), and formal recognition of researchers who assist in understanding and contextualizing the work. The shepherd group will also be responsible for ensuring that the work appropriately credits researchers, techniques, and ideas that it draws on.
The Simons Foundation recently announced a $15 million contribution — of which $5 million comes directly from OpenAI — to help mathematicians navigate the AI era, a sign that the basic science funding ecosystem is beginning to respond to the speed with which AI models are producing results that previously took decades to achieve.
The question is whether this new model of academic "shepherding" will be able to scale beyond already-previously-connected experts, or if it will ultimately reproduce, like so many other innovations in open science, the same asymmetries of access it seeks to avoid.
Sources: OpenAI, The Verge, The Next Web, Interesting Engineering
✓ Independent sources cross-checked and verified before publishing
