OpenAI Claims AI Breakthroughs in Kakeya, ETH, and 4D Quantum Field Theory

OpenAI announced that its internal models have produced three major results in harmonic analysis, computational complexity, and quantum field theory: proving the local smoothing conjecture in all dimensions, thereby deriving the Kakeya and Stein restriction conjectures; finding a subexponential-time algorithm to exactly solve fixed-width SAT, thus refuting the widely used Exponential Time Hypothesis (ETH); and providing rigorous existence proofs for four-dimensional Yang-Mills theory, quantum chromodynamics (QCD), and a class of standard model unification extensions. The first two results have been published as natural language papers alongside Lean-formalized code; the third set of materials remains unpublished, and none of the three have yet undergone independent peer review. What truly matters is not just the results themselves, but a new research workflow emerging: AI rapidly generates proofs, formal systems verify the derivations, and external experts then validate, interpret, and establish academic consensus.

Article author, source: OpenAI

OpenAI refers to these jobs as "discoveries" in the headline and X post, but uses a more cautious description in the body: the current manuscript is best understood as "raw research data" that still requires the academic community to transform it into scientific knowledge that can be understood, reviewed, and cited.

The level of transparency for the three results varies. The harmonized analysis results have been directly published; the computational complexity results have been made public along with the paper and Lean code, and have been verified by an external expert panel; the physical results have only disclosed the conclusions and timeline, with the full proof temporarily held by an expert group. OpenAI has not disclosed the specific model names, reasoning budgets, number of searches, or failed samples used to generate these proofs.

The first result belongs to harmonic analysis.

The local smoothing conjecture investigates how much waves can concentrate during propagation. Intuitively, even if a wave can be highly concentrated at a single moment, this extreme concentration should significantly diminish when observed over a slightly averaged time interval.

This conjecture is significant because it is connected to several core problems, including the Kakeya conjecture and the Stein restriction conjecture. In 2019, Larry Guth, Hong Wang, and Ruixiang Zhang resolved the corresponding local smoothing problem in two dimensions; in 2025, Hong Wang and Joshua Zahl proved the three-dimensional Kakeya conjecture.

OpenAI now claims that its model, using a new auxiliary partial differential equation, has proven the local smoothing conjecture in all spatial dimensions, thereby yielding the Kakeya conclusion and Stein restriction estimates in all dimensions. The natural language proof was generated on October 18, and its Lean formalization was completed the following day.

The publicly available Lean documentation shows that the formalized content includes local smoothing estimates for the Euclidean half-wave operator, estimates for the Kakeya maximal operator, the fact that Kakeya sets have full Hausdorff dimension, and several Bochner–Riesz and spherical restriction estimates.

This is stronger than simply uploading an AI-generated paper, because anyone can download the code and have the Lean kernel re-verify each step of the derivation. However, this still does not mean the mathematical community has accepted all conclusions: experts must still confirm that the formalized statements exactly match the original conjectures, that definitions have not been weakened, and that the natural language paper genuinely explains why the proof holds.

The second result may have a more direct impact on theoretical computer science.

The Exponential Time Hypothesis posits that there is no general algorithm for fixed-width Boolean satisfiability problems whose runtime grows subexponentially with the number of variables. Over the past several decades, numerous papers in graph theory, databases, computational geometry, and computational biology have used ETH as a condition to argue that certain problems are unlikely to admit significantly faster algorithms.

The paper released by OpenAI presents a deterministic algorithm that can precisely compute all satisfying assignments for a fixed-width CNF formula. The runtime of the Lean formalization is a polynomial factor multiplied by:

\(2^{O(n(\log\log n)^2/\sqrt{\log n})}\)

For cases where the number of clauses is linear in the number of variables, the paper provides a stronger:

\(2^{O(n/\sqrt{\log n})}\)

Both are sub-exponential time and require only polynomial space.

It is important to clarify: this does not prove that P equals NP. Although the algorithm is faster than exponential time in the usual sense, it is still not polynomial time; it refutes the ETH regarding how fast SAT can be solved, not the P versus NP problem directly.

This result was led by Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams. The first two were among the original proposers of ETH. In accordance with the working group’s requirements, OpenAI has made all original materials publicly available and has committed to funding a conference open to the entire research community.

The third claim concerns four-dimensional quantum field theory.

OpenAI stated that the model provides rigorous existence proofs for several four-dimensional interacting quantum field theories containing both gauge fields and fermions, following Balaban's approach in constructive quantum field theory, including QCD and a unified extension incorporating standard model fields and fermions.

This work also includes a proof of existence for the pure Yang–Mills theory. While pure Yang–Mills theory is part of the Clay Mathematics Institute’s Millennium Prize Problems, OpenAI’s statement addresses only the subproblem of “existence” and cannot be construed as a complete resolution of the Millennium Prize Problem, which includes the requirement of a mass gap. Consequently, the prize has not been awarded.

The model is claimed to have generated a proof of the Yang-Mills existence and mass gap in September 9, a proof of QCD on September 18, and completed the Lean formalization of the Yang-Mills portion on October 16. The review panel, led by Nobel Prize in Physics laureate David Gross, includes Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov.

However, neither the paper nor the Lean code has been made public, so outsiders can only verify that OpenAI made these claims, but cannot examine the proofs themselves. Therefore, among the three results, this one carries the highest uncertainty.

The role of Lean is to verify that there are no uncaught logical gaps between a set of well-defined axioms, definitions, and propositions. Once the code passes through the kernel, it provides strong evidence of internal consistency.

However, formal verification still leaves two important issues.

First, whether the formalized proposition fully corresponds to the original conjecture discussed in the academic community. If the translation alters the assumptions, scope, or boundary conditions, Lean may correctly prove the formal proposition but cannot automatically determine whether this proposition is the one humans truly care about.

Second, a proof must not only be correct but also explain why the new method works, how it relates to existing literature, and whether it can be generalized. Completing this work still requires domain experts to read, restructure, and write an understandable version.

Therefore, OpenAI proposed four collaboration principles: no requirement for experts to sign non-disclosure agreements; no claim for academic authorship on discoveries generated by internal models; experts should be recognized for their work in verification, explanation, and extension; participation is voluntary and unpaid, with experts free to publish their conclusions independently.

However, this system remains controversial. OpenAI itself acknowledges that designating the first group of "shepherds" creates information asymmetry; its mathematical advisory group has also warned that prioritizing unpublished results for a select few researchers may lead to unequal access, reputation borrowing, and selection bias.

If these three tasks ultimately succeed, their significance lies not only in AI solving a few more problems, but in the potential shift in the pace structure of theoretical research.

The model can generate a large volume of highly complex candidate results in a short time, and Lean can filter out some formally incorrect ones; subsequently, the most scarce resource will become expert attention—who will validate the problem definition, assess the relationship with existing literature, explain new techniques, and determine whether the results are worth further development.

This also means that “AI made the discovery” should not be interpreted as scientists having exited the process. AI may handle search and derivation, formal systems may handle logical verification, while humans remain responsible for problem selection, interpretation of significance, academic credit, and community recognition.