The channel How Did They Prove That? released a video breakdown of a recent result by OpenAI: the existence of four-dimensional quantum Yang-Mills theory has been established as the continuum limit of Wilson lattice theory, with the entire reasoning formalized in Lean 4. The OpenAI-YM-20261028 source code, comprising approximately one million lines, was published on November 11, 2026, and a committee of five physicists declared the proof probably correct. This is essentially the first case where an LLM generated a formalized proof for a Millennium Prize Problem and received independent confirmation.

image
image

What happened

In October 2026, OpenAI announced a proof of the existence of 4D quantum Yang-Mills theory, fully formalized in Lean 4. On November 11, 2026, the OpenAI-YM-20261028 source code was published: approximately one million lines of Lean across 9,477 files, made public on GitHub. The result was evaluated by a committee of five physicists—David Gross, Michael Douglas, Arthur Jeff, Igor Klebanov, and Nikita Nekrasov—who declared the proof probably correct and accompanied the verdict with the publication of the Lean version and a review article. The full article, written by humans, had not yet been released at that time. On November 14, 2026, the channel How Did They Prove That? released a video breakdown of this work. It step-by-step demonstrates how the reasoning is structured: the theory is constructed as the continuum limit of Wilson lattice theory, with the transition to the continuum governed by the Balaban renormalization group; as explained in the video, this construction provides a rigorous mathematical foundation for QCD, confinement, and the mass gap. In parallel, Scott Armstrong's project in the scottnarmstrong/yang-mills-informalization repository is compiling a map of the proof and a 278-page natural language manuscript.

Context

The problem in question belongs to the list of Millennium Prize Problems: it requires rigorously constructing quantum Yang-Mills theory in four-dimensional spacetime mathematically, without relying solely on physical arguments. The traditional constructive path goes through Wilson's lattice formulation: the theory is first defined on a discrete lattice, where all quantities are finite and explicitly expressed, and then the existence of the limiting transition as the cell size shrinks to zero is proven. In this transition, the Balaban renormalization group operates—an apparatus that manages the behavior of the theory as it moves toward the continuum, and on which the discussed construction relies. Lean 4 is an interactive formal verification tool: every logical link in the reasoning is checked by a small trusted kernel, so the correctness of the chain does not depend on trust in its author. This gives rise to a new division of labor: the machine guarantees the correctness of individual steps, while humans examine faithfulness, i.e., the correspondence between the formal statement and the substantive physical formulation. Formalization is currently outpacing human interpretation: the million-line code is already ready, while the version readable by mathematicians and physicists is still being compiled, making the clarity of proofs an independent bottleneck.

Why this matters for the industry

The main signal for the industry is not physics, but the economics of verification. Machine checking of the formal outline remains cheap—essentially a compiler constant—but faithfulness checking has not been eliminated: it was performed by a committee of physicists and remains expensive human work, resulting in the cautious verdict of 'probably correct.' Nevertheless, expensive 'guaranteed intellectual work' is ceasing to be the exclusive prerogative of humans and laboratories: the model produced an artifact where every step is machine-verifiable, and the semantic side is verified by an independent commission. The format of the result is also changing: the OpenAI-YM-20261028 artifact with committee confirmation becomes the benchmark by which industries will have to measure subsequent claims of 'solved problems'—the reference look of verified AI work is now set. For production, this is currently not a product but a pattern demonstration: OpenAI's generative pipeline is not open to the outside, it has no API, methodology, or cost and latency metrics, so there is no direct impact on builds today. The one transferable takeaway is to build eval sets around deterministic machine-verifiable artifacts, not just around evaluating models. If the practice of 'LLM generates proof—committee verifies' is repeated, demand for formalization infrastructure will grow: agent libraries on top of Lean 4, tools for inspecting formal artifacts in the spirit of code review, and expert roles at the intersection of agent pipelines and formal methods; with a reproducible pipeline, the human role in the entire chain will shift toward stating assertions and interpretation.

Why this matters for users

Readers have access to different layers of the story. The essence can be understood through the video by the channel How Did They Prove That? on YouTube—there, the construction with the lattice limit and the Balaban renormalization group is presented visually, without formal code. Those who want to check it themselves can open the OpenAI-YM-20261028 source on GitHub or the committee's review article as a PDF on the KITP website. For those for whom Lean is not suitable, the scottnarmstrong/yang-mills-informalization project is a good fit: a natural language manuscript is being prepared there. Auxiliary services can already be built on the open materials—a navigator for repository files, a chat with links to specific Lean lines, a proof map summary; basic versions of such tools can be assembled in days. For the general audience, this is a clear example of how AI accelerates fundamental mathematics and theoretical physics, not just applied tasks. However, the formulation 'AI solved a Millennium Prize Problem' should be treated with restraint: the generation method has not been disclosed, and the discussion of the news on Hacker News is almost empty—one point and one comment.

What is still unknown / limitations

The generation methodology remains unknown: neither the model, nor the method, nor the volume of computation, nor the share of human editing behind the million lines has been disclosed anywhere, so the conclusion that 'LLM proved a Millennium Prize Problem' cannot yet be considered reproducible. The committee's verdict was formulated cautiously—'probably correct'—and concerns primarily the meaningfulness and accuracy of the formalization; the step-by-step correctness of the code is guaranteed by the Lean kernel only if the artifact actually compiles without hidden assumptions. The full article written by humans has not yet been released, there is virtually no independent technical breakdown by the community, and the disclosure of the methodology and verification of reproducibility are matters for the near future. Until these gaps are closed, the correct formulation of the status is: a strong formal artifact with preliminary human approval, not a closed and reproduced work.

Sources

Author

Look at AI, editorial team