A 250,319-Declaration Lean Project Complicates Review of OpenAI’s Yang–Mills Proof
OpenAI’s proposed proof for the Yang–Mills existence problem has passed Lean’s formal checks, but the mass-gap problem remains unsolved and human review is still underway. In Two Minute Papers, Zsolnai-Fehér says the construction builds on earlier work to control quantum fluctuations across increasingly fine grids; he also warns that a large, difficult-to-navigate codebase will make the proof challenging for the scientific community to assess.

The proposed proof tackles existence, not the mass gap
The Yang–Mills existence and mass gap problem asks whether the mathematics describing the strong nuclear force remains well-defined at arbitrarily small scales, and whether the theory has a mass gap. Károly Zsolnai-Fehér explains that physicists have equations describing the force with great accuracy; the unresolved challenge is proving that the theory holds together as we zoom in. The problem is one of the Millennium Prize Problems, with a million-dollar prize attached.
The approach begins with a discretized model: represent space and time as a grid, then calculate how the force behaves along the connections between grid points. That grid is an approximation, not reality. Making it progressively finer should bring it closer to a continuous limit. Showing that the theory works at that limit would be major progress toward solving the problem.
But refinement does not automatically make the approximation more manageable. In Balaban’s approach, finer grids reveal more quantum fluctuations—changes in the force field at smaller scales. Those fluctuations make the mathematics harder, and the calculations can spiral out of control. The question is whether the theory settles into a meaningful answer as the grid becomes infinitely fine.
Tracking multiple scales is meant to control the fluctuations
OpenAI’s proposed construction builds on Balaban’s approach but tracks multiple grids at once. Instead of examining one scale in isolation, the system calculates how tiny changes on the finest grids affect the coarser ones. Following those effects across scales lets it calculate their combined impact. Zsolnai-Fehér says this shows that the mathematics does not spiral out of control as the grid is refined.
The construction addresses the existence part of the problem. It does not solve the mass-gap part, which remains open.
OpenAI formalized the proof in Lean, a language that allows mathematical logic to be checked step by step. Zsolnai-Fehér reports that the checks passed. The next question is whether the formal proof addresses the intended problem: do its definitions and statement match the standard ones used in physics and mathematics, and does the formalization align with the informal treatment?
A review group led by physicist David Gross is examining those questions. Zsolnai-Fehér describes the initial results as encouraging, but says the review is still underway. The displayed review text identifies checking the definitions, problem statement, and their correspondence to the informal treatment as the group’s first objective.
The repository’s scale complicates human review
The Lean checks passing does not remove the need to inspect the formalization. The review document cited in the source counts 250,319 project declarations, including automatically generated auxiliary declarations. It says 100,168 are used in the dependency tree of the main theorems. The document also notes that the difference is less drastic for source declarations, which exclude auxiliary ones.
The same document flags unconventional names, including “Farewell.dead,” “Cake,” “cookie,” and “anxiety.” The names are memorable, but the concern is practical: a large codebase with many declarations outside the main theorems’ dependency tree can be difficult to understand and review.
Zsolnai-Fehér says unused code does not mean the proof is wrong. It does, however, add to the challenge of making the formalization intelligible. He describes the submission as “a bunch of spaghetti” left for the scientific community to untangle, while also saying he is excited by the prospect of new knowledge. His criticism is about the burden of review, not a claim that the Lean checks failed.
The proposal concerns Yang–Mills existence; the mass-gap problem remains unsolved, so the Millennium Problem has not been fully solved. Human review of the proposed proof is still in progress.
