OpenAI Proved Half of Yang-Mills. The Missing Half Is the One Pauli Would Not Drop.

In the spring of 2019, Yang Chen-Ning, then ninety-six, sat in Tsinghua and told visitors a story from 1954. He and Robert Mills had calculated a result that was “very, very beautiful,” but it contained a wound they could not close. According to their theory, there should exist a particle with charge but no mass. Nature contained no such thing. They agonized over it and decided to publish anyway. Yang gave two reasons. First, the structure was too beautiful to abandon. Second, field theory itself had not yet received a precise mathematical description. It was plagued by divergences, and yet “this did not prevent it from achieving miraculous success in the renormalization program.” Then he added: “Pauli chose not to publish.”

Wolfgang Pauli had reached a similar structure earlier. He had interrupted Yang’s seminar at Princeton with the fatal question: what is the mass of these gauge fields? Yang could not answer. Pauli could not persuade himself that nature would tolerate massless charged particles, so he left the result in a drawer. Yang and Mills took the other road. They wrote down the beautiful structure first and left the wound exposed.

Seventy-two years later, on October 26, 2026, OpenAI announced that its internal model had produced a rigorous mathematical existence proof for four-dimensional Yang-Mills theory. Chinese commentary quickly summarized the news as “filling in Yang’s old regret.” That phrase catches part of the story, but it slides past a central point. Yang’s regret was never a single thing. What OpenAI filled in is the part closest to mathematical rigor, and only half of that.

To understand the weight of that half, start with where Yang-Mills theory sits in twentieth-century physics, and why it remained mathematically suspended for so long.

The equations Yang and Mills wrote down in 1954 demand a nearly unreasonable symmetry. Physical laws must remain unchanged when an internal transformation is performed independently at every point of spacetime. This requirement of local autonomy forces mathematics to grow a new field, and the particles that transmit that field carry charge themselves. They radiate and split. At the time, the structure looked like a mathematical toy with no physical counterpart. It later became the foundation of the Standard Model of particle physics. The electroweak unification of 1979, the confirmation of asymptotic freedom in 2004, and the establishment of the Higgs mechanism in 2013 all rest on this set of equations. The official background note for the 1999 Nobel Prize stated: “In 1954, Yang and Mills constructed the first example of a non-Abelian gauge theory.” Physicists used these equations to calculate numbers that matched experiment to ten decimal places.

But calculational precision and mathematical existence are two different things. The quantum field theory in physicists’ hands is like a vending machine. Insert a coin, press a button, and out comes a result astonishingly precise. Yet no one can open the casing and prove that a legitimate mechanical structure is actually running inside. In continuous four-dimensional spacetime, a field has infinitely many degrees of freedom. The renormalization operation physicists habitually perform, subtracting infinity from infinity, has never received a rigorous mathematical foundation. What Yang meant in 2019 when he said that field theory “has not yet received a precise mathematical description” points to this gap. A theory that is physically precise to a humbling degree has not yet been proved, mathematically, to be a legitimate object.

In 2000, the Clay Mathematics Institute formally named this gap “Yang-Mills Existence and Mass Gap” and attached a one-million-dollar prize. The problem contains two layers. The first is constructive existence: rigorously construct, in continuous four-dimensional spacetime, a nontrivial Yang-Mills field theory satisfying the axioms of quantum field theory. The second is the mass gap: prove that the energy spectrum of this theory has a strictly positive gap above the vacuum, thereby explaining mathematically why the strong interaction is short-ranged and why quarks and gluons are permanently confined inside hadrons. The first asks whether the thing can be built in mathematics. The second asks whether, once built, its lowest energy excitation has a positive lower bound.

OpenAI’s model advanced along a specific technical route, and the route is not new. In the 1980s, the Polish mathematical physicist Tadeusz Balaban developed a multiscale renormalization-group construction intended to control the infinities in field theory through successive block-spin-like scale transformations, pushing a lattice theory step by step toward the continuum limit. Balaban completed a great deal of foundational work at the ultraviolet end, the limit in which the lattice spacing goes to zero, but never closed the entire proof chain in four dimensions. OpenAI claims that its model completed the second half Balaban did not finish, extending the argument from finite volume to infinite volume and obtaining, in continuous four-dimensional space, a rigorously existing gauge field with self-interactions.

From the information disclosed so far, what this proof accomplishes is the first layer of the Clay problem, and it does so in the ultraviolet sense. It argues that as the lattice spacing tends to zero, the theory does not collapse under ultraviolet divergences, and that a legitimate continuum-limit object exists. But the second layer of the Clay problem, the infrared mass gap, is not among the results OpenAI announced. The model proved that the clock can be built. It did not prove that, when struck, the clock has a lowest note, or that the lowest note is greater than zero. Put more precisely, OpenAI filled in half of Yang’s regret about field theory lacking a precise mathematical description. The wound Yang had already walked around in 1954 with physical intuition, where mass comes from, has still not been directly touched by this proof in the strict mathematical sense.

There is a subtler detail that reveals exactly where the boundary of this proof is drawn. OpenAI’s announcement covers not only pure Yang-Mills theory but also quantum chromodynamics with fermions, and even a unified completion of the Standard Model, yet it explicitly excludes the original, un-unified Standard Model. The reason is asymptotic freedom. In pure non-Abelian gauge theory and in QCD, the coupling constant tends to zero at short distances, so the Gaussian principal part in Balaban’s multiscale method can control the nonperturbative remainder, and the ultraviolet limit can converge mathematically. The un-unified Standard Model contains an Abelian U(1) hypercharge gauge field whose coupling constant does not decrease at short distances but instead grows, hitting a Landau pole in the ultraviolet. In other words, the proof could close precisely because it dealt with theories that run increasingly free. The most original form of the Standard Model is, mathematically, the hardest for this toolkit to tame. The detail also marks the physical territory of the result.

What makes this announcement more than a technical advance is the politics of verification it exposes. In announcing three sets of results, OpenAI gave them very different levels of public access. In mathematics, the local smoothness conjecture, the Kakeya conjecture, and the Stein restriction conjecture, the preprint and Lean formalization code were posted directly on GitHub for anyone to inspect. In computer science, the refutation of the Exponential Time Hypothesis, materials were made public, and the original proposer of the hypothesis was invited to participate in review. But the physics proof alone was handed to a five-physicist group to digest first, led by the 2004 Nobel laureate David Gross, with the original manuscript and code not yet public. OpenAI’s explanation was that the material was vast and needed expert organization and understanding before a decision could be made about publication.

The arrangement drew open opposition. A nine-member group of mathematicians called AGMAI, some of whom had been approached by OpenAI and invited to consult, issued a statement demanding that OpenAI immediately release the proofs for Yang-Mills, QCD, and the Standard Model completion, including both natural-language and Lean versions. Their reasoning was direct and sharp. Lean formalization had been completed on October 16, line-by-line verification was over, and from a scientific standpoint there was no longer any reason to delay. “Now that verification is complete, there is no scientific reason for further delay.” AGMAI also noted that the Yang-Mills result has obvious major significance for mathematics. It does not belong only to the physics community.

The dispute touches a question deeper than whether the proof is correct. When the producer of a mathematical-physical result is a private AI company, and verification is placed inside a small expert group selected by that company, where does the public scrutiny on which the scientific community depends go? Lean code is a promise of verifiability, but what it verifies is whether the formalized statement and the proof steps are consistent. It does not verify whether the formalized statement faithfully translates the mathematical proposition that actually matters. In the absence of the original physics proof materials, outsiders cannot even make the latter judgment. OpenAI frames the AI output as raw research data, separating generating a proof from forming scientific understanding. That distinction has epistemological merit, but institutionally it creates a vacuum. Who guarantees that the process of understanding has not quietly reshaped the boundaries of the data?

Return to Yang. In 1954 he chose to publish, accepting a situation in which physical intuition ran ahead of mathematical rigor. He believed the beautiful structure would eventually find its place, and that precise mathematical description could wait. The wait lasted seventy-two years. If OpenAI’s proof survives public scrutiny, it will mark the beginning of the end of that wait in the dimension of existence. The equations physicists have used for more than half a century will finally have been proved, mathematically, to have something there. That is concrete progress and needs public scrutiny, not dismissal with the phrase “only half proved.”

But the phrase “half” sticks because Yang’s regret was never touched in another dimension. In the same 2019 conversation he said: “Among the problems we have not solved, the most important is how to bring gravity in and unify it with the existing model.” What OpenAI proved is the existence of gauge fields and fermions. Gravity is not in it. And the mass gap, the requirement that the lowest note be greater than zero, is not in it either. An existence proof tells you the clock can be built. The mass gap tells you that when struck, it must sound, and that the sound cannot sink to zero. Yang gambled on beauty, betting that the physical world would accept a structure not yet mathematically secure. OpenAI’s model, seventy-two years later, answers that the structure does hold mathematically. It does not answer why it sounds when struck. That latter question may be the one Pauli would not let go.