A rumor about an unpublished OpenAI result changed the publication schedule of three MIT researchers before the AI proof was available to inspect. Dor Minzer and graduate students Yumou Fei and Shuo Wang heard on September 11 that a company result on Unique Games was approaching. On September 14 they posted a rough 95-page paper on a related but different theorem. OpenAI announced its mathematical collection on October 6, including a claimed proof of the Unique Games conjecture.
The chronology is a research event, not simply an AI leaderboard result. The human team's key breakthrough had come in April; they were still turning it into an explanatory manuscript. A nearby corporate claim made establishing public priority more urgent than completing that exposition. Neither theorem makes the other redundant, and neither preprint's appearance substitutes for independent expert review.
Two results, not two versions of the same proof
Subhash Khot introduced Unique Games in 2002. Picture vertices connected by edges, each vertex receiving a label. Each edge has a permutation constraint: knowing one endpoint's label determines exactly one acceptable label at the other. That one-to-one rule gives the game its name.
The conjecture concerns the computational difficulty of approximation. Informally, it is hard to distinguish instances where almost every constraint can be satisfied from instances where only a very small fraction can. It does not say satisfying assignments never exist. It concerns what efficient general algorithms can establish across carefully defined families of inputs.
The distinction affects optimization. Researchers often accept an approximate answer when an exact optimum is too expensive to compute. Unique Games connects many best-known approximation guarantees to underlying hardness barriers. Raghavendra's 2008 framework, conditional on the conjecture, establishes the optimality of a standard semidefinite-programming strategy for broad constraint-problem classes. The method solves a geometric relaxation and rounds it into discrete choices.
The MIT theorem addresses perfect completeness: a starting instance whose constraints can all be satisfied, not merely almost all. In a 2-to-1 game a selected label can allow two compatible choices on the other side; the team's 4-to-1 theorem allows four. It is weaker than proving the original 2-to-1 statement, but reaches an exact-completeness setting not directly covered by ordinary Unique Games.
Why “all” rather than “almost all” matters
Earlier work by Minzer and collaborators reached almost-perfect completeness in 2018. A small remaining gap can still prevent reductions that require a completely satisfiable starting point. Reaching exactly that endpoint is therefore not a cosmetic improvement.
Fei and Wang began working with Minzer in 2025. A newer error-correcting code offered a promising ingredient, but fitting it to the rest of a hardness reduction repeatedly failed. The April breakthrough combined quadratic equations, a middle verification layer and an inner verifier using Grassmann-style encoding.
These are probabilistically checkable proof techniques. A verifier samples a small number of positions in a long encoded proof. A hardness reduction must preserve the gap between cases that should be accepted and those that should be rejected. The interaction among layers, not just the existence of a code, is what makes the construction consequential.
One consequence concerns graph coloring. Even with a promise that a graph is three-colorable, finding a valid coloring using any fixed number of permitted colors remains NP-hard under the theorem. More allowed colors do not necessarily provide an efficient general method. The result also has consequences for independent sets in certain hypergraphs, whose edges can connect more than two vertices.
That is a distinct contribution from OpenAI's claim about the usual Unique Games conjecture. The MIT authors were protecting a separate route and its consequences, not rushing a duplicate of the company's manuscript.
The cost of publishing three days after a warning
Minzer's team said the mathematics was complete but the presentation was not what they wanted to release. The September 14 ECCC preprint contained 95 pages. Minzer described later sections, after Section 6, as almost without connecting words: definitions and intermediate proofs remained, but guidance for readers was sparse.
Exposition is part of research infrastructure. Motivation, transitions, comparisons and explanations help others audit a construction and reuse its ideas. The team could establish a timestamp quickly while still owing the community that work.
They did not know the company's argument, model, precise claim or schedule. The asymmetry was therefore not an ordinary race between two groups following comparable human timelines. Anticipated machine output redirected human effort before specialists could judge it.
OpenAI's release created a second review problem: volume
OpenAI said its internal model attempted roughly 4,000 problems and released 722 manuscripts covering 377 questions. One collection contained 40 theoretical computer science proofs. These figures describe different units—attempts, manuscripts, questions and proofs—and should not be treated as interchangeable counts of settled problems.
The company reported average compute equivalent to about three hours of ChatGPT Pro thinking per result. An average across the release is not an exact cost for Unique Games or a uniform budget for every proof. About half the results had formal verification at announcement, according to the company; its expectation of formalizing the rest was forward-looking.
The Unique Games manuscript claims an explicit deterministic polynomial-time reduction from 3SAT to unweighted instances, with formal artifacts in Lean. It lists implications for cut, covering, ordering, deletion, clustering and constraint-satisfaction problems. Those implications depend on earlier reductions as well as the new theorem.
OpenAI did not present the release as conventional peer review. It acknowledged the need for better references and exposition and said it would support understanding important generated results. Public outputs can be substantial evidence without being a finished community assessment.
Lean checks a formal statement, not every scientific question
Lean verifies that encoded steps follow from encoded assumptions and rules. That is much stronger evidence than a model asserting that its prose is correct. But experts must still compare the encoded theorem with the intended conjecture, inspect imported assumptions and connect the formal code to the manuscript.
Unique Games is sensitive to quantifier order and to how alphabet size depends on the error parameters. A superficially similar statement can be weaker if those dependencies change. Completeness, soundness, explicitness and polynomial runtime must all mean what the standard complexity claim requires.
A checked derivation also does not automatically explain the decisive idea, identify novelty, credit prior work or establish community consensus. Verification and understanding are related stages, not synonyms. The rushed human paper also needs expert reading and improved exposition; human authorship does not remove that obligation.
Reproducibility likewise has layers. Available Lean artifacts and dependencies can let outsiders repeat checking. The unnamed internal model was not publicly available, and selected summaries and aggregate statistics did not provide every prompt, sampling choice and intervention. Reproducing verification is therefore different from reproducing discovery.
The consequential next steps are independent alignment checks of OpenAI's claim, a clearer MIT manuscript and release practices that make machine-scale output reviewable. The event has already changed when researchers publish and how they defend priority. Whether it advances mathematics more durably will depend on turning checked outputs into explanations other researchers can examine, adapt and understand.