Research

OpenAI Says Its Model Refuted the Exponential Time Hypothesis. The Proof Is the Easy Part.

OpenAI claims results in harmonic analysis, complexity theory and quantum field theory, with Lean-checked proofs for two of them. External groups still have to decide what the claims mean.

Evan BrooksEvan Brooks
Heat: 1,620
OpenAI Says Its Model Refuted the Exponential Time Hypothesis. The Proof Is the Easy Part.

OpenAI published three claimed results on October 26, in harmonic analysis, complexity theory and quantum field theory. Two came with machine-checked proofs, and the loudest objection is that a checked proof isn't the same thing as accepted mathematics.

Three OpenAI Claims and Three Different Release Rules

OpenAI announced the results in one post, and the process is as much the story as the mathematics. The company asked outside experts to act as shepherds, then followed three separate release plans.

The math result went public immediately as model output plus a Lean proof. The computer science result came out alongside an external group that has committed to explaining it in public. The physics result is being held while a group led by David Gross writes its own version. Three results, three plans.

OpenAI calls the manuscripts a form of raw research data. That framing is honest about what a language model produces, and it quietly moves the work of interpretation onto people who didn't run the model.

What the Exponential Time Hypothesis Claims

ETH is a guess about the difficulty of Boolean satisfiability, the problem of deciding whether a logical formula can be made true. In plain terms, brute force is close to the best anyone can do.

It matters because hundreds of papers use ETH as a starting assumption. If ETH holds, then a given problem can't be solved faster than some bound. That move shows up in graph theory, databases and computational biology, and pulling the assumption out takes a lot of results with it.

OpenAI says its system proved ETH is false by giving a deterministic algorithm that counts satisfying assignments for fixed-width formulas in subexponential time and polynomial space. The 3-SAT case does the refuting. The model produced the natural language argument on October 16, and the Lean formalization followed a day later.

Russell Impagliazzo and Ramamohan Paturi, who proposed ETH with Francis Zane, lead the group checking it, with Ryan Williams of MIT alongside. OpenAI agreed to publish all raw data and to fund a conference on what the result means.

Local Smoothing, Kakeya and the Harmonic Analysis Result

The math claim concerns Sogge's local smoothing conjecture from 1991, which says the most extreme focusing of a wave must be short-lived. It's the kind of statement that sounds obvious and implies a great deal, since Kakeya and Stein's restriction conjecture both follow from it.

Progress had been slow, and each result covered a fixed number of dimensions. Guth, Wang and Zhang handled two dimensions in 2019. Wang and Zahl settled three-dimensional Kakeya in 2025. OpenAI says its model proved local smoothing in every dimension using a new auxiliary partial differential equation.

The output arrived October 18 and was formalized in Lean the next day. Both the write-up and the Lean file sit in the openai/math repository on GitHub, where anyone can read them.

Yang-Mills Existence and the QCD Claim

The physics result is the one you can't check yet. OpenAI says its model produced a rigorous existence proof for interacting quantum field theories with gauge fields and fermions in four dimensions, including QCD and a unified completion of the Standard Model.

As a subproblem it also claims existence for pure Yang-Mills theory, which is part of a Clay Millennium Prize problem. The company is careful about scope: no gravity, and the result isn't the Standard Model itself, since that theory can't be extended to arbitrarily short distances.

Gross leads the group, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. Jaffe helped formulate the Millennium problem and co-founded the Clay Mathematics Institute. That's about as credible a panel as this field offers.

The group asked OpenAI to hold the AI artifacts while it writes its own account. For now, this claim rests on OpenAI's description and the reputations of five physicists.

What Lean Verification Can and Can't Check

Lean is a proof assistant, software that checks each step of an argument against a fixed set of rules. When a proof compiles, the logic holds. That's a real guarantee, and stronger than most published mathematics ever gets. So what does a compiled proof prove?

It proves that the formal statement follows from the axioms and nothing else. Agreement among mathematicians is another matter.

It doesn't check that the formal statement matches the conjecture people care about. Translating local smoothing into Lean involves choices, and a slip in that translation produces a perfectly valid proof of the wrong theorem.

Nor does Lean tell you whether a result is interesting, readable or useful. Mathematicians have complained that AI-generated proofs are close to unreadable, which makes the human checking step expensive even after a machine signs off.

The Backlash Over OpenAI's Math Releases

The skepticism didn't start this week. On October 6, OpenAI released more than 700 mathematical documents at once. The Association for Human Mathematics responded that releasing 700 files together amounts to a display of power rather than scholarship, and Terence Tao shared the statement.

A separate advisory panel, AGMAI, said in late September that it doesn't endorse testing advanced math problems on proprietary models and asked OpenAI to stop. The Verge noted that the panel repeated that advice the day before this release.

Early discussion on Hacker News ran skeptical, with much of the argument focused on missing detail and the sheer volume of material dropped at once.

What the OpenAI Science Claims Mean for You

Almost nothing this week, and possibly a lot later. No product change, no API, no price. If the ETH refutation holds up, the practical payoff is a map for building faster algorithms. That work hasn't started.

Two things are worth watching. Whether the Impagliazzo group publishes an explanation that independent complexity researchers accept. And whether the physics group's paper appears with the AI artifacts attached, since that's the real test of a release model running against a field that moves slowly on purpose.

Share This Story

Sources

Related AI News

Concentric spiral curves crossed by a vertical line with bright points pinned along it
Research

An Anthropic Researcher Says Claude Proved the Riemann Hypothesis. Stanford Answered Within Hours

A researcher says he and Claude produced a proof of the hardest problem in mathematics, and the code compiles. Verification is a different job, and it has barely started.

Heat: 1,860
A stack of agent nodes linked by a path to a server rack, with a keyhole and a breach mark on the rack
Research

An AI Agent Found a Zero-Day and Owned a University Server in 11 Minutes

An operator gave the agents one goal and no method. They found an undocumented flaw in medical imaging software and owned the server, after 148 failed actions somewhere else.

Heat: 1,520
A funnel carrying small token blocks from one model box to another, with a shield plate blocking part of the flow
Research

OpenAI Cut Reasoning Theft From 54% to 0.2%. The Rest Is an Industry Problem.

OpenAI published the before-and-after numbers for its latest safeguards and they look like a fix. The same post explains why the fix has to be repeated every week, by everyone.

Heat: 1,420
A green chip grid with one highlighted cell and a jagged signal line beside the words Blackwell fingerprints
Research

Blackwell fingerprints in six Chinese models: what a checkpoint can prove

Dek: PyTorch fills billions of empty weights with random numbers, and the GPU doing the filling leaves a mark in the finished file. Someone checked six Chinese checkpoints and found newer silicon than the rules allow.

Heat: 1,250