According to OpenAI, the company's AI models have produced proofs for three problems in mathematics, computer science and theoretical physics that have stood unsolved for decades.
According to one of the proofs, a hypothesis about computing time that hundreds of research papers build on does not hold.
The researchers who formulated the hypothesis are now reviewing the proof, without pay and without nondisclosure agreements.
OpenAI has presented proofs for three problems that researchers have worked on for decades. One is in mathematics, one in computer science and one in theoretical physics. The proofs were produced by the company's internal AI models.
OpenAI calls the manuscripts raw research output that still needs to be understood. That is why the company has asked independent experts to review and explain the results. Each field has been given its own process.
The proofs have also been translated into the programming language Lean. Lean confirms that a proof matches the statement as it is written. Experts still have to decide whether the statement captures the problem one wants to solve. For two of the proofs, the computer check in Lean was done the day after the AI produced the proof.
Earlier in October, OpenAI published 372 new mathematical results from an internal AI model, which we wrote about.
A hypothesis from the 1990s may be wrong
The most talked-about proof concerns the Exponential Time Hypothesis, often called ETH. It is about how long it takes a computer to solve a certain type of logic problem. The task is to decide whether a long list of conditions can all be met at the same time. The hypothesis says that the time it takes in the worst case grows explosively as the problem gets bigger.
Russell Impagliazzo and Ramamohan Paturi at the University of California, San Diego, formulated the hypothesis together with Francis Zane. It has been widely accepted since the late 1990s. Hundreds of research papers in fields such as graph theory, computational biology, databases and computational geometry build on it.
OpenAI's 67-page manuscript describes an algorithm that counts how many solutions such a logic problem has. The algorithm does this faster than the hypothesis allows. If the proof holds, the papers that build on the hypothesis lose their proven foundation. Their results do not automatically become wrong, though.
The result does not settle the P versus NP question, one of the best-known unsolved problems in computer science.
The people behind the hypothesis are reviewing the proof
Impagliazzo and Paturi are now reviewing the proof together with Ryan Williams, a professor at MIT. They are working without pay and have not signed any nondisclosure agreements. They will share what they find as the work goes on.
Williams has described the result as shocking. He says a confirmed disproof would be a major event, since much of theoretical computer science rests on the hypothesis. According to him, the proof suggests a new method that could solve many fundamental problems more efficiently. No one knows yet how much this will matter in practice, according to Williams.
OpenAI is paying for a conference on what the result means for the research field.
Waves in every dimension
The mathematics proof concerns the local smoothing conjecture. Christopher Sogge formulated it in 1991. It is about how waves behave, for example sound, light and seismic waves. A wave can briefly become much stronger when it focuses at one point. Sogge asked how strong such a wave can be on average over a short time.
Until now, there was only a proof for waves in a plane, with two dimensions. It came in 2019 from the mathematicians Larry Guth, Hong Wang and Ruixiang Zhang. OpenAI's 59-page proof covers all dimensions, including the three-dimensional space we live in.
According to OpenAI, several other well-known problems follow from the result. One is the Kakeya conjecture in geometry, which Hong Wang and Joshua Zahl proved for three dimensions in 2025. With the new proof, it would hold in all dimensions. The same goes for Stein's restriction conjecture in Fourier analysis.
Mathematician Alex Iosevich at the University of Rochester calls it a historic event. Sogge was his advisor when he did his PhD. Iosevich expects AI to change theoretical mathematics a great deal.
There is no dedicated review group for the mathematics proof. An outside expert advised OpenAI to publish the material right away, so that researchers can choose for themselves how they want to check it.
A Nobel laureate reviews the physics
The third proof concerns quantum field theory, the theory that describes the smallest building blocks of nature. According to OpenAI, the proof shows that certain quantum field theories in which particles interact with each other can be formulated in a mathematically rigorous way in four dimensions, meaning three dimensions of space plus time.
This includes Yang-Mills theory and quantum chromodynamics, the theory of the strong force between quarks and gluons. Yang-Mills theory is linked to one of the Millennium Prize Problems set by the Clay Mathematics Institute, with a prize of $1 million for whoever solves it. The problem has two parts. OpenAI says the proof solves the first part, that the theory exists mathematically. The second part, known as the mass gap, remains.
According to OpenAI, the proof does not change the predictions of physics. Instead, it would give quantum field theory a rigorous mathematical foundation.
The review is led by David Gross, who received the 2004 Nobel Prize in Physics for research on how the strong force weakens at short distances. The group also includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. Jaffe helped formulate the Millennium Prize Problem on Yang-Mills. The group is holding off on releasing the AI material and is first writing its own account.
The researchers get the credit
OpenAI does not claim any scientific credit for the results. The credit should go to the groups that explain them.
OpenAI is also helping fund a $15 million effort together with the Simons Foundation and Anthropic. The money will go to workshops and projects, among other things, that help mathematicians evaluate results produced with AI.
WALL-Y
WALL-Y is an AI bot created in Claude. A human selects the news, WALL-Y writes the text and a human checks the results before publishing.



