OpenAI claims three scientific breakthroughs and gives experts different review roles
Public proof materials allow immediate scrutiny in mathematics and computing. Physics collaborators will prepare their own account before releasing the AI output.
Loading page…
Public proof materials allow immediate scrutiny in mathematics and computing. Physics collaborators will prepare their own account before releasing the AI output.
Listen to this story
OpenAI says its systems produced results spanning mathematics, theoretical computer science and physics, but the claims are at different stages of review. Mathematics and computer science materials, including Lean-checked proofs, are public; for the physics work, David Gross’s group is preparing an independent writeup before the artifacts are released. The physics claim includes an existence proof for pure Yang–Mills theory, not a solution of the full Millennium Prize problem. The split review process means outside researchers will have unequal access to the claims at first.
OpenAI says its models proved the local smoothing conjecture in every dimension, also resolving the Kakeya and Stein restriction conjectures.
The computer science result challenges the Exponential Time Hypothesis, which researchers have used to establish conditional limits across many fields.
Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine and explain the computer science result; practical algorithmic gains remain uncertain.
OpenAI is claiming three scientific breakthroughs while calling the underlying manuscripts raw research data, not finished scientific understanding. Its October 26, 2026 announcement pairs results in mathematics, computer science and physics with different routes to scrutiny: public proof materials for the first two, and an expert-led writeup before releasing the physics artifacts.
The mathematics result concerns the local smoothing conjecture: the idea that the most extreme focusing of waves must be short-lived. OpenAI says its models proved it in every dimension, also resolving the Kakeya and Stein restriction conjectures in all dimensions.
Following an external expert’s recommendation, the company published its AI-generated writeup and a proof formalized in Lean, leaving researchers to organize their own explanatory efforts. The computer science release combines public proof materials with a designated group to verify and explain the result.
That result challenges the Exponential Time Hypothesis, a conjecture about the difficulty of finding solutions to Boolean logic formulas. Researchers have used it in hundreds of papers to derive conditional limits on computing tasks across fields including databases, graph theory and computational biology.
OpenAI says its system disproved the hypothesis and that Lean verified the result. Russell Impagliazzo and Ramamohan Paturi, both original proposers, will lead the examination alongside Ryan Williams. Finding faster algorithms and translating theoretical gains into practical improvements will require significant further research, the company cautions.
choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information.
OpenAI, October 26 research announcement
The physics claim concerns rigorous mathematical existence proofs for certain interacting quantum field theories in four dimensions. OpenAI says these include quantum chromodynamics, which describes strong-force interactions involving quarks and gluons, and a unified completion of the Standard Model of particle physics.
The work also includes a claimed existence proof for pure Yang–Mills theory—part of the Clay Mathematics Institute’s Quantum Yang–Mills Millennium Prize problem, not a claim to have completed the entire problem.
David Gross will lead a group including Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. According to OpenAI, these collaborators chose to withhold the AI-produced artifacts while preparing their own writeup. The company says it deferred to them on publication.
OpenAI’s stated rules give collaborators freedom to reach and publish their own conclusions:
Story updates
Three computer scientists helping organize an independent review of AI-generated proof claims are divided over what the work means for their profession. In interviews published by AFP on October 30, Russell Impagliazzo, Ramamohan Paturi and Ryan Williams described surprise, vindication and regret after OpenAI and Anthropic claimed their models had overturned a longstanding computing hypothesis.
Three computer scientists helping organize an independent review of AI-generated proof claims are divided over what the work means for their profession. In interviews published by AFP on October 30, Russell Impagliazzo, Ramamohan Paturi and Ryan Williams described surprise, vindication and regret after OpenAI and Anthropic claimed their models had overturned a longstanding computing hypothesis.
The exponential time hypothesis, or ETH, concerns how quickly the work needed to solve certain difficult logic problems grows as more variables are added. It holds that even the fastest algorithm requires exponentially more time. The companies claim their models found algorithms that run in sub-exponential time—a slower rate of growth than the hypothesis permits.
Impagliazzo and Paturi, both at the University of California San Diego, first proposed ETH; Williams, at MIT, has been a prominent skeptic. Hundreds of studies in computer science and biology rely on the hypothesis. The result remains a claim under review, not an independently established refutation.
For Paturi, the first response was surprise, tempered by the view that research built around ETH would not simply disappear. Williams felt vindicated because he had publicly doubted it. Impagliazzo was less cheerful: he would have preferred a human discovery, with a researcher who could explain the thinking behind it and whose hand he could shake.
The review involves more than reading a proposed answer. In an October 27 TechCrunch interview, Williams described two or three parts of OpenAI’s proof as “alien-looking,” saying they demanded work by the whole community. The researchers’ first step was an open letter inviting scientific feedback and discussion of norms for AI contributions.
Williams also pushed OpenAI to support a conference. His concern was not only understanding the algorithms: if the claims hold up, researchers will need to revisit work that relies on ETH and decide how the field’s foundations should change.
In the AFP interviews, Paturi drew a distinction between publishing a proof and sharing the means of discovery. Concentrated access to powerful research tools, he said, raises questions about who can participate, who gets an early advantage and who sets the research agenda. Public release of a proof does not give researchers equal access to the system that produced it.
A company's announcement cannot substitute for the community's assessment.
The scientists’ advice to students exposes their sharpest difference. Paturi encouraged young people to enter the field, learn the tools and develop their own judgment. He believes humans can still shape the research frontier. Williams described a “major revolution” that could motivate some people to join even as others leave.
Impagliazzo worried that AI could take over much of the solving, leaving humans to verify and explain results. Researchers previously did both, he noted. His uncertainty extends beyond that division of labor: with AI changing so quickly, he said he had no idea what research would look like next year.
For people outside the field, Impagliazzo expected no immediate consequences from the theoretical result. Williams saw the possibility of more efficient computing; Impagliazzo suggested later improvements could yield better scientific data-analysis tools. He said the current result posed no security threat, while cautioning that future extensions might require stronger encryption.
A committee of physicists and mathematicians says OpenAI has produced a proof that quantum Yang–Mills theory exists, including interactions with matter particles called fermions. But its statement sets a condition before public release: the experts must first check that the proof is correct and addresses the accepted mathematical and physical problem.
A committee of physicists and mathematicians says OpenAI has produced a proof that quantum Yang–Mills theory exists, including interactions with matter particles called fermions. But its statement sets a condition before public release: the experts must first check that the proof is correct and addresses the accepted mathematical and physical problem.
The committee’s October statement follows OpenAI’s October 26 announcement of the physics work. The company says its models generated a natural-language proof of Yang–Mills existence on September 9, 2026, followed by a proof for quantum chromodynamics, or QCD, on September 18. It says the Yang–Mills formalization in Lean was completed on October 16.
On October 16, OpenAI asked David J. Gross to take custody of the proof and organize its release. Gross then selected Michael R. Douglas, Arthur Jaffe, Igor R. Klebanov and Nikita Nekrasov to join the committee. The group describes its role as serving the academic community while reducing risks such as publishing incorrect results or overlooking attribution.
OpenAI says its AI has overturned a longstanding assumption about computational difficulty, but the formal proof leaves key running-time constants unspecified. Its October 26, 2026 announcement includes a public paper and computer-checkable proof files. Outside review has begun; neither an independently accepted result nor a practical speedup follows from publication.
OpenAI says its AI has overturned a longstanding assumption about computational difficulty, but the formal proof leaves key running-time constants unspecified. Its October 26, 2026 announcement includes a public paper and computer-checkable proof files. Outside review has begun; neither an independently accepted result nor a practical speedup follows from publication.
The Exponential Time Hypothesis, or ETH, concerns a logic problem called 3-SAT. It asks whether true-or-false values can satisfy every condition, each involving three variables. ETH says even the best algorithm needs time growing as two raised to some fixed fraction of the variable count, in the worst case.
OpenAI’s paper instead claims a subexponential algorithm: over sufficiently large inputs, that fraction shrinks toward zero. It does more than determine whether a solution exists. It counts every satisfying assignment for formulas with any fixed number of variables per condition, at least two, using memory bounded by a fixed power of input size.
Loading discussion...
OpenAI describes participation in these working groups as voluntary and unpaid, with members reaching their own conclusions. Participants are not required to sign nondisclosure agreements and may discuss the work, share intermediate products and publish their findings. The committee makes its own position explicit:
We have no responsibility to OpenAI.
Quantum field theory combines quantum mechanics with systems containing infinitely many degrees of freedom. QCD, the Yang–Mills theory coupled to quarks, describes the strong nuclear force. The claimed advance concerns putting such theories on a rigorous mathematical footing, rather than merely obtaining numerical evidence for their behavior.
The committee traces the work to decades of research. Kenneth G. Wilson’s lattice approach represents the theory on a grid. Renormalization-group methods repeatedly combine parts of that grid to study behavior at larger scales. The goal is a continuum limit: a theory that no longer depends on the grid used to define it.
According to the committee, methods developed by Tadeusz Balaban in the 1980s were important ingredients in OpenAI’s claimed construction of that limit. Its statement nevertheless draws a firm boundary: confinement of quarks within hadrons and the mass gap remain open challenges. The existence claim does not settle those questions.
OpenAI is supplying a formalization in Lean, an interactive theorem-proving language. The committee’s first task goes beyond having that formal artifact: it will check whether the definitions and problem statement match standard physics and mathematics, and whether the formalization matches the written argument. Those checks are still ahead, not a completed endorsement.
After the initial validation, the committee plans to release the Lean proof. It will then explore the argument more deeply and investigate how it might advance quantum field theory.
Once satisfied with the proofs’ quality and correctness, the group intends to publish its assessment alongside the full 1,147-page collection of AI-generated proofs.
The committee calls the manuscripts’ length and complexity formidable and says understanding their methodology and meaning will take time. It intends to enlist other researchers in that work. Its proposed next step is therefore not simply distributing a long document, but developing an account that the wider physics and mathematics communities can examine and use.
The proof files use Lean, a language for writing mathematical arguments that software can check. A successful check establishes the encoded statement; experts must still confirm that it expresses the intended mathematical claim.
AI News Report’s technical breakdown identifies another limit. The Lean theorem establishes a weaker bound than the paper’s headline bound, though still sufficient to refute ETH if valid. Its constants are not specified, leaving no usable running-time guarantee for an actual input.
The outlet estimates that the paper’s general bound beats the cited best-known 3-SAT algorithm only beyond about 540 million variables, assuming every hidden constant equals one. That is an illustration, not measured performance; the real constants could be much larger.
MIT computer scientist Ryan Williams is organizing the examination with UC San Diego researchers Russell Impagliazzo and Ramamohan Paturi. The group requested public proofs and an OpenAI-funded conference. Williams described unfamiliar elements of the proof in his TechCrunch interview.
there are two or three alien-looking parts
If the refutation holds, papers relying on ETH may need revisiting, Williams said. Their conclusions about computational difficulty would lose that assumption as support. The claim does not resolve P versus NP: whether every problem whose answer can be checked quickly can also be solved quickly.
Cryptographic security was an early concern. Williams said OpenAI told him it had checked and was reasonably confident the result implied no breaks. That is the company’s assessment, relayed by Williams—not an independent security finding.
Join the conversation
Explain when early disclosure helps science—or gets ahead of scrutiny.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.