OpenAI says its models produced results in mathematics, computer science, and physics, and the company wants outside experts to judge them. According to The Verge, the announcement includes a claimed disproof of the Exponential Time Hypothesis, a proof tied to the Kakeya conjectures, and an existence proof for four-dimensional quantum field theories that include Quantum Chromodynamics (QCD). These OpenAI science discoveries land with an unusual caveat attached: the company is asking academics to check the work before anyone treats it as settled.
That caveat is the real story for readers who follow AI models. A lab is no longer pitching only faster coding or cheaper tokens. It is pitching proofs, in fields where a single wrong step can sink a paper. If you want the pricing and capability backdrop first, our comparison of GPT-6.1 Sol, Sonnet 5.5, and Gemini 4 Argon covers the commercial side. This piece covers the science side, and it separates what OpenAI claims from what outsiders have confirmed so far.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI Announced Across Three Fields
OpenAI announced three results at once. One sits in mathematics, one in theoretical computer science, and one in physics. The mathematics paper is titled “Local smoothing for the Euclidean wave equation.” The computer science paper is titled “The Exponential Time Hypothesis is false.” The physics work is described as a mathematically rigorous existence proof for certain interacting quantum field theories.
The three results share one framing. OpenAI presents them as outputs of its models, then hands them to human experts for scrutiny. Coverage from PANews leads with the Exponential Time Hypothesis, since a disproof there would be the easiest claim for a general tech audience to grasp. Digg carries the story under a title that calls them claimed AI proofs in math, computer science and physics. That choice of word fits the current state of evidence.
The announcement also introduces a process. OpenAI says it contacted external experts to act as “shepherds,” people who organize groups to understand each result. We cover that mechanism in its own section below, because it may matter as much as any single theorem.
The Math Result: Local Smoothing for the Euclidean Wave Equation
The mathematics claim concerns waves. OpenAI says its proof establishes Sogge’s local smoothing for waves in every spatial dimension at least two. Local smoothing is a family of estimates that describe how solutions to the wave equation behave when averaged over time and space. The estimates matter because they control how much a wave can concentrate its energy.
The phrase “every spatial dimension at least two” is the ambitious part. A result that holds in all such dimensions is a different object from one that works in a single case. That breadth is why the downstream consequences are large.
Why Kakeya and Stein Follow From Local Smoothing
OpenAI says the proof yields the Kakeya maximal and dimension conjectures, plus Stein’s spherical restriction estimates, throughout the interior of the conjectured range. Harmonic analysts have long linked these problems. The Kakeya problem asks how small a set can be if it contains a line segment pointing in every direction. Restriction estimates ask how the Fourier transform behaves on curved surfaces such as spheres. Local smoothing sits upstream of both in the standard web of implications.
The qualifier “throughout the interior of the conjectured range” deserves attention. OpenAI’s description does not claim the full boundary cases. Readers should hold that distinction in mind when headlines compress the story into “OpenAI solves Kakeya.” The accurate version is narrower, and OpenAI’s own wording is narrower too.
The Computer Science Claim: Why the Exponential Time Hypothesis Matters
The Exponential Time Hypothesis, or ETH, is one of the load-bearing assumptions of modern complexity theory. It says, roughly, that 3-SAT cannot be solved in subexponential time. Researchers have built hundreds of “conditional lower bounds” on top of it. If ETH falls, many of those conditional results lose their footing, even though they stay true as statements about the hypothesis itself.
OpenAI’s paper gives a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space. The 3-SAT case then refutes ETH. Counting is harder than deciding, so an exact counting algorithm that beats the exponential barrier is a strong form of the claim.
The Lean Verification Detail
OpenAI says the result was subsequently verified in Lean. Lean is a proof assistant, a program that checks each logical step of a formal proof. A Lean check is strong evidence that the formal statement was proved. It does not by itself prove that the formal statement matches the informal claim in the way a human expert would read it. That gap is exactly where the outside group has work to do: confirming that the Lean statement says what the headline says.
The Physics Claim: An Existence Proof for Gauge Theories With Fermions
The physics claim is the boldest. OpenAI describes a mathematically rigorous existence proof for certain interacting quantum field theories that contain gauge fields and fermions in four dimensions. The list includes QCD, the theory of the strong force, and a unified completion of the Standard Model of particle physics.
Physicists use these theories every day to make predictions, and the predictions match experiments with striking precision. What physicists have lacked is a fully rigorous mathematical construction of such theories in four dimensions. An existence proof would fill that hole. Our analysis here is that the claim sits near the territory of the famous Yang-Mills problem, but the announcement as reported does not say it resolves any specific prize problem, and we do not make that claim for it.
One fact limits what anyone can say today. OpenAI says the physics materials were not released in full. Outside readers therefore cannot yet audit the argument end to end. That alone puts the physics result at the back of the verification queue, behind the mathematics and computer science papers that come with fuller write-ups.
Three Results at a Glance
| Field | Title or description | Core claim | Verification step OpenAI cites |
|---|---|---|---|
| Mathematics | “Local smoothing for the Euclidean wave equation” | Sogge’s local smoothing for waves in every spatial dimension at least two | Expert groups organized by shepherds |
| Mathematics (consequences) | Kakeya and Stein results | Kakeya maximal and dimension conjectures, plus Stein’s spherical restriction estimates, in the interior of the conjectured range | Follows from the local smoothing proof |
| Computer science | “The Exponential Time Hypothesis is false” | Deterministic subexponential-time, polynomial-space exact counting for fixed-width CNF, refuting ETH via 3-SAT | Subsequently verified in Lean |
| Physics | Existence proof for interacting quantum field theories | Four-dimensional theories with gauge fields and fermions, including QCD and a unified completion of the Standard Model | Group led by David Gross, materials not released in full |
The Shepherd Model: A New Way to Work With Academics
The second half of the announcement is about process. OpenAI says it contacted external experts to act as “shepherds.” Each shepherd organizes a group of specialists to understand a result, check it, and explain it to the wider field. The company positions this as a new way to work with academics, and The Verge’s headline frames it the same way.
The design addresses a real bottleneck. A machine can emit a long proof faster than any human can read it. Journals and referees run on human time, so the review step becomes the scarce resource. By recruiting known experts before the claims go public, OpenAI tries to turn that bottleneck into a managed pipeline instead of a flood of unreviewed preprints.
There is a tension in the model as well. Shepherds are chosen by the company whose work they examine. Readers may fairly ask how independent the groups are, how they will publish disagreements, and whether a negative finding would carry the same weight as a positive one. The reporting we reviewed does not answer those questions, so we flag them as open.
David Gross and the Physics Shepherd Group
For physics, OpenAI named its shepherds. David Gross will lead a group that includes Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. Gross is a Nobel laureate best known for work on the strong force, which makes him a fitting lead for a QCD claim. Jaffe is widely associated with the rigorous, constructive approach to quantum field theory, the exact discipline an existence proof must satisfy.
The group’s task, per OpenAI, is to shepherd the physics work. That means understanding it first. It does not mean endorsing it. Nothing in our sources says any member has vouched for the result, and readers should not infer endorsement from participation. The company named the group, and the group has not yet published a verdict.
What OpenAI Said in Its Own Words
The company’s announcement page, Working with academic communities to build scientific understanding, carries the key statements. On analysis, OpenAI wrote: “Our models produced a proof of the local smoothing conjecture in every dimension and therefore also resolve the Kakeya conjecture and Stein’s restriction conjecture in all dimensions.”
Note the gap between that sentence and the narrower wording elsewhere, which limits the Kakeya and Stein consequences to the interior of the conjectured range. Both statements come from OpenAI, and a careful reader should expect the technical papers to govern over the summary sentence.
On computer science, the company wrote: “Surprisingly, our system has recently proved that the ETH is false, a result that was subsequently verified in Lean.” The word “surprisingly” is OpenAI’s own. It signals that even the authors did not expect this outcome from the system.
On physics, the same page states: “Our models produced a natural language proof for Yang–Mills existence on September 9th, 2026, and then subsequently produced a proof for QCD on September 18th, 2026.” Two details stand out. The first proof is described as natural language, meaning prose rather than formal code. And the dates show the work is only weeks old, which helps explain why outside review has barely started.
What Is Verified and What Is Not
As of October 27, 2026, independent academic verification of all three results is not clearly established in the cited reporting. That is the single most important sentence for readers who plan to repeat these claims. OpenAI says a Lean check backs the ETH result. It says shepherds are organizing groups for all three. Neither statement equals a completed peer review.
Big claims in any field tend to go through quick scrutiny and sometimes a corrected scope, and nothing in the reporting suggests these will skip that step. Evaluation itself is becoming a craft. Teams that grade model output at scale already borrow statistical tools, as we showed in our tutorial on calibrating an LLM judge with Cohen’s kappa. A proof is a harsher test than a chat answer, but the principle holds: do not trust the grader until you have measured it.
| Question | Mathematics | Computer science | Physics |
|---|---|---|---|
| Result described publicly by OpenAI | Yes | Yes | Yes |
| Machine-checked step cited | Not stated in our sources | Lean verification | Not stated in our sources |
| Full materials released | Not stated as withheld | Not stated as withheld | No, not released in full |
| Named expert group | Shepherds contacted | Shepherds contacted | David Gross, Douglas, Jaffe, Klebanov, Nekrasov |
| Independent confirmation | Unconfirmed | Unconfirmed | Unconfirmed |
Historical Context: Machines and Mathematical Proof
Computers have helped with proofs for decades. Computer-assisted proofs of famous theorems, and proof assistants like Lean, already shape how mathematicians work. What changes here is the direction of travel. In earlier cases, humans proposed the argument and machines checked or extended it. OpenAI says its models produced the argument, and humans are now the checkers.
Complexity theory offers a clean comparison. ETH has stood for roughly two decades as a working assumption, and researchers adopted it precisely because no one could break it. A refutation would be among the most consequential results in the field’s recent history, which is why the bar for acceptance should be high and the review should be slow. The same logic applies to the Kakeya conjectures, which have drawn sustained effort from analysts for a long time.
The surrounding OpenAI story has moved quickly too. Our report on OpenAI resuming frontier training after an 11-day pause, at OpenAI Resumes Frontier Training After 11-Day Pause, shows how much attention the company now pays to controlled conditions for its strongest systems. The science announcement fits a pattern of releasing capability news alongside process news.
Market and Industry Impact
Science claims serve a business purpose, whatever their intellectual merit. Frontier labs compete for enterprise customers, researchers, and investors, and a verified proof is a vivid demonstration that a model can do original work. A false alarm would cost credibility at an awkward moment. Our coverage of OpenAI eyeing up to $30B at a $1.25T valuation shows the capital stakes around the company, and credible science results support the argument that those valuations rest on real capability.
For universities, the announcement poses a staffing question. If AI systems generate proofs faster than departments can read them, review capacity becomes a policy problem. Funding bodies and journals may need new norms for submissions that originate with machines. The shepherd approach is OpenAI’s first draft of such a norm, and the community will likely amend it.
For software and security teams, a disproved ETH would carry practical echoes. Many lower-bound arguments in algorithm design depend on it, so engineers who rely on those arguments to justify design choices may want to track the outcome. The immediate effect on production systems is nil, since an algorithm that runs in subexponential time for exact counting does not, by itself, hand anyone a faster solver for real workloads. We make no claim about practical speedups because the sources do not state any.
Competitive Angle: Science as the New Benchmark
Labs need fresh ways to show what their models can do, and open scientific problems make a vivid candidate. Whichever lab can show a result that experts accept gains a durable talking point. The reverse holds as well: a retracted claim lingers.
OpenAI has also moved into life sciences. Our report on OpenAI buying the four-year-old startup Orchestra Bio shows a company assembling science capability on several fronts at once. The new announcement extends that push into pure mathematics and theoretical physics, domains with no product to sell and everything to prove.
Rivals face a choice. They can answer with their own science claims, or they can stress reliability and cost. Our reporting on Anthropic’s IPO paradox of $16.3B in sales against an $80B-plus loss shows how heavily the leading labs lean on capability narratives to justify spending. A credible proof helps that narrative. An unverified one invites skeptics.
Predictions: What Happens Next
These are our editorial judgments, not reported facts. We keep them qualitative because none of the sources give dates for the review.
- The Lean-backed ETH claim gets scrutinized first. A formal check gives reviewers a concrete target, so we expect the first substantive outside verdict there, probably focused on whether the formal statement matches the informal one.
- The mathematics result gets a scope clarification. Expect experts to pin down exactly which ranges of the Kakeya and Stein conjectures the proof covers, since the interior-of-the-range wording is narrower than the broad summary.
- Physics will lag the other two. With materials not released in full, the Gross-led group cannot publish a thorough assessment until it has the whole argument in hand.
- Other labs will copy the shepherd model. Naming respected experts before publication is cheap and persuasive, so expect similar programs from competitors.
- Journals will publish policy on AI-originated proofs. Pressure on referees makes this close to inevitable, though the timing is unknown.
For ongoing coverage of model releases, benchmarks, and lab announcements, see our AI and machine learning section.
How Readers Should Treat These Claims
Hold two thoughts at once. The claims are specific, technical, and come with named experts and a formal check in one case. That is more than most AI hype offers. They are also unconfirmed by independent parties, and the physics materials are partly withheld. Both statements are true on October 27.
Practical advice follows. Attribute the claims to OpenAI whenever you repeat them. Avoid verbs like “solved” until named outside experts say so. Watch for the shepherd groups’ write-ups, which will carry more weight than the announcement. And read the actual papers if your work depends on the underlying conjectures.
Frequently Asked Questions
What did OpenAI announce?
OpenAI announced three results in mathematics, computer science, and physics, plus a process for working with academics through expert “shepherds.” The Verge reported the announcement as three big science discoveries and a new way to work with academics.
Does OpenAI claim the Exponential Time Hypothesis is false?
Yes. OpenAI says its system proved ETH false and that the result was subsequently verified in Lean. The paper gives a deterministic algorithm that exactly counts satisfying assignments of fixed-width CNF formulas in subexponential time and polynomial space, and the 3-SAT case refutes the hypothesis. Independent academic confirmation is not yet clearly established.
What does the Kakeya result mean?
OpenAI says its local smoothing proof for waves in every spatial dimension at least two yields the Kakeya maximal and dimension conjectures, plus Stein’s spherical restriction estimates, throughout the interior of the conjectured range. Experts still need to confirm the scope.
Has OpenAI proved a QCD result?
OpenAI describes a mathematically rigorous existence proof for certain interacting quantum field theories with gauge fields and fermions in four dimensions, including QCD and a unified completion of the Standard Model. The company says the physics materials were not released in full, so outsiders cannot yet audit it.
Who are the shepherds?
Shepherds are external experts OpenAI contacted to organize groups that understand each result. For physics, David Gross will lead a group including Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. Participation does not equal endorsement.
Are the results peer reviewed?
Not yet, based on the cited reporting. Independent academic verification of all three results is unconfirmed. A Lean check backs the computer science claim, which is a formal verification step but not a substitute for expert review.
Which OpenAI model produced the proofs?
The announcement speaks of “our models” and “our system.” The sources we reviewed do not tie each proof to a single named model, so we do not assign one.
Why does this matter for software engineers?
Complexity assumptions like ETH underpin many arguments about what algorithms can and cannot do. If the refutation holds up, those arguments change. For now, nothing changes in production systems, and the sources claim no practical speedup.
