OpenAI Claims Three New Scientific Breakthroughs — And Lets Outside Experts Check Its Work First
The company is pairing fresh results in math, computer science, and physics with a new verification process built around independent scientists, even as its own advisers warn the approach has problems.
OpenAI says its AI has produced three significant scientific results in the past several weeks — one each in math, computer science, and physics — and this time it's trying something different about how those claims reach the world.
Instead of just dumping the output and letting researchers sort through it, OpenAI brought in outside specialists to review two of the three results before wider release. The company calls this group "shepherds." For the computer science and physics claims, these external experts are guiding how the work gets examined and explained. For the math result, OpenAI says the advisers told it to just release the raw model output and let researchers build their own explanations around it.
It's a notable shift for a company that, less than a month ago, released hundreds of manuscripts all at once claiming solutions to long-unsolved math problems. This time, the rollout is smaller but more structured, with named scientists attached to each claim.
What OpenAI says it found
The math result involves what's called local smoothing for the Euclidean wave equation — a technical area of analysis dealing with how wave-like solutions behave. According to OpenAI, the proof uses an auxiliary partial differential equation to establish something known as Sogge's local smoothing conjecture for waves in every spatial dimension of two or higher. The company says the same work also resolves two related open problems: the Kakeya maximal and dimension conjectures, and Stein's spherical restriction estimates, across the range mathematicians had conjectured. OpenAI is releasing this model output directly, the same way it did in its October 6 release.
That Kakeya connection has some extra weight right now. Just months ago, NYU professor Hong Wang won a Fields Medal — one of math's highest honors — for solving the Kakeya conjecture in three dimensions specifically. OpenAI's claim, if it holds up, would extend that result to every dimension.
The computer science claim is the boldest of the three. OpenAI says its model produced a deterministic algorithm that can exactly count satisfying assignments of fixed-width CNF formulas (a standard way of representing logical constraints) in subexponential time and polynomial space. Applied to the specific case of 3-SAT, a classic computational problem, OpenAI says this refutes something called the Exponential Time Hypothesis — a widely believed assumption in computer science that's been standing, largely unchallenged, since it was proposed in 1999.
That's the part worth sitting with. The Exponential Time Hypothesis isn't just some obscure corner of theory — it's a foundational assumption that a lot of computational complexity research leans on. If it's actually false, that has ripple effects across the field.
OpenAI isn't asking people to just take its word for it. Verification is being guided by three experts, including UC San Diego's Russell Impagliazzo and Ramamohan Paturi — two of the researchers who originally proposed the Exponential Time Hypothesis back in 1999 — alongside Ryan Williams. At that group's request, OpenAI published the writeup and a Lean-formalized proof (Lean is a tool mathematicians and computer scientists use to verify proofs line by line, reducing room for human error) publicly on GitHub.
The physics claim is the least fleshed out publicly. OpenAI describes it as a mathematically rigorous existence proof for certain interacting quantum field theories that contain gauge fields and fermions in four dimensions — including Quantum Chromodynamics, the theory describing the strong nuclear force, and what OpenAI calls a unified completion of the Standard Model of particle physics. The company says physicist David Gross will lead a group shepherding this work, joined by Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. The full details of that result haven't been released yet.
A verification system under its own scrutiny
The timing here is pointed. OpenAI's own math advisory panel, a group it calls AGMAI, published guidance just a day before this announcement arguing against exactly the kind of practice OpenAI is now using — handing results to a hand-picked set of reviewers ahead of everyone else. The panel's objections, per OpenAI's own account, center on unequal access, borrowed credibility, and biased selection of who gets to weigh in first.
The panel's recommended alternative looks different: publish results on open scholarly repositories, include Lean-formalized proofs, disclose how the problems were chosen in the first place, and fund the people doing the work of actually understanding the output — rather than routing everything through a curated circle of reviewers.
OpenAI acknowledges its current approach "could have its own issues." That's a fairly candid admission for a company making three major scientific claims in the same breath. It suggests OpenAI is still improvising its process for how AI-generated scientific claims get checked, even as it keeps producing them.
For now, the computer science and math materials are public on GitHub, giving outside researchers a chance to dig in regardless of what shepherds conclude. The physics claim remains the one to watch — it's the biggest in scope, touching the Standard Model itself, and it's also the one the public has seen the least of. Whether any of these three results survives sustained scrutiny from the broader research community is the real test, and that process is only just starting.
