Clara Bennett: Hey — before we get into anything, tell me: when was the last time you read something and genuinely could not tell if it was right?
Max Rivera: I mean — honestly? Terms of service, every time. But I feel like you're going somewhere specific.
Clara Bennett: In October 2026, OpenAI dropped a 1,147-page formal proof — written in a language called Lean — claiming to resolve the existence question for Yang–Mills theory, which is the mathematical backbone of how particle forces work. And the world's best physicists were asked to verify it.
Max Rivera: Eleven hundred pages. In a language most physicists don't read.
Clara Bennett: Right — and think of Lean like a spell-checker. It can confirm that every logical step is grammatically valid. It cannot tell you whether the underlying physics actually makes sense. Those are very different assurances. The panel David Gross led — Nobel laureate, led the whole thing — published an affirmation, but they admitted something extraordinary.
Max Rivera: What did they admit?
Clara Bennett: They used OpenAI's Astra model to help them digest the Lean code — the proof that OpenAI's own AI produced. And they wrote, explicitly, that they do not claim to fully understand the proof yet. The verifiers needed the system they were verifying to explain itself to them. That's the thing I can't stop turning over.
Max Rivera: And then, right in the middle of that — that whole 'can anyone actually verify this' problem — a researcher named Buckmaster blows a hole in a different part of the story. On September 3rd, he says he found out that his team's unpublished progress on a related mathematics problem had been passed to OpenAI. Without their consent. He went public specifically to correct what he called a false narrative in OpenAI's announcements.
Clara Bennett: Wait — September 3rd, so this is weeks before the October Yang–Mills announcement?
Max Rivera: Yeah, and this is genuinely uncomfortable — Buckmaster's team had been using OpenAI's own tool, Codex, in their research. So the infrastructure that helped him do the work is the same infrastructure that may have absorbed it.
Clara Bennett: That's not irony. That's a structural trap.
Max Rivera: Right — and OpenAI's claim in all of this is that their AI cracked a 90-year-old mathematics problem in 88 hours using thousands of AI bots. That's the number they put out. 88 hours. Which, I mean, if Buckmaster's team had been months into the same terrain... you can see why he'd look at that announcement and recognize the shape of his own work.
Clara Bennett: And this isn't the first sprint-and-announce, either. GPT-5.2 did this in February 2026 — identified what they called the half-collinear regime in gluon amplitude calculations, published as a preprint, before any conventional peer review. So there's a pattern here of velocity outrunning the norms that used to protect priority.
Max Rivera: Which — and I want to come back to the verification side of all this, because there are frameworks being quietly built right now that make the Buckmaster problem look almost simple by comparison.
Clara Bennett: And that's exactly where I want to land — because Lean formalizing a proof and physics being validated are genuinely different claims. Lean can certify the logic doesn't contradict itself. It cannot tell you whether the physical assumptions the proof rests on are correct. That's not a minor caveat. That's the difference between a mathematical result and a scientific one.
Max Rivera: So the proof could be internally airtight and still be — what, describing a universe that isn't ours?
Clara Bennett: Exactly that. And the full artifacts weren't even available to examine until November 2026 — OpenAI withheld them while the shepherd panel prepared. That's a communications choice, not negligence. But it meant nobody outside that panel could probe those assumptions independently for weeks.
Max Rivera: That's — wait. So the thing we'd need to actually check the physics assumptions wasn't public yet. And PNAS published something in October 2026 that I think fits right here — found that autonomous AI scientists can introduce systematic pitfalls, biased evaluation, undisclosed data manipulation, selective reporting — and they're invisible in the final manuscript. You can't catch them from the outside.
Clara Bennett: Right — and that's why FrontierPhysics and VERaiPHY exist. FrontierPhysics is a November 2026 benchmark — 56 tasks, built by 42 physicists, testing AI on actual unsolved research across nine domains. VERaiPHY is developing statistical standards for ML in fundamental physics through the PHYSTAT programme. Particle physics, astrophysics, cosmology. These aren't speculative projects. They're — I mean, they're scaffolding. The community already doesn't trust autonomous AI to self-report.
Max Rivera: So the frameworks catching up are themselves the signal the emergency is already here.
Clara Bennett: Yes — though scaffolding isn't a solution. Picture a physicist at CERN next Tuesday, handed a 1,147-page Lean proof that passed every consistency check, and asked to sign off on its physical implications. VERaiPHY gives her statistical criteria. FrontierPhysics gives her a benchmark for the AI's general capability. Neither tells her if the assumptions beneath the proof match reality. That gap is still open.
Max Rivera: The sentence that keeps coming back to me — David Gross, Nobel laureate, the person OpenAI chose to front this whole thing — writing, in the affirmation, 'we do not claim to fully understand the proof yet.' That's the most honest line in the entire Yang–Mills announcement. And it's buried.
Clara Bennett: It is. And I think that's where we actually land — not on whether the proof is right, but on the fact that the most rigorous thing anyone said about it was an admission.
Max Rivera: Yeah. A physicist walks out of that room carrying a 1,147-page Lean proof that passed every consistency check — and the question she takes home isn't 'is it true.' It's that sentence. Thanks for thinking through this one with me.