Eliza Ward: Hey. You okay? You've got a look.
Brian Reed: Yeah, I'm — let me see, how do I put this. I woke up to, like, four different people sending me the same Scientific American headline. Which doesn't usually happen.
Eliza Ward: The 'dueling proofs' one.
Brian Reed: The dueling proofs one. October 27th — OpenAI and Anthropic both announce they've refuted the Exponential Time Hypothesis. Same day. Same problem.
Eliza Ward: And that's not all OpenAI announced. They also — wait, this is the part that actually stopped me — they put out results in mathematics and physics. Including something that touches the Yang-Mills problem. That's one of the Clay Millennium Prize Problems. A million dollars.
Brian Reed: On the same day Anthropic filed its IPO.
Eliza Ward: On the same day. Right.
Brian Reed: The part I don't get — do we actually know if they proved the same thing? Like, 'dueling' implies comparability. But the sources don't establish whether those two ETH proofs overlap at all.
Eliza Ward: We don't. And that's actually the thing — the ETH question is almost a sideshow at this point, because with the Yang-Mills result, OpenAI put the mathematics on GitHub but withheld the physics proof entirely.
Brian Reed: Hold on. They released some of it but not the part that touches the Millennium Prize?
Eliza Ward: Correct. And here's the plain version of why that matters: in science, the proof is the claim. Without it, you haven't demonstrated a result — you've announced that you're claiming one. Those are completely different things.
Brian Reed: It's like a chef announcing a new dish exists but not letting anyone taste it. Or even see the recipe. You've announced an event, not a result.
Eliza Ward: And AGMAI — they said this out loud. October 26th, the day before all this dropped, they put out a statement designating expert shepherds specifically to manage AI-generated results. Then OpenAI announced anyway, without releasing the paper, and AGMAI came back and said there is — I want to get this right — 'no scientific reason for delaying' its release. That's not a vague criticism. That's a named organization saying the withholding is unjustified.
Brian Reed: And nothing happened. The paper's still not out.
Eliza Ward: Right. There's one verification thread that actually exists — David Gross, Nobel laureate, leading an external review group. Which, I mean — that's a serious name. But even that group doesn't have a public paper to work from, as far as the sources show.
Brian Reed: So the review is happening, but the thing being reviewed isn't accessible to anyone else. And Constantin Kogler at the Institute for Advanced Study — he actually welcomed the outside consultation but specifically called out the restricted access. So even people who like the idea of having David Gross involved are saying the closed-door part is the problem.
Eliza Ward: And to be precise — OpenAI's own framing is that the manuscripts need further academic review. Which, wait, that's — if they need further review, why make public announcements implying resolved breakthroughs? Those two positions don't sit together. OpenAI also characterized the Yang-Mills result as only partially touching the Clay problem, not fully resolving it. So the announcement was already overstating what they're even claiming internally.
Brian Reed: And that's the part where I think the circulating take just doesn't hold. The one I keep seeing — 'of course two frontier labs converged on the same famous problems, they have identical training data and identical capabilities.' That's the clean version. But hang on — the ETH announcements were four minutes apart.
Eliza Ward: Four minutes.
Brian Reed: Four. And Kevin Roose's Atlantic piece — September 2026, so this is documented before October 27th — describes a pattern of OpenAI and Anthropic releasing products within minutes or hours of each other. That's not new. That's a habit.
Eliza Ward: Right, but — okay, I want to be careful here. I'm not saying it's coordination. I genuinely don't know that. What I'm saying is the convergence story requires you to ignore that pattern and also ignore that Anthropic filed its IPO the same day. An IPO filing is a document where you disclose risks. Announcing a scientific breakthrough the morning of that filing is — I mean, that's not a timing coincidence you can just set aside.
Brian Reed: So the question is whether Sam Altman and Dario Amodei are both just on the same scientific clock, or whether the announcement is doing work that the science isn't finished doing yet.
Eliza Ward: And the announcement strategy actually answers that. If you were confident the proof survives review — actually confident — you release the paper. Full stop. You don't announce a Yang-Mills result to the press and then tell David Gross's review group something that amounts to 'the manuscripts need further work.' Those are contradictory positions held simultaneously.
Brian Reed: No, I don't buy that as fully resolved either — I mean, labs sit on papers for PR timing all the time, that's not automatically — actually, wait, yeah it is different here. Because the Clay Prize has a million dollars attached to the Yang-Mills problem. The stakes of being wrong aren't just reputational.
Eliza Ward: And that's before we even get to what this does to the researchers who were closest — that's a whole separate problem, and honestly the part that I think makes this worse for everyone downstream.
Brian Reed: The researchers. Yeah. So Sussman and Alpöge — they had updated their GitHub the night before OpenAI's announcement. The night before. They were close enough that they were committing code to a public repo on the same problem.
Eliza Ward: And then woke up to find it gone.
Brian Reed: Gone. And — I mean, Michael R. Douglas at Harvard called OpenAI's QCD result 'a major advance.' That quote is out there. The human team working the same problem? Nothing. No equivalent recognition. Douglas is talking about the AI result as if it's the event, and the actual researchers who were days away are just... not in that sentence.
Eliza Ward: And QCD — quantum chromodynamics, the strong nuclear force — that's not abstract. If OpenAI's claim on that holds, it's verifiable physics. But verification requires the paper, which doesn't exist publicly yet.
Brian Reed: Right — and think about what that actually means structurally. Picture a grad student, three years into her own Yang-Mills project, sitting down at her laptop Friday morning and pulling up the news. Her work isn't wrong. It's just invisible now. There's no norm that says 'here's what your contribution is worth when an AI crossed the line first.' The Clay Millennium Prize framework has no established process for adjudicating AI-generated proofs. Nobody built that rule yet.
Eliza Ward: And the ETH refutation compounds that — because ETH touches P versus NP, which touches cryptography. Attribution on a foundational result in that space isn't just academic credit. It's — I mean, careers, funding lines, entire research programs.
Brian Reed: So what do we actually watch for? Like, concretely.
Eliza Ward: Whether David Gross's review group releases publicly — that's the first signal. If it stays internal, the verification bottleneck just becomes permanent. And whether AGMAI's shepherd process develops any actual enforcement mechanism. Right now they can issue statements. That's it.
Brian Reed: And the Clay Mathematics Institute hasn't said anything yet about how they'd even evaluate a machine-generated proof for a million-dollar prize. That question is not hypothetical anymore. It's sitting on someone's desk.
Eliza Ward: And that's — I mean, that's where I land. Both claims from OpenAI and Anthropic are still provisional. Not peer-reviewed, not independently verified as of October 27th. That's not my read on them, that's just the state of the record. So if you ask me what I actually think right now: I can't tell you the science is wrong. I also can't tell you it's right. What I can tell you is the announcement strategy created an incentive structure that is — wait, structurally opposed is the right phrase — opposed to the patience verification actually requires. Those two things are in direct conflict and nobody's resolved that.
Brian Reed: Right — there are two possible stories here. Either the proofs survive David Gross's review and get released, and then this was genuinely new, genuinely fast, and the field has to absorb what that means. Or the paper stays locked, the review stays internal, and the announcement was — I mean, being first was itself the product. The speed was the thing being sold. And we don't know yet which story this is. That's not me hedging. That's actually what the evidence leaves open.
Eliza Ward: The Gross review. That's the signal. Public release or not.
Brian Reed: And if it stays behind closed doors, we'll know something — just not what they proved.
