Brian Reed: Hey. Good to be back.
Eliza Ward: Yeah, you too. Though — okay, I have to ask. Did you see the OpenAI thing yesterday?
Brian Reed: The blog post. Yeah. Three scientific discoveries, they're calling them — math, computer science, physics. October 27.
Eliza Ward: Right. And the math output just — it's out. Public. But the physics findings? Withheld. Still. That's the first thing that caught me.
Brian Reed: Wait — so they announced all three together but only released one of them?
Eliza Ward: Essentially, yeah. And then — actually, hold on — the BBC had this separate detail. OpenAI also claimed their AI agents solved something related to the Navier–Stokes equations. In 88 hours. The prize body said it had apparently been settled but the review would take time.
Brian Reed: Eighty-eight hours for Navier–Stokes. That's — I mean, that's one of the most famous unsolved problems in math. And they're just... saying it like that.
Eliza Ward: Which brings me back to: why isn't the physics out yet?
Brian Reed: Because physics is where the claims get slippery. I mean — math, a proof either checks out or it doesn't. But physics existence proofs, quantum field theories, quantum chromodynamics — those require interpretation. Someone has to say what the result even means.
Eliza Ward: Right — but what actually stops me is this. OpenAI itself said outside academics must verify, interpret, and formalize these results. Their words. Which means what they announced October 27 isn't a discovery. It's a candidate.
Brian Reed: Say that differently — what do you mean candidate?
Eliza Ward: Like — okay, think of it this way. An AI generating a proof is a student handing in an exam answer. It might be right. But someone still has to mark it. The grade doesn't exist until the marking happens. That's what independent verification actually means here.
Brian Reed: And the Exponential Time Hypothesis — disproving that would be a decades-level result in computational complexity. That's not a student answer you just... skim.
Eliza Ward: No. And open science norms exist exactly for this reason — extraordinary claims go out broadly, preprints, peer review, the whole community marks the exam. Not five hand-selected shepherds. David Gross, Michael Douglas, Arthur Jaffe, Igor Klebanov, Nikita Nekrasov — I'm not questioning their credentials, but they were chosen by OpenAI.
Brian Reed: There's also — wait, the Codex thing. OpenAI said it couldn't rule out that anonymised product-use data helped its models, then later said specific Codex prompts could not have influenced the system in any way. That's the same company asking us to trust the independence of its results.
Eliza Ward: Which is why who does the marking matters enormously. If the marking is controlled — paced, filtered, shepherded — the grade can come out whenever OpenAI decides it should.
Brian Reed: So the question isn't whether the AI got the answer right. It's whether anyone outside OpenAI gets to check.
Eliza Ward: But that's actually where the circulating take breaks down — because a lot of people are saying, no, this is responsible disclosure, OpenAI brought in field experts, they're respecting disciplinary norms. And I get why that reads as reasonable.
Brian Reed: The charitable version. Yeah. David Gross won the Nobel, the shepherd group is real, they're not inventing credentials.
Eliza Ward: Right — except The Verge reported in September that the independent mathematical advisory group would 'not be responsible for advising on how to pace' releases. That's the sentence. OpenAI keeps the timing lever entirely.
Brian Reed: So they can't decide when.
Eliza Ward: They can't decide when. Which means — wait, think about what that actually makes them. They're not gatekeepers. They're explainers with a corporate deadline.
Brian Reed: It's like getting a mass email from a biotech company saying a panel of Nobel laureates will explain our new drug trial results to you when they're ready. You'd know immediately the panel isn't running anything.
Eliza Ward: Exactly — and the choice of which shepherds matters too. Gross, Michael Douglas, Arthur Jaffe, Igor Klebanov, Nikita Nekrasov — OpenAI selected all five. So the right to choose who frames a result is itself a form of scientific authority. That's not independent review, that's... I mean, that's curation.
Brian Reed: And the math going out immediately — no shepherds, just public — that's the tell, actually. The least interpretively contested finding gets open science treatment. The most dramatic one, quantum chromodynamics existence proofs, stays locked. That's strategic, not principled.
Eliza Ward: And the August convening is where that pattern started — forty mathematicians, a reported assurance, Lindsay McCallum later saying OpenAI is 'not aware of' that promise. That thread goes somewhere darker, and we need to get to it.
Brian Reed: And that August thread — I mean, most people are going to think this whole dispute started October 27. But the mathematician convening was August 2026. That's three months of this.
Eliza Ward: Three months where the pattern repeated and — wait, actually — the September advisory group announcement makes it worse, not better. The Verge called it 'messy and confusing.' Members said OpenAI had 'learned little from its mistakes.' That's not journalists editorializing, that's people inside the group.
Brian Reed: Learned little. Between August and October.
Eliza Ward: Right — and the spokesperson denial is the specific thing that stands out to me. Lindsay McCallum says OpenAI is 'not aware of' the assurance given to those forty mathematicians. That's not a correction. That's a non-denial denial. 'Not aware of' covers a lot of ground.
Brian Reed: It also puts forty researchers in an impossible position — like, what do they do with that? They remember the conversation, the company doesn't. There's no mechanism to resolve it.
Eliza Ward: Which is why the unresolved status of everything right now is actually the concrete thing to watch. Physics — not fully released. The computer science interpretation group — not yet assembled, as of October 27. The math output is public but not peer-reviewed. So we have three announced results, zero of them verified, and one of the groups that's supposed to do the verifying doesn't exist yet.
Brian Reed: The computer science group isn't assembled — so who's sitting with the Exponential Time Hypothesis result right now?
Eliza Ward: OpenAI. That's — I mean, that's the answer. Nobody else has it in a form they can evaluate. And disproving ETH would rewrite complexity theory. That's not a result you hold in a drawer.
Brian Reed: So the signal to watch is pretty simple — does the physics result ever fully release, and does that computer science group actually get named? Because if neither happens in the next few months, the August-to-October pattern isn't a learning curve. It's the model.
Eliza Ward: And if it is the model — that's the part I can't resolve. Because if the shepherded physics results come out six months from now and they're real, verified, important — every other lab watched how this landed. And the lesson isn't 'OpenAI got lucky.' The lesson is: pick the right five people, control the timing, and you've shaped how a discovery enters the record. That's not a one-time choice. That's a template.
Brian Reed: And then it's not about OpenAI's judgment anymore. I mean — that's the shift. It stops being 'did they do this responsibly' and becomes 'who else gets to do this now?' Because there's no shared definition of responsible disclosure sitting out there. Open science norms say the whole community marks the exam. The shepherd model says — actually, no — we'll choose the markers.
Eliza Ward: Two genuinely incompatible versions of scientific authority.
Brian Reed: And what would settle it — I mean, what would actually settle it — isn't whether OpenAI's intentions were good. It's whether the three October 27 claims get independently verified. Not explained by David Gross, not contextualized by the shepherd group. Actually verified. The Exponential Time Hypothesis, the quantum chromodynamics existence proofs, the harmonic analysis conjectures — none of those have cleared that bar yet.
Eliza Ward: And if they do clear it — through a process OpenAI still controlled — I'm not sure we'll even be able to tell the difference between 'this was right' and 'this was managed well enough to look right.' That's where I'm stuck.
