Brian Reed: Hey, good morning. I want to start with something that came across my feed that I genuinely don't know what to make of.
Eliza Ward: The Jacobian conjecture thing.
Brian Reed: Yeah. An 87-year-old open math problem — posed by Ott-Heinrich Keller in 1939 — and someone just posted on X that they broke it.
Eliza Ward: Levent Alpöge. He's a mathematician at Anthropic. The post went up August 6th. He says Claude Fable 5 helped him find a counterexample — compact, polynomial, three dimensions — and that it's verifiable by hand.
Brian Reed: Okay, and — wait, what is the conjecture actually asking? Like, in plain terms.
Eliza Ward: Think of it as asking whether a certain kind of mathematical transformation always has an undo button. Very specific conditions — nonzero constant Jacobian determinant — and the question is: does that guarantee reversibility? Alpöge says no. He found one that doesn't.
Brian Reed: And New Scientist is calling this the most significant math problem resolved with AI help. That landed fast.
Eliza Ward: That landed fast — and that's actually the problem. Gemini Grounded checked ScienceDaily and other reputable publications as of August 6th, the same day. No independent confirmation. Anywhere.
Brian Reed: Wait — none? Like, not even a preprint on arXiv?
Eliza Ward: X. That's it. No journal, no preprint server, no department seminar. Just a post. So New Scientist is calling it the most significant math problem resolved with AI assistance — before a single external reviewer has touched it.
Brian Reed: And the 'verifiable by hand' framing is doing a lot of heavy lifting there. Because — wait, let me think about what that actually means. It means you don't need a computer to check the steps. It does not mean anyone has sat down and checked the steps. Those are completely different sentences.
Eliza Ward: Right. Verifiable by hand is a property of the result. Verified by hand is an event that either happened or didn't. We have the first. There's no evidence yet of the second.
Brian Reed: And the reason that gap matters more here specifically — Beniamino Segre claimed a proof of the Jacobian conjecture. Found to contain errors. Wolfgang Gröbner, same conjecture, claimed a proof. Also errors. Two credentialed, serious mathematicians. This problem has a documented history of looking solved before it was.
Eliza Ward: Hold on — Segre and Gröbner, those were proofs, not counterexamples. Alpöge is doing something structurally different. A counterexample is — actually, no, that's not the full defense I thought it was. A subtle error can live in a counterexample just as easily.
Brian Reed: Exactly. The form is different but the verification problem is identical. And Keller posed this in 1939 — eighty-seven years of people being wrong about it. The history isn't trivia, it's the reason absence of confirmation actually means something.
Eliza Ward: So what's confirmed: Alpöge posted a claimed counterexample on X on August 6th using Claude Fable 5. What's not confirmed: that the counterexample holds. The next concrete signal is whether a peer review or independent verification actually surfaces — and right now that clock hasn't started.
Brian Reed: But that clock thing — hang on — there's a layer under it that we haven't named yet. Alpöge works at Anthropic. The same Anthropic that built Claude Fable 5. So the person making the mathematical announcement is employed by the company whose product is being celebrated.
Eliza Ward: Right. That's the structural fact. And weeks after Claude Fable 5's public release.
Brian Reed: So the result functions as both a mathematical claim and — I mean, whether or not Alpöge intended this — a product demonstration. Those two things are happening at the same time in the same post.
Eliza Ward: And the take circulating right now is: historic AI milestone, full stop. That doesn't hold up. Not because the math is wrong — we don't know yet — but because it flattens something important. The announcement itself is doing two jobs.
Brian Reed: Okay but — is it possible both things are true? Math is sound, incentives are compromised?
Eliza Ward: Yes. Actually — that's almost exactly the problem. It's not about bad faith. It's structural. Anthropic has direct reputational and commercial stakes in Claude Fable 5 being seen as capable of frontier mathematical discovery. That stake existed before Alpöge sat down with the problem.
Brian Reed: And the sources describe Claude's role how, exactly? Because I've seen 'collaborator,' I've seen 'tool,' I've seen 'helped uncover.' Those are not the same thing.
Eliza Ward: Unspecified. That's the honest answer. Did Claude generate the insight and Alpöge verified it, or did Alpöge direct the model toward an algebraic structure he already suspected? Those two scenarios have completely different implications for what we're claiming the model can do — and we can't tell which one happened. Think about a graduate student in algebraic geometry who reads this on X in August 2026, downloads Claude Fable 5 to try it on her own stuck problem — she has no way to close that gap. None.
Brian Reed: And honestly that gap gets harder to close because — the pace problem is the part that comes next, and I think it makes this worse. But the Jacobian announcement isn't even an isolated event.
Eliza Ward: Right — and that's exactly what Quanta Magazine was reporting three days before this. August 3rd. AI solving legendary Erdős problems. So by the time Alpöge posts on August 6th, there's already a pattern accumulating, not a single event.
Eliza Ward: Three days. And that's the structural problem. It's not that one result is unverified — it's that they're arriving faster than the field has any mechanism to process them.
Brian Reed: Because mathematics has — I mean, what does mathematics actually have? There's no body whose job is to audit this at speed. Peer review at a journal like Annals of Mathematics can take a year. Sometimes longer. And these claims are living on X for months before that process even starts.
Eliza Ward: There's no FDA equivalent. No — wait, actually that framing undersells it. The FDA at least has a defined mandate. Mathematics has journals, referees, seminars — all designed around a pace where one hard result appears every few years and everyone has time to look at it.
Brian Reed: So picture an algebraic geometer — say, sitting in a department in Bonn, September 2026 — and she's looking at the Jacobian claim, the Erdős results, and she wants to know which of these are real. Who does she call? What's the institution? Because if the answer is 'check X and hope someone credible replied to the thread,' that's — that's genuinely new territory.
Eliza Ward: And the coverage framing makes it worse. New Scientist calls Claude Fable 5 a research collaborator contributing novel conceptual moves — not a calculator. If that framing takes hold before the result is verified, it reshapes what people think these models can do based on a claim that hasn't cleared the first bar.
Brian Reed: So what are the concrete things to actually watch here? Not in the abstract — like, what are the specific next events?
Eliza Ward: A preprint on arXiv. Then a journal submission. Then independent algebraic geometers — not at Anthropic — confirming the counterexample holds. Those are the three checkpoints the evidence points to. If none of those happen by early 2027, the Quanta framing and the New Scientist framing are running ahead of a result that was never actually closed.
Brian Reed: And the Jacobian conjecture itself — Keller posed it in 1939, it's still formally unresolved in the literature. As of August 6th. Whatever Alpöge posted, that's the technical status. The literature hasn't moved.
Eliza Ward: That's the honest place to land. The conjecture is unresolved in the literature. The counterexample is a claim on X. And — I mean, those two things can coexist for months. Maybe longer. If AI results keep arriving at this pace and the people announcing them work for the companies building the tools, I don't actually know what mathematical verification means as a practice going forward. That's not rhetorical. I genuinely don't know.
Brian Reed: No, I don't think there's an answer yet.
Eliza Ward: Watch for the arXiv preprint. That's the first concrete move — not the X thread, not New Scientist. If Alpöge submits and independent algebraic geometers outside Anthropic confirm it, that's the signal. Until then the question is live.
