Eliza Ward: Hey — rough week for anyone who spent the last decade working on the Connes conjecture, I'll tell you that much.
Brian Reed: Wait, that's — okay, I need context, what happened?
Eliza Ward: OpenAI, August 1st through 3rd of this year, publishes — not a product, a research post. 'Ten advances in mathematics and theoretical computer science.' The model behind it is Astra, which isn't even out yet. And one of the ten results is a disproof of Alain Connes's rigidity conjecture. Not a partial result. A disproof. Two different groups, same von Neumann algebra structure — the conjecture said that was impossible.
Brian Reed: And the proof is — what, we're taking OpenAI's word for it?
Eliza Ward: No, and this is the part that actually separates this from every prior AI math claim. The proofs are in Lean 4 — a formal proof assistant — published on GitHub, Apache 2.0. Sorry count zero. Zero. That means no step in any of the ten proofs is unverified. You can run it locally, it either checks or it doesn't.
Brian Reed: Ten proofs. The total cost OpenAI reported — I saw the $2,000 figure and I honestly assumed that was per problem.
Eliza Ward: All ten. $2,000 combined, at GPT-5.6 Sol API rates, discovery phase. And that includes — wait, let me make sure I'm saying this right — the non-sofic group construction, which resolves a question Mikhail Gromov left open in 1999. Twenty-seven years. First explicit proof that non-sofic groups exist at all.
Brian Reed: So the math world sees all of this on August 3rd and — nothing? Nobody's gone on record in 48 hours?
Eliza Ward: Silence isn't rejection — but it isn't acceptance either. And that gap is what makes this difficult.
Brian Reed: Hang on though — think of it like a spell-checker running on a legal contract. It can guarantee every sentence is grammatically correct. It cannot tell you whether the contract is fair, or even whether it solves the problem the lawyers actually cared about. Lean 4 is the spell-checker. It certified the logic. That's real. But significance? That's a different question entirely.
Eliza Ward: Right — and the label on the GitHub repo makes this concrete. It says 'agent-reviewed.' Not community peer-reviewed. Those are not the same thing.
Brian Reed: Which means — wait, so who actually vouched for whether the Connes disproof solves the conjecture as mathematicians meant it?
Eliza Ward: Nobody yet. Forty-eight hours out, formal peer review is still pending.
Brian Reed: So picture someone whose whole career is von Neumann algebras. August 3rd, they open GitHub, they pull the Connes proof locally — it verifies in seconds. It's correct. And then they're just... sitting there. Because there's no process for what comes next. The institution doesn't have a playbook for 'machine proved it, now what.'
Eliza Ward: They can't say it's wrong — Lean 4 verifies every step down to the axioms, that's what sorry-count zero actually means. But they also can't say it's understood. Those are two genuinely different things.
Brian Reed: Logical correctness isn't mathematical significance. That's the sentence. And I don't think the announcement separates those cleanly.
Eliza Ward: No — and Gromov introduced soficity in 1999. Twenty-seven years, nobody could prove non-sofic groups existed. Astra produces the construction. The proof verifies. Is that resolved? Formally, yes. Professionally? The field hasn't said.
Brian Reed: But that 'professionally unresolved' picture gets weird fast — because within 24 hours Levent Alpoge, an Anthropic researcher, says he reproduced five of the ten proofs. Using Fable. On a generic prompt. No internet access.
Eliza Ward: Wait — five of ten. That's half.
Brian Reed: Right, so — is that corroboration or is that deflation? I mean, classically, if a rival lab independently finds the same answers, that's a validation signal. The proof strategies hold up. But if Fable can do it on a generic prompt with no internet, then what exactly is Astra's advantage here?
Eliza Ward: Okay, here's where I actually think the hot take survives — partially. The Lean 4 sorry-count-zero is still doing structural work regardless of who solved it. Alpoge reproduced the results. He didn't re-formalize ten proofs into machine-checkable Lean 4 certificates overnight. That pipeline — the 249-page manuscript, the GitHub repository — that's still Astra's.
Brian Reed: So corroboration, then. On the math. Not on the capability claim.
Eliza Ward: On the math, yes. And actually — wait, that might matter more than the deflation story. If Fable independently lands the same proofs, the strategies are probably sound. The Erdős unit distance disproof, Ehrhart's volume conjecture — those aren't flukes if a second model hits them cold.
Brian Reed: The part that I can't square is that Astra hasn't even launched. OpenAI framed this whole announcement as a sneak peek — GPT-5.7, maybe GPT-6, somewhere down the road. So we're debating the significance of a model that isn't a product yet, while Anthropic's already-deployed Fable is apparently doing half of it on demand.
Eliza Ward: Which — yeah, that reframes the competition entirely. The breadth is still real though. Sphere packing, lattice cryptography, combinatorics, von Neumann algebras — these results land in completely different departments simultaneously. The peer-review burden isn't just slow, it's scattered across communities who don't talk to each other.
Brian Reed: And the $2,000 figure — that's the part we haven't fully broken open yet, and I think when we do, the story looks different than the headline.
Eliza Ward: The $2,000 is real — I want to be precise about that — but it's the discovery phase at Sol API rates, and the 249-page manuscript, the Lean 4 formalization, the human researcher hours to set up the GitHub repository: none of that cost is broken out. So when the headline reads '$2,000 for ten theorems,' what it actually means is '$2,000 for the bottleneck step.' Which might still be the important thing! But the full pipeline cost is invisible.
Brian Reed: The framing does a lot of quiet work there.
Eliza Ward: It does. And here's what that leaves us with — actually, let me phrase this carefully — what's defensible is: discovery is now cheap. Understanding what you discovered is not. Those are different bottlenecks. And the field has no process for the second one when the first one moves this fast.
Brian Reed: The Apache 2.0 license lowers the barrier to engage — anyone can pull the Lean proofs, run them locally. But the volume is the problem, right? Sphere packing, lattice cryptography, Erdős combinatorics, von Neumann algebras — those communities don't share reviewers.
Eliza Ward: Ten results, scattered across departments that genuinely don't talk. That's not slow peer review — that's no existing peer review structure at all.
Brian Reed: So the 48-hour silence — I mean, I want to be careful not to overread it — but it's not awe. It's closer to an institution realizing it doesn't have a form to fill out.
Eliza Ward: Wait — that's actually the calibrated version. Not 'AI solved math.' Not 'the proofs are suspicious.' The honest claim is: Lean 4 guarantees logical correctness down to the axioms, the $2,000 covers discovery only, and the gap between 'verified' and 'accepted' is now an institutional problem rather than a technical one. The field has to decide whether human understanding is a prerequisite for acceptance — and it's never had to decide that before.
Brian Reed: And the closest-vector result — the lattice cryptography one — I keep seeing people jump to 'this breaks post-quantum encryption,' which is just wrong.
Eliza Ward: No — it strengthens worst-case hardness evidence. That's the opposite direction. It's a result that makes lattice cryptography look more secure, not less. Nothing in the announcement constitutes an attack on deployed post-quantum systems. That's the headline that shouldn't exist.
Brian Reed: Fine. But the shape of this, when you step back — ten results, from a model that hasn't launched, reviewed by nobody in the community yet, for what amounts to less than a month of one researcher's salary. I mean, that's not nothing. That's just... also not finished.
Eliza Ward: Fine. $2,000 and a lot of uncosted human labor. That's actually — yeah, that's the honest version of the headline. And if the proofs survive peer review, which the Lean 4 verification makes likely on the logic, the field faces something it hasn't faced before: absorb machine-generated results, or decide that human understanding is a prerequisite for acceptance. That's not a technical question anymore.
Brian Reed: The silence after 48 hours — I keep landing there. It's not awe. It's a field that built every review process for one paper at a time, now staring at ten, from something that isn't even a product yet.
Eliza Ward: That's the one. Good think-through.
