Nobody saw the model.
That’s the part I keep circling back to. On September 8, 2026, OpenAI announced that an internal model had produced a proof resolving the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The method, as described, involved roughly 10,000 concurrent AI agents. Then in October, the company published 722 mathematical manuscripts on GitHub, crediting an unnamed internal frontier model it has not released. Alongside those came claims of hundreds more resolved problems, including a solution to the four-dimensional Kakeya conjecture and improvements on some of the world’s most important computer algorithms.
I build bots for a living. I spend my days wiring up agents, watching them argue with each other, debugging the moments where one of them confidently hands garbage to the next. So my reaction to all of this is not the reaction of a mathematician. It’s the reaction of someone who has looked at an orchestration graph at 2 a.m. and thought: how would I ever verify this?
Ten thousand agents is an architecture decision, not a magic number
The headline figure gets treated like a flex. To me it reads like a design choice with specific consequences. When you run agents at that concurrency, you are making a bet that search breadth beats depth of reasoning in any single thread. You’re also accepting a verification burden that scales with your fan-out.
Anyone who has built even a twelve-agent pipeline knows the failure modes:
- Agents that reach the right answer through wrong reasoning, which passes a check on output but poisons anything downstream
- Duplicate work that looks like independent confirmation but shares a common ancestor prompt
- Orchestrators that pick the most confident-sounding branch rather than the most correct one
- Logs so large that nobody actually reads them, which means the audit trail exists in theory only
None of that means the proofs are wrong. It means the architecture is the story, and the architecture is exactly what wasn’t published. We got 722 artifacts and no system to interrogate.
Why the pushback makes sense from a builder’s seat
The claims landed badly with parts of the mathematical community. Twenty-five Fields Medal winners argued that the push by AI companies to solve famous problems as a benchmark harms the science of mathematics and the mathematical community.
I take that seriously, and not out of deference. I take it seriously because it’s a critique of benchmark-driven development, and I’ve watched benchmark-driven development distort my own work. When you pick a target because it’s legible to outsiders, you optimize for the target. You stop asking whether the target measures the thing you care about. Famous open problems are extremely legible. They make for a clean announcement. They are also not a representative sample of what mathematics is for.
Swap the domain and the pattern is familiar to anyone in this space. A model tops a coding leaderboard and the write-up implies general software competence. Then you deploy it and discover the leaderboard was a narrow slice of a much messier job. Solving a celebrated problem tells you something real. It does not tell you the system is a mathematician.
What responsible sharing would look like
I don’t think the answer is to publish less. Dropping 722 manuscripts on GitHub instead of staging them through a press conference is, in some ways, the more open move. The problem is what came with them, and what didn’t.
If I were shipping agentic results into a field with its own standards of evidence, here’s what I’d consider table stakes:
- Name the model and pin the version, so results are attributable to something reproducible
- Publish the orchestration setup, including how agents were spawned, how their outputs were merged, and what the tie-breaking rules were
- Separate generated candidates from machine-checked proofs, and say plainly which is which
- Report failures at the same volume as successes, because the ratio is the only honest signal of reliability
- Invite review before publication from people who will push back, not after
That last one matters most. A field already in shock is not a field well positioned to evaluate a sudden flood of output. Volume without context shifts the verification cost onto the people least equipped to absorb it, and then their exhaustion reads as resistance to progress.
The practical takeaway for the rest of us
Most of us aren’t running 10,000 agents. But the lesson scales down cleanly. Breadth of search is cheap now and getting cheaper. Verification is not. If your pipeline generates more claims than your checking layer can handle, you haven’t built a solid system, you’ve built a very fast way to produce unverified assertions.
Build the checker first. Publish how it works. Let people break it.
🕒 Published: