Ten thousand agents solved in 88 hours what humanity could not solve in 90 years. Twelve hours earlier, two flesh-and-blood mathematicians had announced nearly the same thing. And their drafts were sitting inside the rival’s machine.
The coincidence has a name: Navier-Stokes, existence and smoothness, one of the Millennium Problems.
The brutal announcement and the pending verification
OpenAI’s numbers are of another scale: an internal model more capable than Astra, about 10,000 agents, nearly 3 million messages exchanged, 130 billion output tokens, a Lean formalization covering two of the four required claims. The bill closes at 10 million dollars of compute to not claim the 1-million prize.
And the parenthesis the bill does not close: the solution has not been independently verified nor accepted by the Clay Institute. The technical feat is remarkable; the status is a claim, not an accepted proof.
The problem is not the math, it is the provenance
The dispute the case opened is the year’s most important for anyone using AI in intellectual work: Tristan Buckmaster and Levent Alpöge had been placing drafts of their work into Codex for a year. When Buckmaster asked whether the model had been trained on those sessions, he got no answer about training.
OpenAI denies access to specific user data but admits it cannot rule out that anonymized derivatives of usage helped improve the models. Read that sentence carefully: anonymizing removes the name, never the authorship of the idea. It is Andreas Thom’s synthesis, from Dresden, who asked about training and access for his sessions, received a one-line answer, and turned off training on June 29: a button that is a promise without external audit and only works forward.
The raw material of the next race
The conclusion is uncomfortable and simple: whoever places an original draft into a chatbot hands over the raw material of the next race and still watches from outside.
The case generalizes beyond mathematics. Every original draft (product, strategy, research, code) pasted into a third-party tool is the same move: the company’s most valuable raw material crossing the border under a retention policy written by the recipient. The key inventory maps who can turn things off; the circulation inventory maps what is being taught for free.
The inventory no lab does for you
At Tech86, the exercise that accompanies cases like this is the circulation inventory:
- Map the flows: where sensitive data touches AI (drafting, review, code, contracts).
- Classify the untouchable: the data that never reaches third-party models (original, strategic, client-owned).
- Verify the promise: turning off training is the vendor’s button; zero retention is worth what an audit proves.
- Establish the original-draft rule: it leaves the company only through tools with proven isolation, or it does not leave.
It is the same spirit as the incident-response checklist: preparation precedes the event. In the Navier-Stokes case, the event was the news; the lesson is for the draft you pasted yesterday.
Conclusion
88 hours, 10,000 agents, 10 million dollars: the technical proof that agent swarms solve Millennium Problems. And a provenance dispute proving something else: AI’s input is other people’s intellectual work, and its border is your data policy.
The question the case leaves every company using AI: if someone asked your vendor tomorrow "was this model trained on my drafts?", what would the answer be, and who would audit it? The faucet closes with inventory, not with trust.