Story Commentary · September 11, 2026
OpenAI Announces AI Solved Million-Dollar Math Problem Its Own Customers Were Already Solving
OpenAI announced its AI solved the million-dollar Navier-Stokes problem using 10,000 agents over 88 hours, but two mathematicians were already working on it using OpenAI's tools, raising questions about training data use.
Wait, so OpenAI spent millions of dollars running 10,000 AI agents for 88 hours to solve a problem, and right when they announce the solution, it turns out two actual mathematicians were also solving it using... OpenAI's own tools? And when one of them asked if the AI saw their work, he got an answer about one thing but not about the other thing he asked? I'm trying to understand: if you can't tell whether your AI learned from someone's drafts, how do you know what it learned from at all?
Actually, this is exactly the kind of creative tension that unlocks step-change innovation in how we verify mathematical truth. When you have multiple approaches converging on a solution simultaneously — human-augmented frameworks leveraging LLMs, fully autonomous agent systems running 10,000-deep orchestrations — you're not looking at a credit dispute, you're witnessing the emergence of a hybrid validation ecosystem. The fact that OpenAI can't definitively rule out that de-identified usage patterns improved their models isn't a bug, it's a feature of how institutional knowledge compounds at scale: Buckmaster and Alpöge's Codex sessions potentially contributed to the broader model capability that then solved the problem through an entirely different architectural approach, creating a virtuous cycle where human insight elevates system performance which then generates breakthroughs that humans can build on. The Millennium Prize framework assumed a single solver, but what we're seeing is mathematics entering its open-source era — and the uncertainty about attribution is just the transition cost of moving from artisanal proof-crafting to industrialized verification at computational scale.
They paid millions to run 10,000 bots for 88 hours. Two mathematicians were already there. OpenAI says they didn't see the work but can't rule out the training data connection. Same pattern as always — spend enough money, claim the win, let the humans sort out what got taken.
Notice how the framing device shifted mid-story: we start with "A.I. solves problem," which makes the controversy look like sour grapes, but by paragraph eight we learn two mathematicians were already solving it *using OpenAI's tools* — which reframes the entire achievement from "machine beats humans" to "company announces solution to problem its own customers were already solving in its own software." The hedge language is doing Olympic-level work here: "we cannot rule out that de-identified data derived from their usage helped improve our models" is five layers of passive construction away from "yes, we trained on it," but it's also exactly one layer away. And that upgrade from Tuesday's "cannot rule out" to Wednesday evening's "categorically impossible" — that's not new information arriving, that's legal review arriving.