Elsewhere · This week
AI is everywhere. Then it went after a million-dollar math problem.
From phone cameras to fluid equations — the same week the tools got ordinary, a prize-class dispute got loud.
NEHALEM, Ore. — Artificial intelligence stopped feeling like a lab toy between the spell-check that rewrites a grant sentence and the photo app that clears fog off a coast shot. In 2026 the tools draft emails, sketch websites, summarize board packets, and sit in the same browsers people use to pay a water bill. That everyday flood is the easy part. The hard part showed up this week in mathematics.
On Sept. 8, OpenAI published a claim that an internal AI system had resolved the Navier–Stokes existence and smoothness question — one of the Clay Mathematics Institute’s Millennium Prize Problems, each carrying a $1 million purse. The company posted a write-up and a Lean formalization, framed as evidence of how fast models are moving, not a cash claim.
“We do not intend to claim the Millennium Prize for this result,” OpenAI wrote on its site.
Timing made the week combustible. OpenAI said that on Sept. 1 it heard rumors Millennium Prize problems had been resolved — rumors it later traced to Tristan Buckmaster, a mathematics professor at New York University, and Levent Alpöge, a researcher employed by Anthropic who has said the math project was not Anthropic’s assignment. The pair had been pushing related fluid-equation questions — including forced Euler blowup — and, according to reporting, had been using OpenAI’s Codex while they worked.
OpenAI says it then threw a large multi-agent effort at the open Millennium problems with a new internal model still in training. Agents reached a Navier–Stokes resolution in roughly 88 hours, the company says, plus Lean verification. It says its researchers did not see Buckmaster and Alpöge’s unpublished work, while conceding it “cannot rule out” that de-identified data from product use helped improve models.
Buckmaster has publicly argued OpenAI fought dirty — that a parallel corporate sprint after hearing rumors, plus the chance models trained on researchers’ own Codex sessions, turns academic priority into a scoop war. TechCrunch and other outlets covered those allegations this week alongside OpenAI’s denials that it accessed specific user data.
Two truths can sit on the same desk. First: the public is already inside an AI boom that is not waiting for prize committees. Nonprofits outline budgets with chat tools; local reporters clean notes; designers stub a page. The “state of AI” is less a breakthrough than a thousand quiet substitutions — afternoon work now done in twenty minutes, for better and for worse. Second: when the industry that sells those tools also trains on the traces of human research labor, credit stops being etiquette and starts being infrastructure. A Millennium Problem is supposed to be settled by mathematics the field can check. Lean helps. Provenance — whose intermediate ideas fed which model — is harder to audit when the training loop is private.
OpenAI’s post emphasizes that its Euler-related claim differs from Buckmaster and Alpöge’s forced-Euler result, and that it offered a concurrent announcement and priority recognition after the fact. That is academic manners at industrial speed. Another version is simpler: if researchers build in the open, or nearly open, inside a vendor’s tools, the vendor should not treat rumor-triggered compute as a clean-room race.
Coast Desk is not here to crown a winner of Navier–Stokes. Clay’s rules and the mathematics community will do the slow work. For a North Coast reader, the pattern is the news: AI is already in the phone and the grant file, and the same stack is colliding with century-old open problems and the humans who have been grinding on them for years. The tools will keep finishing sentences. The question this week is whether the people who did the hard thinking still get named when they do.