The most surreal AI story of the week did not arrive as a research paper — it arrived as a keynote in Abu Dhabi. At the ADIPEC 2026 opening ceremony, OpenAI CEO Sam Altman said the company's next-generation GPT, the internal model that recently proved the Navier–Stokes equations, will ship "soon," and described how it got there: a swarm of 10,000 AI agents, with no human in command, worked for 88 straight hours to produce a mathematical proof, then spent another 17 hours turning it into a machine-checked Lean formalization. If even half of that is as described, the boundary between "AI assists mathematicians" and "AI is the mathematician" just shifted by a lot.

A Swarm With No Boss

The detail that makes this genuinely strange is the orchestration — or the lack of it. According to Altman's recount, there was no top-level "foreman" agent breaking the problem into tasks while subordinates executed mechanically. Instead, 10,000 agents picked their own directions, shared progress in real time, borrowed each other's ideas, and decided their own next steps. This is the opposite of how most multi-agent systems are built today, where a controller plans and workers comply. The result, he said, was a four-day, four-night autonomous push that a human mathematician would have struggled even to scope.

What Reportedly Got Solved

The headline claim is that this was the first Millennium Prize problem to fall: the Navier–Stokes existence and smoothness problem, the grand challenge of fluid dynamics. Beyond that, Altman said the same capability touched Riemann, Hodge, and the Kakeya conjecture across all dimensions, overturned the exponential time hypothesis, and produced a proof of Yang–Mills existence. The output was not a single note but a pile of 722 research manuscripts. Even read as an ambitious marketing narrative, the shape of the claim — many agents, many subfields, one continuous run — is the interesting part.

"Smarter Than Most Top Scientists"

Altman reportedly told the audience that "AI is now, most of the time, smarter than top scientists," and described the model behind this as "only a few months old." That is the kind of line that sounds like hype until you remember that the same stage has been used to preview real releases before. The fun here is the gap between the dry language of prize committees and the sci-fi reality of agents radioing each other in the dark.

Key Data Points

  • 10,000 autonomous agents collaborated with no central controller
  • 88 hours of continuous work to produce the proof
  • 17 hours for Lean formal verification of the result
  • 722 research manuscripts generated during the run
  • 1st Millennium Prize problem reportedly cracked (Navier–Stokes)
  • Subfields touched included Riemann, Hodge, the Kakeya conjecture, the exponential time hypothesis, and Yang–Mills existence

Why It Matters

Even allowing for showmanship, the signal is real: frontier labs are now talking about agent swarms that run for days without human direction and produce original mathematics. Whether or not the Navier–Stokes proof holds up to scrutiny, the demo of autonomous, self-organizing research changes how we should think about the pace of discovery. The disappointing flip side is the same one we always hit — we are being told the result exists, but the proof trail for outside experts is the part that still has to survive contact with reality.

Originally published on threerouter. Sources: Xinzhiyuan via NetEase, Phoenix Tech AI Morning Post.