Peter Diamandis read Sam Altman’s reaction to OpenAI’s proposed mathematical breakthrough on the Moonshots panel. The company’s chief executive was not simply celebrating. He was arguing for caution.
“I did not expect a result of this magnitude to happen so soon,” Altman wrote, according to the passage Diamandis read aloud. After describing the need to pace progress for safety, Altman called the result “the strongest evidence yet for that urgency.”
The result was OpenAI’s 8 September 2026 announcement of a proposed proof concerning the Navier–Stokes equations, the subject of one of the Clay Millennium Prize Problems. The company reported approximately 10,000 coordinating AI agents, an 88-hour research run and another 17 hours for formalization and verification.
Diamandis put the question to the panel: genuine alarm, or marketing? The disagreement turned on what, exactly, should have been surprising—the ability to tackle a grand mathematical challenge, or the resources now apparently needed to do it.
What OpenAI says its system proved
The Navier–Stokes equations describe fluid motion through velocity, pressure and viscosity—the fluid’s internal friction. The prize problem concerns three-dimensional, incompressible flow: fluid treated as not changing volume under pressure. It asks whether suitably smooth conditions always permit a smooth solution that continues indefinitely, or whether a mathematical breakdown can occur.
OpenAI reports a construction in which the fluid starts at rest and a smooth external force pushes it. Its total energy remains finite, but its velocity becomes unbounded in finite time. That is the proposed singularity: a breakdown of the smooth mathematical description, not a claim that an ordinary cup of coffee will physically reach infinite speed.
The external force is part of the claim, not a detail to leave out. Charles Fefferman’s official formulation offers four acceptable targets. Alternatives A and B ask for global smooth solutions for arbitrary qualifying starting conditions without external forcing, either in ordinary three-dimensional space or in a periodic domain—a setting that repeats in space. Alternatives C and D ask for breakdown under suitable smooth starting conditions and smooth forcing.
OpenAI identifies its result with C and D. Forced flow is therefore not automatically outside the prize formulation. Whether the proposed proof satisfies all the required conditions is a separate question from whether forcing is allowed.
The company says it produced both an analytical proof and a formalization in Lean, software used to express mathematical statements precisely and check their logical derivations. This goes beyond a simulation that appears to show a blowup. Fefferman’s formulation explains why numerical evidence is difficult to trust here: extreme instability near a suspected singularity can undermine reliable inference from a simulation.
Formal verification nevertheless checks the statement and assumptions actually encoded. Acceptance as a solution to the Millennium Prize problem also requires scrutiny of how that formal statement corresponds to the official mathematical target. OpenAI says it will not claim the prize; its announcement is not an independently accepted prize solution. Clay allocated $1 million to each of its seven problems when it announced them in 2000.
The swarm, and the humans in it
The reported system was not one chatbot producing a proof in a single answer. AI agents—software systems that can pursue tasks through multiple steps—explored different formulations and shared intermediate insights. Humans directed the consolidation of work between groups.
That arrangement allowed separate lines of investigation to feed into one another, rather than leaving thousands of attempts isolated. The human role also means the reported result came from a coordinated research process, not an unattended machine handed a conjecture.
OpenAI separates the 88 hours spent reaching the result from the further 17 hours spent formalizing and verifying it. Discovery and checking were distinct parts of the run. The panel’s discussion of “test-time compute” concerns this broader possibility: spending more computation while a system works on a problem, rather than only while training the underlying model.
“The whole thing is in the pitch deck”
Alex Wissner-Gross said he took Altman at his word and thought he was probably surprised. He described a gap between rapid changes in mathematical research and the ordinary world outside the window, calling it a “normalcy overhang.” On the street, things could still look like business as usual while the panel believed scientific capabilities were changing sharply.
The discussion also included rumors of solutions to the Hodge and Birch and Swinnerton-Dyer conjectures, two other Millennium Prize problems. Those were presented as speculation, not additional confirmed breakthroughs.
Emad Mostaque compared the moment to Tom Hanks catching COVID: a development that made something people already knew about suddenly feel real. He forecast that AI could tackle an extraordinarily broad range of solvable, verifiable problems within a year, while acknowledging that the systems remained uneven in their abilities.
Another panelist objected to the spectacle of surprise. Laboratories had raised billions, recruited exceptional researchers and explicitly set out to build artificial general intelligence—broadly capable AI intended to transform work. “The whole thing is in the pitch deck,” he said. Where was the institutional preparation to match that ambition?
His suggested response was to create incentives that reward restraint if restraint is the goal. He also said he opposed slowing down and doubted it could be enforced. The objection was as much about preparation and credibility as about the pace itself.
A surprise about price, not capability
A different response separated the expected destination from the cost of reaching it. One panelist said he had anticipated AI solving grand challenges within a few years. What surprised him was seeing such work happen on “budgets of only a few million dollars.” That was the panelist’s characterization, rather than an itemized cost established by the research details supplied here.
On that reading, a laboratory can expect a capability eventually and still be startled when it becomes practical sooner, with less computation than anticipated. More affordable research runs would make it easier to repeat the process across additional problems—the prospect behind both the panel’s enthusiasm and Altman’s pacing argument.
OpenAI’s chief scientist Jakub Pachocki had made a more concrete proposal in his 6 September essay: enforceable safety thresholds, potentially overseen by outside auditors, governments or international institutions, with continued scaling dependent on confidence in safety.
The panel also invoked OpenAI’s charter as evidence that cooperation was not a new idea. Its actual commitment is narrower than a general promise of coordinated slowdown: if another value-aligned, safety-conscious project approaches AGI first, OpenAI says it will stop competing and assist that project, with details negotiated case by case.
For now, the reported research process still has people consolidating work between agent groups and a separate formal-checking stage after discovery. Those are concrete points of human involvement in this run. Pachocki’s proposal would add a further decision before continued scaling: whether the evidence of safety is strong enough to proceed.