For nearly a century, one of mathematics’ most formidable questions defied the world’s greatest minds. An army of AI agents may have solved it in less than four days. OpenAI deployed around 10,000 autonomous agents, powered by an internal model not yet available to the public, which produced a solution to the Navier–Stokes Millennium Prize Problem in just 88 hours. They exchanged 2.7 million messages and generated approximately 130 billion output tokens; formalising and verifying the proof in Lean required only another 17 hours. Although Lean can certify every step of the formal proof, mathematicians must still confirm that the statement encoded in the system corresponds exactly to the original problem. If they do, this will not merely be the most important proof ever produced by AI: it will mark the moment when machines entered the highest realm of mathematical discovery. https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
The achievement is, however, surrounded by a profoundly disturbing dispute. OpenAI acknowledges that it launched the operation after hearing rumours that two Millennium Prize Problems had been solved, rumours subsequently linked to the still-unpublished work of Tristan Buckmaster and Levent Alpöge. According to Buckmaster, they had used OpenAI models and placed all their project drafts in Codex sessions. OpenAI states that neither its researchers nor its agents specifically accessed this work, but admits that it cannot rule out the possibility that de-identified data derived from their use of its products helped improve the models. Both groups also relied heavily on the mathematical programme previously developed by Diego Córdoba and Luis Martínez-Zoroa, further complicating the attribution of priority and scientific credit.
This episode extends far beyond mathematics. The same company that provides scientists with research tools also owns the models, computing infrastructure and financial resources required to transform itself almost instantly into their competitor. Once it becomes aware of a promising opportunity, it can mobilise thousands of agents and spend several million dollars, resources beyond the reach of any academic research team. Even if no improper access occurred, the asymmetry of power has become impossible to ignore.
This episode gives disturbing concrete form to the civilizational choice at the heart of my manifesto, Civilizational Futures. AI can serve as “intellectual scaffolding”, liberating intelligence and accelerating discovery; but when a handful of corporations controls the models, computing infrastructure and armies of autonomous agents—while receiving scientists’ unpublished work—it acquires the power not merely to assist science, but to appropriate it. The Navier–Stokes episode may therefore be both a triumph of machine-augmented reason and a warning that humanity’s greatest instrument for distributing intellectual power could become its most formidable mechanism for concentrating it.
In this emerging order, the scientist supplies the question, the drafts, the corrections and perhaps even the decisive insight; the platform absorbs the value, unleashes an army of agents and emerges claiming the breakthrough. The question is therefore no longer whether AI can conduct science. It is whether scientists are paying to train the very machines and enrich the very corporations, that may strip them of priority, authorship and control over their own discoveries. Until enforceable safeguards exist, entering unpublished research into a commercial AI system should be treated for what it may become, a transfer of scientific power from its creator to a vastly more powerful corporate rival.
PS - That warning has now been corroborated from inside the AI industry itself. Jacob Coxon, who has just resigned from Anthropic after a stint at OpenAI, explicitly cited the Navier–Stokes breakthrough while accusing both companies of racing towards self-improving superintelligence and “gambling with our lives.” Evan Hubinger, a current lead in Anthropic’s alignment division, said Coxon was right—and conceded that the company still has no plan for aligning superintelligence. When the builders themselves speak this way, leaving such power in private corporate hands can no longer be defended as innovation policy. It becomes the civilizational gamble at the heart of Civilizational Futures: whether AI liberates human intelligence or concentrates unprecedented power in a handful of corporations answerable to no one.
