The Forecast Factory
In 1922, Lewis Fry Richardson imagined a vast hall, a theatre in the round, filled with sixty four thousand human beings, each computing the weather for one small patch of atmosphere. A conductor stood at the centre on a raised dais, shining a spotlight on any section that fell behind. Richardson called it a “forecast factory.” He published the idea in a book ahead of its time, and died in 1953, three years after a machine called ENIAC had vindicated his method.
Last week, OpenAI filled a room that Richardson would have recognised. Not with people but with ten thousand artificial intelligence agents, each trained on a model the company says is significantly more capable than anything it has released to the public. They set the agents to work on the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems posted by the Clay Mathematics Institute in the year 2000, each worth a million dollars to the first solver. The equations describe how fluids move. They underpin weather forecasting, aircraft design, blood flow modelling, and ocean current prediction. For ninety years, a complete proof of their smoothness, the guarantee that solutions do not blow up into infinities, has eluded mathematicians.
Eighty eight hours later, the agents had exchanged nearly three million messages and consumed a hundred and thirty billion output tokens. OpenAI announced it had resolved two of the four statements required by the Millennium Prize. The company estimated the effort would have cost roughly ten million dollars at its own published pricing. It then added, with studied modesty, that it did not intend to claim the prize.
The modesty did not last the afternoon. On the same day, Tristan Buckmaster, a mathematics professor at New York University, released a statement claiming that he and Levent Alpöge, a mathematician employed by Anthropic, had been working toward solutions to the same problem. They had been using OpenAI’s own Codex tool to assist their research. Buckmaster said he had learned on 3 September that information about their progress had reached OpenAI. He said the company did not begin its own work on the Navier-Stokes equations until after that information arrived. He had not yet read OpenAI’s full proof, he said, but felt compelled to go public because the alternative was to let a sequence of announcements say something he knew to be false.
[https://www.wired.com/story/openai-navier-stokes-math-discovery-academics/]
OpenAI congratulated what it called the “concurrent work” of Buckmaster and Alpöge. It said it had not seen any of their research until they published it. It acknowledged, in a sentence that deserves to be read twice, that while unlikely it could not rule out that anonymised data derived from their usage of its products had helped improve its models.
That sentence is the story.
The question is not whether ten thousand agents can prove a theorem. The question is whether the company that sold the mathematicians their tools then used the residue of that work to beat them to the result. It is a question that would have appalled Richardson, who destroyed his own research on atmospheric turbulence when he discovered it was being exploited for poison gas dispersal. Richardson believed that knowledge carried obligations. He would not have recognised a world in which the obligation ran in the other direction, in which the user owed a duty of caution to the tool.
Israel has reason to pay attention. The country’s artificial intelligence sector, anchored by the Technion, the Weizmann Institute, and the sprawling alumni network of Unit 8200, depends on a compact between talent and platform. Israeli researchers publish openly, train models on American cloud infrastructure, and license commercial tools from the same handful of companies they compete against in publication and in patent filings. The Buckmaster affair, whatever its resolution, is a warning that the compact may have a trapdoor. If a platform can absorb the working patterns of its most capable users, strip the identifying marks, and feed what remains back into a competing model, then the researchers are not customers. They are inputs. For a country whose principal export is human capital refined into intellectual property, that distinction is existential.
The Navier-Stokes equations describe, among other things, the transition from laminar to turbulent flow, the moment at which a smooth, predictable current breaks apart into chaos. Mathematicians have spent ninety years trying to prove that the equations do not produce singularities, points at which the mathematics itself collapses. The irony is that the race to solve the problem may have produced a singularity of a different kind, a point at which the distinction between tool and competitor collapses, and the platform becomes the thing it was built to serve.
Richardson’s forecast factory was never built. It did not need to be. The forecast arrived anyway, decades later, on machines he never imagined, solving problems he had sketched on paper during a war he refused to fight. The factory was always a thought experiment, a room filled with people who did not exist.
OpenAI’s room held ten thousand agents that did exist, briefly, for eighty eight hours. They exchanged three million messages. They consumed ten million dollars of compute. Then the proof was written, the press release drafted, and the agents switched off.
The room is empty. The equations remain.
Companion Appendix: Applications in Finance and Geopolitics
The Navier-Stokes equations are not confined to wind tunnels and weather stations. Their mathematical structure, and the new proof techniques that a resolution would unlock, have direct and underappreciated applications across financial economics and geopolitical risk analysis. The table below maps six fluid dynamics concepts to their counterparts in each domain.
| Fluid dynamics concept | Finance application | Geopolitical application |
|---|---|---|
| Turbulence cascade (energy dissipating from large to small scales, per Kolmogorov 1941) | Mandelbrot’s multifractal models of asset returns. Volatility clustering across time horizons mirrors the energy cascade. A smoothness proof would tighten the mathematical foundations under heavy tailed return distributions that currently lack the rigour of Gaussian frameworks. | Escalation dynamics in conflict. A localised incident (a border skirmish, a cyberattack) cascades through alliance networks and media cycles in a pattern structurally similar to turbulent energy transfer. Better cascade mathematics improves early warning models. |
| Laminar to turbulent transition (the moment smooth flow breaks into chaos) | Regime switching in markets. The transition from low volatility (“laminar”) to crisis (“turbulent”) regimes is poorly modelled by existing threshold models. Navier-Stokes regularity results could yield sharper criteria for when and why markets break. | State failure and revolution. Political scientists model the transition from stable governance to institutional collapse using analogous threshold dynamics. The Arab Spring, the Soviet dissolution, and the 2019 protest waves all exhibited abrupt laminar to turbulent transitions. |
| Singularity and blowup (the mathematical question of whether solutions can become infinite in finite time) | Flash crashes and liquidity black holes. The 2010 Flash Crash and the 2020 Treasury dislocation are financial singularities: prices moved toward infinity (or zero) faster than any market mechanism could arrest. Understanding when Navier-Stokes solutions blow up maps onto understanding when orderly price formation collapses. | Nuclear escalation thresholds. Deterrence theory assumes rational actors will stop short of catastrophic exchange, but the mathematical question is identical: can the system reach a singularity, a point of no return, from smooth initial conditions? |
| Viscosity as a stabilising force (viscosity damps turbulence; without it, the Euler equations are far more unstable) | Regulation and transaction costs as market viscosity. Tobin taxes, circuit breakers, and capital requirements function as viscosity in financial systems. The proof’s finding that Navier-Stokes solutions can blow up even with viscosity present suggests that regulation alone may not prevent financial singularities. | Diplomacy and institutional friction as geopolitical viscosity. The UN, the WTO, and bilateral treaties function as viscosity in the international system. The proof implies that even with these stabilising institutions in place, the system can still produce blowup under certain initial conditions. |
| Nonlinear PDE techniques (the proof methods, not just the result) | Derivative pricing and free boundary problems. The Black-Scholes equation is a linear PDE; real markets require nonlinear extensions (stochastic volatility, jump diffusion). New techniques for Navier-Stokes regularity could yield better existence and uniqueness results for the PDEs that quants actually solve. | Contagion and diffusion models. Financial contagion through interbank networks and geopolitical contagion through alliance obligations are both modelled using nonlinear diffusion equations structurally related to Navier-Stokes. |
| Continuum modelling of flow (treating discrete particles as a continuous fluid) | Market microstructure. Almgren-Chriss optimal execution models frame the aggregate impact of order flow as a continuous process. Better turbulence mathematics could improve models of how liquidity flows and, critically, where it breaks down, which is precisely the regime that matters most for institutional risk. | Refugee and capital flows. Researchers have modelled mass displacement and capital flight using diffusion and gravity models structurally analogous to fluid dynamics. Improved Navier-Stokes mathematics would sharpen predictions of when orderly flows become chaotic surges. |
A note on the proof’s direction. OpenAI’s claimed result is that the Navier-Stokes equations can blow up: smooth initial conditions can produce a singularity in finite time. This is the negative direction, and its applied implications are sobering. It suggests that in any system governed by fluid-like dynamics, whether physical, financial, or geopolitical, catastrophic breakdown is not a pathology but a mathematical possibility inherent in the governing equations, even when stabilising forces (viscosity, regulation, diplomacy) are present. The question for practitioners is not whether blowup can happen, but under what initial conditions it will.