Math rarely produces a public scandal. The fight between OpenAI and parts of the scientific community over its claimed solution to the Navier–Stokes problem may be the exception. The immediate dispute concerns priority, provenance, and who or what deserves credit for the breakthrough. For most people, that will feel like academic theatre. Look closer, however, and the argument exposes a question that should concern every leader: as AI becomes capable of solving problems once beyond human reach, will it extend our ability to think, or slowly erode our ability to decide what is worth thinking about?
Navier–Stokes is one of seven Millennium Prize Problems selected by the Clay Mathematics Institute in 2000 as some of mathematics’ most important unsolved challenges. Each carries a prize of $1 million. Only one has been officially resolved: the Poincaré Conjecture, which concerns the topology of three-dimensional spaces. Russian mathematician Grigori Perelman solved it in 2003, then famously declined both the Fields Medal and the prize money. The Navier–Stokes problem has proved more stubborn. Its proposed solution raises a more unsettling possibility: machines may produce breakthroughs faster than people can turn them into ideas others can understand and use.
Henri Poincaré understood that the value of science depended on what it enabled us to think and discover next. Collecting facts, he argued, was like gathering stones. However large the pile, someone still had to work out how to build a house. Scientific progress came from the thinking that connected observations, revealed a pattern, and allowed us to anticipate something we had yet to see. Yet Poincaré recognized that our choice of worthwhile problems reflected the limits of our minds. A greater intelligence might find something simple and useful where we saw only hopeless complexity. He acknowledged the possibility, then dismissed its practical relevance: “we can not use that superior mind, but only our own.” AI changes that condition. We can increasingly draw on intelligence beyond our own, which makes the development of our own understanding a choice we can no longer take for granted.
Seen from this vantage, there is something compelling about Terence Tao’s warning about a “severe misalignment” in mathematics. The declaration, initially signed by 25 Fields Medallists, asks us to examine what mathematical discovery is for.. Famous mathematical problems give researchers something to pursue while they develop ideas that other people can understand, teach, and carry into new fields. Solving one has traditionally signaled that an important advance in understanding has occurred. AI, however, now makes it possible to produce the result while leaving much of that human work unfinished. A proof can be correct long before its central ideas become accessible.
Tao’s argument reaches beyond the prospect of mathematicians losing their jobs. Ultimately, we have to decide what we want from discovery itself. Is success a growing inventory of secrets uncovered, or an expanding capacity to understand the world and recognize what else deserves our attention? The two have often advanced together. Their separation is what makes this moment complicated.
What I found most interesting about the whole controversy was the detail about how OpenAI actually co-ordinated the workflow between their teams and their AI models. Their researchers assigned groups of AI agents to the remaining unsolved Millenium problems, including easier questions that might open a route into harder ones. When nearly 100 agents made progress on the Euler equations after about 50 hours, the humans recognized its significance and redirected resources toward Navier–Stokes. They also used Codex to consolidate insights across groups and feed them back into the search. The successful group involved roughly 10,000 concurrent agents. The Navier–Stokes effort consumed approximately 130 billion output tokens. OpenAI disclosed no dollar cost, but directing computation on that scale made the researchers’ judgment consequential.
Even with thousands of agents available, someone had to decide where the next hour of effort belonged. Token costs aside, the attention of the world’s best AI engineers is probably even more scarce. The careful allocation of both machine, and human taste matters. At OpenAI, researchers chose where to begin, then revised their priorities in response to what the machines found. To me, this suggests an expanding frontier of shared effort. Human judgment was operating through a research system of extraordinary reach, with machine discoveries changing what the humans believed was possible. The next question is what the researchers learned through that process, and whether it made them better at choosing where to go next.
Business leaders face the same design challenge: how to make working with AI develop the judgment their organizations will need next. A magical AI answer engine can make employees more productive while dismantling the mental dojos in which they learned to frame problems. The junior analyst who works through a complicated sales forecast is also learning which assumptions deserve closer scrutiny. If a machine supplies the solution immediately, the organization needs another way to develop that judgment. This will require deliberate interventions: giving junior colleagues a chance to form a hypothesis before consulting AI, asking them to investigate disagreements, and having experienced people explain why an apparently convincing answer worries them.
The balance matters. Time saved on repetitive work should create room for more demanding thought, including investigations that would previously have been impossible. Rewards will also need attention. The person who steps back and discovers that a team is solving the wrong problem may create more value than the person who generates twenty polished answers. Leaders need to recognize that contribution without turning curiosity into another reporting requirement. The way we arrange human and machine judgment should help people become better at deciding where effort belongs.
Poincaré’s insight offers a useful test for what comes next: does a discovery expand our capacity for further thought? AI may help us reach answers that would otherwise remain beyond us. It may also lead us towards questions whose significance we initially struggle to understand. Learning to navigate that unfamiliar territory could become a new kind of intellectual apprenticeship, developing capacities we do not yet possess. That is an ambition worth designing our organizations around.
The most valuable answers will change what we become capable of asking.