Something remarkable happened in September 2026.
OpenAI announced that an internal AI system had produced a proof addressing the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems established by the Clay Mathematics Institute.
According to OpenAI, a system involving roughly 10,000 concurrent AI agents worked on the problem. The agents exchanged millions of messages and, after several days of computation, produced an analytical proof together with a formalization in Lean. OpenAI described the result as establishing finite-time singularity for a forced formulation of the three-dimensional Navier–Stokes equations.
Three days later, the Clay Mathematics Institute issued a remarkably cautious but important statement: it said that the Navier–Stokes problem had “apparently been settled.” It also stressed that evaluation and the assignment of credit would follow its deliberately slow review process. In other words, this should not yet be described as an officially awarded Millennium Prize solution.
Five days in September 2026
A capability claim, an institutional hedge, and two very different warnings — in under a week.
- 8 September OpenAI announcesA proof of finite-time singularity, produced by ~10,000 agents and formalized in Lean.
- 11 September Clay Institute hedgesThe problem is "apparently settled" — but review is deliberately unhurried, and the prize is not awarded.
- 11 September 25 Fields Medalists objectNot that AI should leave mathematics, but that the goals of AI companies and mathematicians are "severely misaligned."
- 12 September Amodei calls for pacingFrontier development should slow enough for safety mechanisms and institutions to keep up.
Whatever the final mathematical judgment, I find the event important for another reason.
Only a few years ago, much of the public discussion about generative AI concerned whether a language model could write an email, generate an image, summarize a document or produce working computer code.
Now we are discussing whether AI systems can contribute to mathematical problems that have resisted generations of human researchers.
That changes the question.
The question is no longer simply:
How useful can AI become?
It is increasingly:
How much intelligence should humanity allow itself to create?
And on this question, I am increasingly uncomfortable with the idea that our objective should be to deliberately place a permanent ceiling on AI capability.
Humanity is very far from finished
We sometimes discuss advanced AI as though humanity had already solved its major problems and our remaining challenge were simply to prevent machines from becoming too capable.
Reality looks very different.
We still cannot cure many cancers. We cannot prevent or reverse many neurodegenerative diseases. We understand only part of how our own brain works. We do not know how to substantially extend healthy human lifespan. We still face major problems in energy production, climate adaptation, materials science, food production and medicine.
The scale of one unsolved problem
Global cancer burden, IARC/WHO estimates.
These are not abstract research benchmarks. They are human lives.
And they remind us of something easy to forget during debates about AI risk:
The world without more advanced AI is not a risk-free world.
People already die because there are diseases we do not understand, medicines we have not discovered and technologies we have not yet invented.
Perhaps intelligence itself is a scarce resource
We usually think about scarcity in terms of money, energy, natural resources, compute or time.
But perhaps one of civilization’s most important scarce resources is simply: high-quality problem-solving intelligence.
A brilliant researcher has only one brain and 24 hours in a day.
A scientific team can read only a fraction of the literature relevant to its field. Researchers can investigate only a limited number of hypotheses, perform a limited number of experiments and explore a tiny fraction of possible molecules, materials or mathematical constructions.
AI begins to challenge these limits.
AlphaFold 3, for example, demonstrated high-accuracy prediction across complexes involving proteins, nucleic acids, small molecules, ions and other biomolecular structures.
DeepMind’s GNoME system used machine learning to explore enormous regions of materials space, identifying 2.2 million crystal structures predicted to be stable relative to previous databases, including 381,000 candidates on the updated stability hull.
In medicine, a generative-AI-discovered drug candidate for idiopathic pulmonary fibrosis reached a randomized Phase 2a clinical trial, with the results published in Nature Medicine in 2025.
And now the Navier–Stokes announcement gives us another glimpse of what might be possible. OpenAI reports that thousands of agents could explore different approaches simultaneously, exchange intermediate results, discard unsuccessful paths and combine promising ideas.
No individual mathematician can work as ten thousand mathematicians at once. No individual scientist can read every relevant publication, test millions of hypotheses and work continuously.
This is why I believe advanced AI should ultimately be viewed as something much larger than a productivity technology. It could become an amplifier of civilization’s capacity to discover.
But solving a problem is not the same as understanding it
The Navier–Stokes episode also revealed something important on the other side of the debate.
On September 11, 2026, 25 Fields Medal recipients published an extraordinary collective warning. Their argument was not simply that AI should disappear from mathematics. In fact, they explicitly recognized AI’s potential to accelerate genuine mathematical study.
Their concern was deeper. They argued that “the goals of the AI companies and the goals of the mathematical community are severely misaligned.”
Why?
Because mathematics is not only the production of correct answers.
A famous unsolved problem is valuable partly because attempting to solve it forces mathematicians to invent concepts, discover structures, develop methods, explain ideas and transmit understanding to the next generation.
If an AI produces thousands of correct proofs that almost nobody understands, mathematics may acquire more answers without acquiring proportionally more understanding.
The Fields Medalists therefore make a distinction that I think should become central to the broader AI debate:
Output is not the same as understanding.
This does not convince me that we should make AI less capable. It convinces me that we should become much better at deciding what we want capable AI to optimize for.
Instead of asking an AI only:
Solve this problem.
we may eventually need to ask:
Solve this problem, identify the genuinely new concepts behind the solution, connect them with existing knowledge, produce independently verifiable proofs, explain them at several levels of abstraction, identify which human ideas contributed to them, and help humans understand what has been discovered.
That is a much more ambitious form of alignment. And it is precisely the kind of alignment I would like to see.
The danger of moving too fast is real
At the same time, concerns about rapid development cannot simply be dismissed.
Anthropic CEO Dario Amodei recently argued that frontier AI development should slow sufficiently for safety mechanisms and institutions to keep pace. His proposals include independent monitoring inside frontier laboratories, industry-wide rules and international coordination.
BBC reporting also noted that Sam Altman and Elon Musk expressed support for the general call for greater caution.
OpenAI itself has recently adopted similar language. It has said that increasingly capable systems may require deliberate decisions about the pace of development and has already temporarily slowed scaling in response to cybersecurity concerns surrounding advanced models.
I understand this argument.
A system capable of accelerating medical research may also accelerate biological-weapons research. A system capable of discovering cybersecurity vulnerabilities could help defend critical infrastructure — or attack it. A system capable of operating computers autonomously at enormous scale creates risks that ordinary software does not.
These are serious questions.
But there is another question that receives much less attention.
What is the risk of moving too slowly?
When discussing AI safety, we constantly ask: What happens if AI advances too quickly?
We should also ask: What happens if useful intelligence advances too slowly?
Both columns are costs
The debate is usually conducted as though only the left column existed.
Cost of moving too fast
- Medical research capability that also accelerates biological-weapons research
- Vulnerability discovery that can attack infrastructure as easily as defend it
- Autonomous operation at a scale ordinary software never reached
- Safety mechanisms and institutions left behind by capability
- Answers produced faster than anyone can understand them
Cost of moving too slowly
- A cancer treatment arriving ten years later than it might have
- Alzheimer's, biological aging and antibiotic resistance left unsolved longer
- Clean energy, batteries and carbon removal delayed
- Scientific questions nobody has the capacity to even ask
- Discoveries postponed, with the delay charged to people now living
Imagine, hypothetically, that a future AI system could help researchers discover an effective treatment for a major cancer ten years earlier than would otherwise have been possible. Those ten years would have a human cost.
Imagine that AI could accelerate research into Alzheimer’s disease, biological aging, clean energy, battery technology, carbon removal or antibiotic resistance. Delaying such capabilities also has consequences.
Of course we do not know that AI will achieve these breakthroughs. We should not pretend that scaling today’s language models automatically produces cures for cancer or unlimited scientific discovery.
But uncertainty works in both directions.
We cannot calculate the risks of AI while assigning a value of zero to discoveries that slower development might postpone. There is an opportunity cost of caution just as there is an opportunity cost of recklessness. Our decisions should consider both.
Safety should not mean weakness
This leads me to a distinction that I think is essential: AI safety is not the same thing as limiting intelligence.
We should invest aggressively in alignment. We should develop interpretability. We should require rigorous evaluations. We should improve cybersecurity and containment. Highly autonomous systems should be monitored. Independent researchers should be able to test important safety claims. Governments and international institutions have legitimate roles when potential externalities become large.
And when there is concrete evidence that a particular capability creates an unacceptable risk that cannot yet be controlled, temporary pacing may be justified.
But this is different from deciding in advance that artificial intelligence must never become substantially more capable than human intelligence.
Humans have always built technologies that exceed our biological limitations. Cars move faster than we run. Aircraft fly. Telescopes see things our eyes cannot see. Computers calculate at speeds no human brain can approach.
We did not make these technologies useful by forcing them to remain within the boundaries of human biology. We developed ways to control them.
Intelligence is clearly more complicated because a highly autonomous intelligent system can itself plan, adapt and act.
Nevertheless, I think our fundamental challenge should be:
How can we build extremely capable intelligence while keeping meaningful human control over its goals and use?
rather than:
How can we guarantee that machines never become too intelligent?
More intelligence — but whose goals?
The Fields Medalists’ warning adds another dimension to this question.
Even a perfectly controllable AI can be directed toward the wrong objective.
A company may optimize for impressive demonstrations. A researcher may optimize for publication. A government may optimize for strategic advantage. A user may optimize for personal profit.
But humanity may need AI to optimize for very different outcomes: better health, deeper scientific understanding, clean and abundant energy, more resilient societies, greater access to education, and eventually perhaps the ability for human civilization to survive beyond Earth.
The alignment problem is therefore not only about preventing an AI from disobeying us. There is a prior question: What should we ask it to do?
That question cannot be delegated entirely to AI companies. Nor can it be answered entirely by governments, scientists or markets. It is ultimately a societal question.
Beyond Earth
I also believe our ambitions should extend further into the future.
Human civilization currently depends on a single planet. That makes us extraordinarily successful — but also extraordinarily fragile.
A genuinely self-sustaining human settlement away from Earth would require solving thousands of interconnected problems involving biology, medicine, radiation protection, energy, agriculture, robotics, manufacturing, recycling and autonomous maintenance. Many of these systems would need to function for long periods without immediate assistance from Earth.
No single human discipline can solve all of this. Perhaps sufficiently advanced AI could help integrate knowledge across fields at a scale that human organizations struggle to achieve.
This does not mean that AI will take us to Mars. It means that if humanity wants ambitions measured in centuries rather than product cycles, greater intelligence may be one of the resources we need most.
Do not confuse caution with a ceiling
I therefore find myself in an unusual position.
I agree with many AI researchers that advanced AI could become dangerous. I agree that frontier systems require stronger monitoring. I agree that there may be moments when development should temporarily slow because safety mechanisms are clearly inadequate. And I agree with the Fields Medalists that producing answers without understanding, attribution or human intellectual development would be a poor vision of scientific progress.
But none of these points leads me to the conclusion that humanity should permanently constrain how much intelligence it can create.
Quite the opposite. They make me think that we need to become better at directing intelligence, rather than merely limiting it.
The Navier–Stokes episode may eventually prove to be historically important, or it may turn out to be one step in a much more complicated scientific story.
But imagine where the trajectory leads if systems improve another ten times, one hundred times or one thousand times in their ability to perform scientific reasoning.
What questions could they help us ask? What diseases could they help us understand? What mathematics could they uncover? What new forms of energy or matter could they help us discover? What parts of the universe could they help us reach?
We do not know.
That uncertainty is precisely why both reckless acceleration and permanent prohibition seem unsatisfactory to me.
Humanity is not finished
Humanity still has an extraordinary amount left to understand.
Cancer remains with us. Aging remains with us. The brain remains largely mysterious. Our civilization remains confined to Earth. Most of the universe remains beyond our reach. And countless scientific questions remain unanswered.
Artificial intelligence may not solve all of these problems. Some may prove far harder than we imagine. Some may require experiments, institutions, political choices or human qualities that intelligence alone cannot provide.
But abandoning the possibility of much greater intelligence because it might become dangerous would also carry a risk.
My preferred objective is therefore neither accelerate at any cost nor slow down at any cost. It is more demanding:
Build as much beneficial intelligence as humanity can responsibly control, and improve our ability to control it as aggressively as we improve the intelligence itself.
Safety and capability should not be enemies. Understanding and automation should not be enemies. Human scientists and AI should not have to be enemies. The objective should be to make them reinforce one another.
Because the final purpose of artificial intelligence should not be to produce better chatbots, win benchmarks or replace human intellectual work. It should be to expand the frontier of what humanity can understand and what humanity can achieve.
Humanity does not need less intelligence.
We need more intelligence — and the wisdom to decide what to do with it.
References
- OpenAI. On the Navier–Stokes Millennium Prize Problem. September 8, 2026.
- Clay Mathematics Institute. Navier–Stokes Announcement. September 11, 2026.
- Collective of 25 Fields Medalists. A Severe Misalignment of AI in Mathematics. French version published in Le Monde, September 11, 2026.
- Castelvecchi, D. OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’. Nature, September 2026.
- BBC News. Anthropic boss Dario Amodei calls for AI development to slow down. September 12, 2026.
- Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).
- Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023).
- Xu, Z. et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine 31, 2602–2610 (2025).
- IARC/WHO. Global cancer statistics 2024. Published July 8, 2026.