OpenAI’s AI Agents Tackle a Major Math Problem, but Terence Tao Warns of a Bigger Risk

-When Terence Tao raises concerns about artificial intelligence and mathematics, it’s worth paying attention.

The UCLA mathematics professor and Fields Medal winner has spent decades working at the highest levels of modern mathematics. So when he discusses what increasingly capable AI systems could mean for mathematical research, the issue is more complicated than the usual debate over whether machines will eventually replace human mathematicians.

That debate became particularly relevant on September 8, following an announcement from OpenAI, the company behind ChatGPT. OpenAI said its AI systems had successfully constructed a challenging mathematical example related to the famous Navier–Stokes equations.

The achievement is impressive. But Tao’s reaction highlights a different question: What happens to mathematical research when AI can explore thousands of ideas simultaneously and potentially complete work that another researcher has only just begun?

Why the Navier–Stokes Problem Matters

The Navier–Stokes equations are among the most important mathematical tools used to describe the behavior of fluids.

They can be used to model things such as the movement of water and air, currents, flow patterns, and turbulence. The equations themselves are well established, but mathematicians still don’t fully understand every aspect of their behavior.

One of the most famous questions concerns whether the equations can develop a singularity.

In simplified terms, imagine starting with a smooth and well-behaved fluid. The question is whether the mathematical model can evolve into a situation where the fluid’s velocity becomes unbounded within a finite amount of time.

A real-world fluid obviously cannot accelerate to infinite speed. But mathematical equations can produce singularities, and determining whether such a singularity can actually occur under the Navier–Stokes equations is a fundamental problem in mathematics.

The question is so important that it was selected as one of the Millennium Prize Problems established by the Clay Mathematics Institute.

What OpenAI Claims Its AI Systems Found

OpenAI says its researchers constructed a specific scenario in which a fluid initially at rest is subjected to a smooth external force.

According to the company, its AI system produced a mathematical proof establishing the desired result. The proof was then checked using Lean, a formal verification system designed to verify mathematical statements and the logical steps used to prove them.

That distinction matters.

This isn’t simply an AI model generating an answer that looks mathematically convincing. Formal proof systems such as Lean can check whether the individual logical steps satisfy the rules of the underlying mathematical system.

OpenAI also isn’t claiming that it has solved the entire Millennium Prize problem or that it is entitled to the associated prize money.

Instead, the result represents a specific mathematical construction connected to the broader Navier–Stokes problem.

Around 10,000 AI Agents Worked on the Problem

Perhaps the most striking part of OpenAI’s announcement is the scale of the experiment.

The company says it deployed approximately 10,000 AI agents to explore potential solutions. These agents were able to investigate different approaches and exchange their findings, while an internal model coordinated the overall process.

The system reportedly reached its solution after roughly 88 hours, followed by an additional verification stage.

That approach is fundamentally different from asking a single chatbot to solve a difficult mathematics problem.

Instead, thousands of software agents can effectively explore different branches of a problem at the same time. Some approaches will fail, others may produce partial results, and a smaller number can eventually lead to something useful.

For mathematical research, that could become a powerful new tool.

But it also introduces problems that have little to do with whether AI is “smarter” than humans.

Terence Tao Sees a Problem Beyond AI vs. Humans

Tao’s concerns focus on how mathematical research itself could change.

A major part of mathematics isn’t simply solving problems that have already been defined. Researchers must first determine which questions are worth asking.

A useful mathematical problem can reveal something fundamental, connect seemingly unrelated areas, or provide techniques that can later be applied elsewhere.

That’s very different from generating thousands of solutions to arbitrary mathematical exercises.

An AI system capable of solving enormous numbers of predefined problems could therefore become extremely productive without necessarily replacing the human ability to identify the most important questions.

And that distinction could become increasingly important as AI systems improve.

What If AI Finishes Someone Else’s Research?

Tao also points to a more immediate concern for mathematicians.

Imagine that a researcher publishes or shares a promising idea. Another research group then feeds that idea into a large AI system equipped with thousands of agents.

Instead of spending months or years developing the idea, the second group could potentially have AI systems explore thousands of variations simultaneously and finish the work much faster.

The original researcher could suddenly find themselves watching someone else publish the completed result based on an idea they had only recently shared.

That creates an uncomfortable incentive.

Researchers might become increasingly reluctant to share promising ideas before they are fully developed.

And that’s potentially harmful to mathematics, where conversations, collaboration, informal discussions, unfinished arguments, and exchanging partial results have traditionally played an important role.

AI Research Needs to Show More Than the Final Answer

Tao has another criticism that could become increasingly important as AI-generated mathematics becomes more capable.

AI companies tend to emphasize successful results. But researchers can learn just as much from failed approaches.

If an AI system tries thousands of methods and only one succeeds, knowing which approaches failed could be extremely valuable.

It could tell mathematicians:

• Which strategies are unlikely to work
• Which assumptions lead to dead ends
• Which intermediate techniques are reusable
• How the successful approach was discovered
• Which parts of the problem remain difficult
• How future researchers might attack similar problems

Without that information, an AI-generated proof can look like a black box with a successful answer at the end.

Researchers see the destination, but not necessarily the path.

Publishing the Failed Attempts Could Be More Valuable Than It Sounds

For human mathematicians, an unsuccessful approach isn’t necessarily wasted effort.

A failed proof can reveal why a particular assumption doesn’t work. An abandoned technique can inspire another researcher to modify it. Even a dead end can help establish which parts of a problem are genuinely difficult.

AI systems could generate enormous amounts of this information.

But if companies publish only the final successful proof, much of the potentially useful research knowledge remains hidden.

Sharing the experiments, failed approaches, intermediate discoveries, and successful reasoning could therefore make AI a much more valuable research partner.

Instead of simply announcing that an AI system solved a problem, researchers could see how the system explored the mathematical landscape to get there.

AI Could Change Mathematical Research Without Replacing Mathematicians

The OpenAI experiment illustrates both sides of the AI revolution in mathematics.

On one hand, thousands of AI agents working in parallel could dramatically accelerate certain types of research. Problems that previously required enormous amounts of human trial and error could potentially be explored much faster.

On the other hand, mathematical progress isn’t measured only by the number of problems solved.

The harder question is determining which problems deserve to be solved in the first place.

That’s where human mathematical judgment remains particularly important.

The future may therefore look less like AI replacing mathematicians and more like mathematicians working with increasingly powerful systems capable of exploring vast numbers of possibilities.

But for that future to benefit the broader mathematical community, transparency will matter.

If AI companies publish only their victories, researchers may gain impressive answers without gaining the knowledge needed to reproduce, understand, or extend them.

If they also publish the failures and the reasoning that led to successful proofs, AI could become something much more valuable than an automated answer machine.

It could become a genuinely new tool for mathematical discovery.

Final Thoughts

OpenAI’s Navier–Stokes work demonstrates just how far AI-assisted mathematics is beginning to move. Deploying thousands of agents to explore a difficult problem in parallel would have sounded unrealistic only a few years ago.

But Terence Tao’s warning points to a deeper issue.

The biggest impact of AI on mathematics may not be that machines solve problems faster than humans. It may be that the entire process of discovering, sharing, and developing mathematical ideas changes.

If researchers become afraid to share unfinished ideas because AI systems can rapidly turn them into completed results, collaboration could suffer.

The solution may be greater transparency: publish not only the final proof, but also the experiments, failed attempts, and reasoning that produced it.

That would give mathematicians something more valuable than a finished answer: a map of how the answer was found.