Imagine dropping a small stone into a perfectly calm lake.
You see the ripples moving outward.
It looks simple.
Now imagine trying to predict the exact position of every ripple, every tiny swirl, every change in speed, and every interaction between those ripples several minutes later.
Suddenly, it doesn't look so simple.
Now make the problem even harder.
Instead of a calm lake, imagine ocean waves, smoke coming out of a chimney, air moving around an aircraft, water flowing through a pipe, or turbulent air around an F1 car.
All of these are fluids.
And fluids are surprisingly difficult to predict.
For more than a century, scientists have had a set of equations that describe their motion remarkably well. They are called the Navier–Stokes equations.
They are used everywhere in modern engineering and science.
And yet, there is a basic mathematical question about these equations that remained unanswered for decades:
If a fluid starts out smooth and well-behaved, will it always remain smooth?
Or can the mathematics eventually produce something so extreme that the solution effectively breaks down?
That question became one of the famous Millennium Prize Problems.
And now, in September 2026, OpenAI says an internal AI system has produced a solution.
But the interesting part isn't simply:
"AI solved a 100-year-old problem." The interesting part is understanding what the problem actually is, how we got here, what OpenAI claims to have proved, what other mathematicians were working on, and how we should decide whether an AI-generated mathematical discovery is actually correct.
Let's start from the beginning.
To understand why Navier–Stokes became necessary, we need to go back to one of the biggest ideas in physics.
In the 17th century, Isaac Newton developed his laws of motion.
His famous second law is usually written as:
F = ma
Force equals mass multiplied by acceleration.
This is an incredibly powerful idea.
If you know the forces acting on an object, you can calculate how its motion changes. It works beautifully for things like a thrown ball, a moving car, or a planet orbiting the Sun.
But there is a problem.
What happens when the thing you're studying isn't one object?
What happens when it is water?
Or air?
A glass of water isn't really one object.
It contains an enormous number of molecules.
Each molecule is moving.
Each molecule interacts with neighboring molecules.
And all of those interactions collectively produce what we see as fluid motion.
Trying to calculate the position and velocity of every individual molecule would be hopeless for any practical problem.
So scientists needed another way to think about fluids.
Instead of following every molecule, imagine dividing the fluid into extremely tiny imaginary boxes.
You don't care about the exact molecule inside each box.
You care about the average behavior of the fluid in that small region.
You can ask:
This is the basic mathematical view behind fluid dynamics.
And it is an incredibly powerful simplification.
We don't need to know what every molecule is doing.
We can describe the large-scale behavior of the fluid.
In the 18th century, mathematician and physicist Leonhard Euler developed equations describing the motion of an ideal fluid.
Euler's equations were a major step forward.
But there was something missing.
Real fluids aren't perfect.
They have viscosity.
And viscosity turns out to be extremely important.
You've probably experienced viscosity without knowing the word.
Pour water.
It flows quickly.
Pour honey.
It moves much more slowly.
The difference is largely related to viscosity.
In simple terms:
Viscosity is a measure of a fluid's resistance to flowing or to layers of fluid sliding past each other.
Think about putting your hand into water and moving it.
The water pushes back.
Now imagine doing the same thing in something much thicker.
There is more resistance.
Inside a flowing fluid, different layers can move at different speeds.
Viscosity describes the internal friction between those layers.
And this becomes very important in Navier–Stokes because viscosity tends to smooth out differences in motion.
If one region of a fluid is moving much faster than another, viscosity tends to reduce that difference. You can almost think of it as nature's internal smoothing mechanism.
In the 19th century, Claude-Louis Navier and later George Gabriel Stokes developed equations that incorporated this viscous behavior into the mathematical description of fluid motion.
This gave us what we now call the Navier–Stokes equations.
They combine several basic ideas:
Velocity — how fast the fluid is moving.
Acceleration — how that motion changes.
Pressure — how pressure differences push the fluid around.
Viscosity — how the fluid resists changes in motion.
External forces — things such as gravity or other forces acting on the fluid.
Put very simply, Navier–Stokes is trying to answer:
Given what the fluid is doing now, what will it do next?
That sounds straightforward.
It isn't.
You may see the Navier–Stokes equation written like this:
Change in fluid motion = Fluid’s own motion + Pressure + Viscosity + External forces
Don't panic.
We don't need to understand every symbol to understand the story.
Very roughly:
The left side describes how the fluid's motion changes.
The right side contains some of the things causing that change: pressure, viscosity, and external forces.
The important part isn't memorizing the equation.
It is understanding that the equation creates a feedback system.
The fluid's current motion affects its future motion.
And its future motion affects the flow again.
That is where things become difficult.
Let's imagine a tiny swirl inside a fluid.
That swirl moves.
As it moves, it changes the surrounding fluid.
The surrounding fluid then pushes back on the swirl.
The swirl stretches.
It may become smaller and faster.
That changes the surrounding flow again.
Now imagine millions of these interactions happening simultaneously in three dimensions.
This is where nonlinearity enters the story.
"Nonlinear" sounds intimidating, but the basic idea is simple.
In a linear system, doubling something might simply double the result.
In a nonlinear system, doubling the input can produce a much more complicated change.
Fluid motion contains nonlinear interactions.
And those interactions can amplify small changes.
This is one reason turbulence is so difficult.
Think about smoke rising from a cigarette.
At first, the smoke may rise in a fairly smooth column.
Then, a few centimeters higher, it starts wobbling.
Then it develops little swirls.
Those swirls break into smaller swirls.
The flow becomes chaotic.
That is turbulence.
Turbulence is not simply "random movement."
It is the result of complicated interactions across many different sizes and time scales.
Large structures can break into smaller structures.
Those smaller structures can break into even smaller ones.
This process is sometimes called an energy cascade.
Energy moves from larger scales of motion toward smaller scales.
And suddenly we have a huge mathematical problem.
How do you describe all of this exactly?
You've probably heard of the butterfly effect.
The popular version is:
A butterfly flaps its wings and eventually causes a hurricane.
That sentence is more of an illustration than a literal scientific claim.
The real idea is about sensitivity to initial conditions.
If two systems start almost exactly the same, their future behavior can eventually become very different. Fluid systems can show this kind of sensitivity.
A tiny difference in the initial flow can grow over time.
That means even if our equations are perfect, prediction can still become difficult.
And this gives us an important distinction:
A system can be governed by deterministic equations and still be extremely difficult to predict in practice.
The equations may tell us exactly how the system evolves.
But tiny differences in the starting conditions can make long-term prediction incredibly difficult.
Pretty much everywhere fluids matter.
Engineers use fluid dynamics to understand how air moves around wings and aircraft bodies.
That helps with questions about lift, drag, stability, and efficiency.
Air moving around a car affects aerodynamic drag and downforce.
Engineers can simulate airflow around different designs before physically building them.
The atmosphere is a giant fluid system.
Understanding air movement is fundamental to weather and climate modelling.
Water flowing around a ship affects resistance and fuel efficiency.
Blood is a fluid.
Fluid dynamics can help researchers study blood flow through arteries and other biological systems.
Pipelines, pumps, turbines, engines, cooling systems and chemical processes all involve fluid movement.
Airflow around spacecraft and atmospheric vehicles creates complex fluid-dynamic problems.
So Navier–Stokes isn't some equation that exists only inside mathematics departments.
It is part of the mathematical foundation behind a huge amount of modern engineering.
The Clay Mathematics Institute describes it simply: these equations govern the flow of fluids such as water and air.
This is the part people often misunderstand.
Engineers can use Navier–Stokes.
Computers can simulate Navier–Stokes.
Scientists can calculate approximate solutions.
So what exactly remains unsolved?
The mathematical question is much deeper.
Imagine starting with a perfectly smooth fluid.
No infinite values.
No weird discontinuities.
Everything behaves nicely.
Now let the Navier–Stokes equations evolve that fluid forward in time.
The question is:
Will the solution remain smooth forever?
Or:
Can a singularity develop in finite time?
That is the famous existence and smoothness problem.
The Clay Mathematics Institute describes the central questions as whether solutions exist and whether they remain smooth; the official problem concerns three-dimensional incompressible flow.
This word sounds dramatic.
But in mathematics, it has a very specific meaning.
A singularity is a point where the mathematical solution develops behavior that becomes unbounded or otherwise ceases to remain regular.
In this particular problem, one possible scenario is that the velocity becomes arbitrarily large in a finite amount of time.
Again, this doesn't mean a real piece of water literally travels at infinite speed.
It means the mathematical solution develops an infinite quantity.
That raises a fundamental question:
Does the mathematical model permit this to happen from perfectly reasonable starting conditions?
Nobody had been able to prove the answer.
You might wonder:
"If we can study fluid flow on a computer, why can't mathematicians simply solve the equation?"
One major reason is three-dimensionality.
A two-dimensional fluid is already complicated.
A real fluid has three dimensions.
Now a vortex can stretch, twist, bend and interact with other vortices in ways that simply don't exist in the same form in two dimensions.
Imagine taking a rubber band and stretching it.
Now imagine that rubber band is actually a spinning tube of fluid.
Stretching the vortex can intensify the rotation.
That creates more complicated motion.
That complicated motion feeds back into the rest of the fluid.
This is one of the key difficulties in three-dimensional fluid dynamics.
In 2000, the Clay Mathematics Institute announced seven Millennium Prize Problems.
Each problem came with a $1 million prize.
The goal wasn't to create seven difficult puzzles for mathematicians to solve for fun.
These were fundamental questions at the edge of mathematical knowledge.
Navier–Stokes was one of them.
The prize exists partly because solving such a problem can create new mathematical ideas that become useful far beyond the original question.
And there is an interesting historical detail here.
The Navier–Stokes equations themselves are more than 150 years old.
But the precise mathematical question about existence and smoothness has remained open for roughly 90 years.
So the equation isn't "100 years old and nobody knows how to use it."
Quite the opposite.
We use it constantly.
What remained mysterious was what the equations guarantee mathematically under all allowed conditions.
Now we arrive at September 2026.
OpenAI announced that an internal AI system had produced what it describes as a solution to the Navier–Stokes existence and smoothness problem.
According to OpenAI, its system produced an analytical proof showing that a smooth three-dimensional fluid can develop a singularity in finite time.
The proposed mechanism involves a vortex.
The vortex becomes increasingly elongated and concentrated.
Its central region shrinks while the velocity increases.
At the same time, the construction keeps the total energy finite.
That is important because simply inserting an infinite force into the system would not demonstrate the kind of breakdown the problem is asking about.
OpenAI says the singular behavior instead emerges from the dynamics of the fluid itself.
If the argument survives mathematical scrutiny, it would mean the correct answer to the Millennium Problem is not: "Smooth solutions always remain smooth."
Instead, it would establish that finite-time breakdown can occur.
This part is almost as interesting as the mathematics.
OpenAI says it used a large system of coordinating AI agents.
The agents explored different approaches, communicated with each other, ran code and consolidated useful ideas.
The group working on Navier–Stokes involved roughly 10,000 concurrent agents.
OpenAI says the agents generated around 2.7 million messages and approximately 130 billion output tokens during the Navier–Stokes effort.
The result reportedly emerged after about 88 hours.
Then a separate formalization process used Lean to verify the proof structure, taking another 17 hours.
That is very different from asking ChatGPT:
"Can you solve Navier–Stokes?"
This was closer to building a giant virtual research team.
Suppose I give you a 100-page mathematical proof.
You read it.
It looks convincing.
But somewhere on page 67, I accidentally make a logical mistake.
You might not notice.
Computers can help here.
Lean is a formal proof system.
Instead of simply writing:
"Therefore this result follows."
you express mathematical statements in a precise formal language that Lean can check.
It doesn't replace mathematical understanding.
But it can verify that formalized logical steps actually follow according to the rules of the system.
OpenAI says it released both a written proof and a Lean formalization of its proposed Navier–Stokes result.
That makes this episode particularly interesting.
AI isn't only generating text.
It is being used to generate mathematical reasoning that can then be translated into a machine-checkable form.
OpenAI says its system has resolved the Millennium Prize problem.
But the Clay Mathematics Institute has its own process.
Its rules say that before it considers awarding a Millennium Prize, a proposed solution must be published in a qualifying outlet, at least two years must pass, and the solution must receive general acceptance from the global mathematics community.
And this is exactly how mathematics is supposed to work.
A company announcing:
"We solved it."
is not the final step.
Other mathematicians need to inspect the argument.
They need to try to break it.
They need to reproduce the logic.
They need to look for hidden assumptions.
They need to understand whether the proof actually establishes the exact statement required by the problem.
Interestingly, the Clay Mathematics Institute itself published an announcement on September 11 saying the problem had "apparently been settled", while emphasizing that its evaluation process is deliberately unhurried.
So the story is moving very quickly.
But mathematics moves slowly for a reason.
Around the same time, mathematicians Tristan Buckmaster and Levent Alpöge, with Alpöge associated with Anthropic, were working on a related problem.
Their work concerned the forced Euler equations.
Remember Euler?
He gave us an important version of fluid equations before viscosity was incorporated into Navier–Stokes.
So we can think of the relationship roughly like this:
Euler equations → idealized fluid without viscosity
Navier–Stokes → fluid with viscosity
The difference sounds small.
Mathematically, it is huge.
Buckmaster and Alpöge produced a result showing a form of finite-time blow-up for a forced Euler problem.
OpenAI says its own Navier–Stokes result is different: its construction does not rely on external forcing in the same way.
Because the timing was unusual.
OpenAI says that on September 1 it heard rumors that two Millennium Prize problems had been resolved.
Those rumors turned out to be connected to Alpöge and Buckmaster's work.
OpenAI then decided to test its internal model against open Millennium Prize problems.
The Navier–Stokes effort eventually produced its claimed solution.
OpenAI says it later investigated whether Buckmaster's earlier Codex prompts could have influenced its internal system and concluded that they could not.
It also says the two proofs are significantly different.
This is an important area where we should be careful.
There are multiple claims being made by different researchers.
The safest way to understand the situation is not:
"AI stole the mathematicians' work."
Nor:
"AI independently solved everything with no connection whatsoever."
Instead:
There was concurrent research on closely related mathematical problems, unusual timing, and a dispute about possible influence. OpenAI says its investigation found no access to the relevant unpublished work and says the proofs differ significantly.
The mathematics itself needs to be examined.
Suppose the proof is eventually accepted.
Then something remarkable has happened.
A machine-assisted system has helped solve a problem that humans have struggled with for generations.
But that creates another question:
What does it mean to understand a proof that a machine discovered?
Imagine an AI gives mathematicians a proof containing hundreds of pages of complicated reasoning.
The proof is formally verified.
Every individual step checks out.
But perhaps only a small number of people can actually understand the big idea behind it.
Is that still mathematics?
I think this is going to become one of the most interesting questions of the AI era.
Because mathematics isn't only about getting the correct answer.
It is also about understanding why the answer is correct.
The Clay Mathematics Institute itself describes the value of proof in terms of not just certainty, but understanding.
For centuries, mathematicians did most of the exploration themselves. They calculated.
They experimented.
They wrote proofs.
They searched for patterns.
They tried again.
AI changes the economics of that process.
You can potentially have thousands of agents exploring different approaches simultaneously.
One agent searches for counterexamples.
Another tries a geometric argument.
Another works with inequalities.
Another tests numerical examples.
Another searches for connections with known theorems.
Another tries to formalize the proof.
Another tries to destroy the whole argument.
This last one is particularly important.
Maybe the future of mathematical AI isn't simply:
AI → produces proof → humans accept it.
Maybe it becomes:
AI → proposes proof → other AIs attack it → mathematicians inspect the surviving ideas → formal systems verify them.
That would look much more like a scientific research ecosystem.
There is a lesson here that has nothing specifically to do with fluid mechanics.
We are entering a world where machines can produce answers much faster than humans can verify them.
That is both incredibly useful and slightly uncomfortable.
Imagine an AI gives you a beautiful explanation.
It has equations.
It has references.
It has graphs.
It sounds confident.
It may even be completely wrong.
The Navier–Stokes story gives us a very good rule for the AI age:
A convincing answer is not the same thing as a verified answer.
And this applies far beyond mathematics.
If AI writes your code, test it.
If AI summarizes research, check the sources.
If AI gives you financial information, verify the numbers.
If AI gives you scientific conclusions, look at the evidence.
If AI gives you a mathematical proof, ask:
Can someone independently verify it?
There is something almost poetic about this problem.
For decades, mathematicians have been asking whether a mathematical system describing fluids can suddenly develop behavior that our equations cannot keep under control. Now AI is entering the picture and creating a similar question for us.
Can humans keep our understanding under control when machines become capable of producing mathematics that we cannot easily reproduce ourselves?
The answer doesn't have to be scary.
It could actually be exciting.
Maybe AI will become a new kind of mathematical microscope.
Something that lets us see structures that were always there but were too complicated for humans to discover.
Maybe it will help mathematicians solve problems that once seemed impossible.
Maybe it will also make mistakes that take years to uncover.
Probably both.
And that is why the most important part of the Navier–Stokes story may not be whether an AI has finally solved a Millennium Prize Problem.
It may be what happens after the AI gives us the answer.
Because the real test isn't:
"Can AI produce an answer?"
We already know it can.
The much harder question is:
"Can we understand, verify, and trust what it has produced?"
And perhaps that is the next great problem AI has given mathematics.
Not another equation.
Not another million-dollar prize.
But a question about how humans and machines discover knowledge together.