The most familiar space has three dimensions: up and down, in and out, left and right. Three coordinates are required to mark a point. The thought that nothing could be simpler is both natural and wrong.
In 1961 Stephen Smale proved the generalized Poincaré conjecture for five and higher dimensions. The original three-dimensional conjecture remained unresolved for another forty years. When Grigori Perelman finally established the conjecture in 2002–2003, he required mathematical techniques not available to Smale—chiefly Richard Hamilton’s work on Ricci flow.
The hierarchy of difficulty had run backward.
Something similar is true of the Navier–Stokes equations. Known since the nineteenth century, they describe the flow of fluids. In two dimensions, their behavior is comparatively well understood. In three, an additional possibility appears: vortices may stretch as well as turn. Stretching intensifies rotation; viscosity opposes it. Given smooth initial conditions, can the competition between rotation and viscosity drive the velocity field to a singularity—a place where solutions cease to remain smooth—in finite time? If so, the vortices have won. If not, viscosity has prevailed. The question is at the heart of the problem posed by the Clay Mathematics Institute in 2000—one of seven Millennium Prize Problems.
On September 8, 2026, OpenAI announced that one of its internal mathematical models had found an answer. There exists a smooth external force for the three-dimensional incompressible Navier–Stokes equations such that, starting from a fluid at rest, the resulting solution remains smooth for a finite time and then becomes singular. The velocity becomes unbounded while the total kinetic energy remains finite.
The Navier–Stokes equations begin with two quantities that are familiar enough: velocity and pressure. At every point x and at every time t, let u(x,t) denote the velocity of the fluid, and let p(x,t) denote its pressure. Velocity is a vector: it tells us both how fast a small portion of fluid is moving and in which direction. Pressure, by contrast, is a scalar: it has a value but no direction. The equations describe how the velocity field changes under the influence of pressure, viscosity, and an external force.
For an incompressible fluid of constant density, the Navier–Stokes equation may be written as
1 ∂u∂t + (u⋅∇)u = -∇p + νΔu + f.
Equation 1 is nothing more than Newton’s second law of motion applied to a fluid. The expression on the left describes acceleration; the expressions on the right, the forces that produce it.
The first term on the left, ∂u/∂t, records the change in velocity at a fixed point. But a fluid does not remain at a fixed point. Witness the nonlinear second term, (u⋅∇)u, which records the change in velocity as the fluid moves from place to place. On the other side of the equation, -∇p represents the acceleration caused by differences in pressure; νΔu, the effect of viscosity. Here ν > 0 measures how viscous the fluid is, while the Laplacian Δ compares the velocity at a point with nearby velocities. Its effect is smoothing: sharp differences in velocity tend to be rubbed away. Finally, f represents an externally applied force.
One further equation is required to represent incompressibility:
2 ∇⋅u = 0.
A fluid may enter and leave a small region, but it cannot accumulate there. Whatever flows in must flow out.
The mystery of the Navier–Stokes equation lies in the contest already visible in these symbols. The nonlinear term can transfer motion to smaller and smaller scales. Viscosity works in the opposite direction, dissipating motion at those scales. In two dimensions viscosity prevails. In three dimensions, the geometry permits vortices to stretch, and stretching can intensify vortices.
The question left open for nearly a century was whether this process could outrun viscosity altogether.
There is one important qualification. Mathematicians have long known that the Navier–Stokes equations possess solutions in a weaker sense. Jean Leray proved this in 1934. What remained unknown was whether perfectly smooth initial data must always produce a solution that remains smooth for all time. The Millennium problem was not whether the equations have solutions. It was whether smoothness can break down.
Vorticity
Imagine the velocity field as a collection of small arrows, one attached to every point of the fluid. Nearby arrows need not point in exactly the same direction. Differences among nearby arrows can reveal a local tendency to rotate. The mathematical operation that extracts this rotation from the velocity field is the curl, and the resulting quantity is the vorticity:
3 ω = ∇×u.
Like velocity, vorticity in three dimensions is a vector. Its direction gives the axis about which the fluid is locally rotating, and its magnitude measures the strength of that rotation.
Taking the curl of the Navier–Stokes equation produces an equation for vorticity:
4 ∂ω∂t + (u⋅∇)ω = (ω⋅∇)u + νΔω + ∇×f.
Pressure has disappeared. It entered the original equation as a gradient, and the curl of a gradient is zero.
The two terms on the left-hand side of 4 describe changes in vorticity as it is carried along by the fluid. On the right-hand side, viscosity again has a smoothing effect, and the last term records any vorticity introduced by an external force. But there is now a new term,
5 (ω⋅∇)u,
and it is this term that distinguishes three dimensions from two.
Vortex Stretching
Imagine a thin tube formed by neighboring vortex lines. In an incompressible fluid, if such a tube is stretched lengthwise while preserving its volume, it must become narrower. The stretching term in Equation 4 shows that this deformation can also intensify the vorticity within the tube. A stronger vortex may then be stretched still further by the surrounding flow. A fluid in motion can increase the very rotation that determines its subsequent behavior.
In two dimensions this mechanism is absent. Vorticity vectors are perpendicular to the plane in which the fluid moves, while velocity varies only within that plane. The stretching term vanishes. Vorticity can be transported from one place to another, and viscosity can diffuse it, but the flow cannot amplify vorticity by stretching a vortex out of the plane.
This is the reason that the two-dimensional equations are better behaved.
In three dimensions, a vortex may turn, stretch, narrow, and intensify. Viscosity is still trying to smooth it away. The unresolved question is whether stretching can concentrate vorticity ever more strongly on ever smaller scales, until smoothness finally fails, or whether viscosity must always intervene before that can happen.
To see exactly what would count as settling this question, we must turn to the formal statement of the Millennium problem.
Four Alternatives
Charles Fefferman, the Princeton mathematician and Fields Medalist who formulated the Millennium problem for the Clay Mathematics Institute, gave mathematicians four possible ways to settle it. Two concerned flow throughout three-dimensional space; two concerned periodic flow, in which the pattern of motion repeats indefinitely in every spatial direction. In each setting, an answer could take either of two forms: proving that smooth solutions remain smooth for all time, or constructing a solution whose smoothness breaks down. The four possibilities can be stated as follows:
Statement A: In three-dimensional space, smooth initial data with no external
force always produce a smooth solution for all time.
Statement B: The same
conclusion holds for periodic flow.
Statement C: In three-dimensional space,
there exist smooth initial data and a smooth external force for which a smooth solution breaks
down in finite time.
Statement D: The same conclusion holds in the periodic
case.
Statements A and B establish global regularity: viscosity ultimately prevents singularity. Statements C and D establish finite-time breakdown. The OpenAI construction takes the route provided by Statement C.
In statements A and B, the external force is zero.
Not so in Statements C and D. There the external force is crucial. At first sight, allowing an external force may seem to make the problem too easy. Push a fluid hard enough, and why should it be surprising that something goes wrong?
But that external force is subject to essentially the same constraints as the initial data.
It must be smooth in space and time, and it must decay rapidly at large distances. It cannot become singular.
Singular behavior must arise from the dynamics of the Navier–Stokes equation itself.
The Construction
In OpenAI’s construction, a fluid begins at rest:
6 u(x,0) = 0.
There is no initial vortex waiting to explode. The fluid’s motion is produced by a smooth external force, and the resulting velocity remains smooth for a finite interval of time. During that interval, a vortex develops and becomes increasingly concentrated. At time t = 1, the velocity becomes unbounded, even as the total kinetic energy remains finite.
This may seem paradoxical. For a velocity field u, its kinetic energy, apart from an inessential constant depending on density, is measured by
7 ∫ℝ3 |u(x,t)|2 dx.
A velocity field can become unbounded even if its total energy remains finite. The system’s energy depends not only on how large the velocity becomes but also on the volume over which those larger velocities occur. A function may grow without bound on a region whose volume is shrinking rapidly enough for the integral of its square to remain finite. As t approaches 1, the region in which the motion is becoming extreme contracts; at the same time, the velocity within it increases. There is no contradiction between finite energy and unbounded velocity, and no paradox either. The singularity is not produced by putting an infinite amount of energy into the fluid. It is produced by concentrating the fluid’s motion into an ever smaller region.
The example constructed by OpenAI describes a flow in which a vortex spirals inward while being stretched along its axis. As the central region becomes narrower, the fluid moves faster.
The geometry resembles the mechanism suggested by the vorticity equation, but a suggestion is not yet a proof. Many plausible scenarios have failed because viscosity, pressure, or some other term in the Navier–Stokes equation becomes large enough to destroy the mechanism before the singularity can form.
The decisive difficulty is not simply to invent a velocity field that blows up. It is to invent one that blows up while still satisfying the Navier–Stokes equations with an external force that remains perfectly smooth.
These constraints turn the Navier–Stokes equation into a balancing problem. Near the developing singularity, several terms in the equation become very large: acceleration, the pressure gradient, the nonlinear transport term, and the viscous term. Taken separately, they appear to demand an unbounded external force. The construction is designed so that their largest contributions cancel. What remains after cancellation is smooth.
The singularity is confined to the solution, not smuggled into the equation that produces it.
Oscillatory Disturbances
Those cancellations do not arise from the collapsing vortex alone. Left to itself, the vortex comes close to doing what is required, but not close enough. The central flow can be arranged so that its leading terms satisfy the Navier–Stokes equation, yet the balance still fails in an annular (ring-like) region where that flow joins the surrounding fluid. The external force required to correct this failure would itself become unbounded as t approaches 1.
The OpenAI construction makes use of an old idea in mathematics and physics. Opposing effects may cancel when added; their interactions need not cancel at all. A collection of small motions can therefore have almost no average motion of its own and still produce a substantial net effect. Disturbances taken one by one have essentially no average motion around the vortex. Their quadratic effects, however, need not average to zero. A disturbance carrying slightly more angular momentum outward and one carrying slightly less angular momentum inward can together produce a net outward transport of angular momentum. Similar disturbances can transport momentum along the axis. The amplitudes, orientations, and locations of the oscillations are specified as part of the construction so as to supply the momentum flux missing from the background vortex.
The external force does not perform the singular balancing act. The nonlinear motion of the fluid does.
To see how this is possible, it helps to look more closely at the scale of the collapsing vortex. Let
8 τ = 1 - t.
Thus τ measures the time remaining before the singularity. As t approaches 1, τ approaches zero.
The central vortex does not contract at the same rate in every direction. Its radial width is of order
9 τ1/2,
while its length along the axis is of order
10 τ1/2 - h,
where h is a small positive number. Both dimensions tend to zero, but the radial width contracts faster. The region containing the most violent motion becomes a progressively thinner column. Speeds within that column increase. The azimuthal velocity—the velocity around the axis—and the axial velocity grow roughly as
11 τ-1/2 - h;
and become unbounded as τ tends to zero. The radial velocity grows more slowly, on the scale of τ-1/2. The geometry and the velocity are thus coupled: the region contracts, the vortex spins and stretches more rapidly, and the characteristic speeds diverge. These estimates also show how the finite-energy condition can be satisfied: the total kinetic energy must remain finite even as the velocity becomes unbounded.
The volume of the central region is of order
12 τ3/2 - h.
The square of its largest characteristic velocity is of order τ-1 - 2h. Multiplying the two gives a contribution to the kinetic energy of order
13 τ1/2 - 3h.
For h < 1/6, as required by the construction, this quantity tends to zero. The speed inside the collapsing region is becoming arbitrarily large; the amount of kinetic energy concentrated there is becoming arbitrarily small.
This disparity of scales is not incidental to the proof. It is what allows the vortex to approach a singularity without violating the finite-energy requirement, and it creates increasingly intense shear in the annular region surrounding the core. That shear, in turn, supplies the mechanism by which the oscillatory disturbances can do their work.
Shear Amplification
Shear surrounding the core does more than provide a setting for oscillatory disturbances. It supplies their energy. Each disturbance is introduced with a small amplitude by an external force and carried by a background flow whose velocity changes sharply with position. The disturbance can then extract energy from this shear and grow. The external force supplies the seed; the fluid supplies the amplification.
Rotation around the axis provides the first example. A small fluid element moving outward may carry slightly more angular momentum than the surrounding fluid at its radius. Its angular momentum then differs from that of the surrounding flow, altering its motion in a way that can push it farther outward. Suppose that the angular velocity of the background vortex decreases sufficiently rapidly with the radius. The outward displacement then increases the element’s excess angular momentum relative to its surroundings. This strengthens the displacement still further. An inward-moving parcel with a corresponding deficit behaves in the opposite direction. The result is a feedback mechanism. Given the constraints imposed by the OpenAI construction, the amplification becomes exponential.
Viscosity is acting at the same time. It damps rapid spatial oscillations, and eventually it wins. But before it does, there is an interval during which shear amplifies the disturbance faster than viscosity suppresses it. As the background shear deforms an oscillatory disturbance, its radial wavelength becomes shorter. As the wavelength shortens still further, viscous damping becomes stronger, and the disturbance dies away. The initial wavelength is chosen so that these events occur in the required order: first amplification, then damping.
This amplification is precisely what the construction requires. A disturbance may begin at very small amplitude, grow large enough for its nonlinear interactions to transport the required angular momentum, and then disappear rapidly enough so that the external force used to create it remains smooth. A second family of disturbances performs the analogous task for momentum along the axis. With their relative strengths appropriately adjusted, the two families supply the components of momentum transport that the background vortex by itself lacks.
No single disturbance need survive until t = 1. Instead, a succession of them is introduced on progressively smaller spatial and temporal scales. Each performs its part of the correction and is damped away; later disturbances operate closer to the contracting core and closer to t = 1. The sequence tracks the collapse inward. What would otherwise be a singular residual in the Navier–Stokes equation is absorbed, scale after scale, by the nonlinear motion of the fluid itself.
What Counts as an Answer
A mathematical problem is what it is. A mathematical theorem says what it is. The problem formulated by Charles Fefferman did not ask only whether smooth solutions of the unforced Navier–Stokes equations remain smooth for all time. It also allowed the opposite conclusion to be established. The distinction was part of the problem from the beginning.
There are good reasons for regarding the unforced problem as deeper, more natural, or more closely connected with the behavior of physical fluids. None of them alters the statement of the Millennium problem. Once a mathematical question has been formulated precisely, judgments about which of its alternatives is more interesting cannot change what counts as an answer.
This is one reason the solution is so striking. The OpenAI model did not reformulate the problem, weaken its hypotheses, or introduce an exception not contemplated in the original statement. It followed the statement itself.
For twenty-six years the problem had been surrounded by judgments about which outcome was likely, which formulation was natural, and which part of the question mattered most. The model was bound by none of these judgments. It confronted instead a mathematical statement, with several alternatives explicitly left open.
One of those alternatives led somewhere.
