Headlines announced the solution of a million-dollar problem, one of the Clay Mathematics Institute’s Millennium Prize Problems.

The results tell a more precise story. Regularity for the Navier–Stokes equations remains open in the unforced case. Several groups have shown that a singularity can form under tightly controlled conditions. OpenAI has also published a separate theorem on blowup for unforced Euler; an independent group has found a numerical candidate for a stable singularity in the same regime.

These results give us a better map of what is known.

The equations#

The Euler equations describe an ideal fluid with no viscosity. The Navier–Stokes equations add viscosity, which dissipates energy and makes the model closer to a real fluid.

Both equations describe the evolution of a velocity field u. Analysts ask whether a smooth initial field keeps all its derivatives smooth or develops a singularity in finite time. In that case, a velocity, a vorticity, or one of its derivatives becomes unbounded.

A singularity does not describe a particle reaching infinite speed. It describes the continuum field used by the equation losing regularity at one point and one instant: up to that moment the solution stays smooth, then one of its norms or one of its derivatives diverges. The term blowup names exactly this unbounded growth. In a real fluid, a molecular description would have to take over before that mathematical limit.

The Fefferman formulation of the Clay problem distinguishes four cases:

Case Domain External forcing
A three-dimensional space absent
B periodic domain absent
C three-dimensional space present and smooth
D periodic domain present and smooth

Proving any one of the four alternatives satisfies the mathematical criteria for the prize. Cases C and D allow researchers to construct a singular solution and define the forcing as the residual left in the equation. In cases A and B, the forcing must be zero.

OpenAI’s result#

OpenAI’s paper Finite Time Blowup for Navier–Stokes constructs, for every positive viscosity, a smooth compactly supported force, zero initial data, and a smooth solution up to t = 1.

As that time approaches, the velocity diverges in the L∞ norm. The L² energy remains bounded and the singularity concentrates at the origin. Compact support also produces the periodic version of the result.

The paper proves alternatives C and D in Fefferman’s formulation. The external force is part of the construction. The work does not show that unforced Navier–Stokes develops a singularity, and it does not show that unforced solutions remain regular. A and B remain open.

The sentence “OpenAI solved Navier–Stokes” therefore needs a qualifier. OpenAI proved the forced case described by C and D. The announcement used “Navier–Stokes” without putting that distinction first.

The proof is public and formalized in Lean, but it has not yet received independent mathematical review. Formal verification certifies that the statement follows from its assumptions; acceptance by the community is a separate step.

OpenAI also published Finite Time Blowup for the Euler Equation, which proves blowup for unforced three-dimensional Euler with smooth compactly supported initial data. This is a separate result from the Navier–Stokes paper and does not solve the Clay problem.

The amplification cascade#

The OpenAI result and the Alpöge–Buckmaster result follow a strategy developed in the mathematical program of Diego Córdoba and Luis Martínez-Zoroa.

Researchers start with an approximate solution and add a high-frequency correction. The correction must grow enough to destroy the solution’s regularity. Its residual must remain small enough to fit inside a smooth forcing term.

Each layer is transported by the previous flow and prepares the next one. The frequency grows at every step. The solution accumulates the amplification; the forcing accumulates smaller corrections. In the limit, the solution develops a singularity while the forcing remains smooth.

The hard part is controlling every derivative of the residual. A simulation can show a very steep profile. This construction must prove that the sequence of profiles converges in the way required by the equation.

Alpöge and Buckmaster’s result#

The paper Blowup for the Euler Equations with Smooth Forcing studies the three-dimensional incompressible Euler equations.

The authors construct a solution with smooth initial data and a force that is smooth in space and time, supported inside a solid torus. In finite time, the gradient of the circulation, the vorticity, and the time integral of the vorticity norm diverge. The circulation and the meridional velocity components remain bounded.

The result concerns forced Euler. It does not concern unforced Euler, and it does not transfer automatically to Navier–Stokes, where the viscous term is present.

The work closes a technical step in the Córdoba and Martínez-Zoroa program: a sequence of approximations can produce a singularity while keeping the total forcing smooth.

The independent PINN candidate#

The independent group of Adarsh Ganeshram, Valentin Duruisseaux, and Anima Anandkumar is looking for a stable singularity in the three-dimensional Euler equations on R³, without a boundary and without forcing. Their manuscript, Stable Singularity of the Euler Equations on R3 without forcing, describes the candidate. Their work follows a different route from OpenAI’s unforced Euler theorem.

A physics-informed neural network, or PINN, receives the differential equation and searches for a function that satisfies it. The group used a PINN to find a self-similar profile. They then converted the profile into splines and used interval arithmetic to control residuals and derivatives.

The reported numerical residual reaches about 10^-10. The candidate shows the required damping behavior across most of the domain and a strategy for the off-axis region, where the estimates become harder.

The manuscript separates established results from assumptions that remain to be closed. The authors formalized the reduction of the equations, the linearization, the modulation, and conditional stability theorems. They have not yet proved that an exact profile exists near the spline or that all constants have the positive margin required.

The work therefore provides strong evidence and a computer-assisted proof framework. Unconditional blowup for unforced Euler remains unproved.

Four consequences#

Smooth forcing can accompany blowup#

The OpenAI and Alpöge–Buckmaster results construct singularities with forcing that is smooth in space and time. The force does not hide the problem behind an irregularity.

This answers the mathematical question in cases C and D. It does not show that a real fluid spontaneously produces the same behavior.

Unforced Euler is the closest frontier#

OpenAI proved blowup for the unforced three-dimensional Euler equations with smooth compactly supported initial data. The PINN candidate searches for a stable solution in the same setting, but has not yet reached an unconditional theorem.

The move from Euler to Navier–Stokes remains open. Viscosity changes the balance at small scales; a mechanism that works for Euler does not transfer automatically.

The equations describe a limit of the continuum model#

Fluid equations work because, at macroscopic scales, we can ignore molecular details. Near a singularity, that approximation may stop applying.

Effective field theory for fluids offers a complementary view. The symmetries of the system allow researchers to derive the Euler and Navier–Stokes equations after removing microscopic details that do not matter at large scales. A mathematical singularity can therefore mark the limit of the continuum model without implying infinite velocity in a real fluid.

A real fluid consists of molecules. Before reaching the infinity of the equations, the model would have to give way to a microscopic description.

AI accelerates research, verification decides#

The three projects use different forms of AI:

  • OpenAI coordinates thousands of agents that explore and assemble proofs;
  • Alpöge and Buckmaster use LLMs to accelerate an analytical strategy already present in the literature;
  • Ganeshram, Duruisseaux, and Anandkumar use a non-generative network to search for a solution profile.

Verification remains the decisive step. Lean certifies that an assertion follows from its assumptions. It does not certify that the assertion represents the question the community meant to ask, or replace understanding of the result.

In the Navier–Stokes case, mathematical correctness and mathematical relevance follow different paths. Formal systems can certify the first. The second requires reading, comparison with the literature, and human interpretation.

The open question#

Researchers have constructed singular configurations with rigorous control even when the forcing is smooth. OpenAI proved blowup for unforced three-dimensional Euler; an independent group is searching for a stable singularity profile with a PINN, but has not yet closed the proof. The results show how an iterative amplification strategy can turn successive approximations into formalized theorems.

The question remains open for Navier–Stokes:

Can a smooth three-dimensional solution, starting from smooth initial data with no forcing, develop a singularity in finite time?

Research has clarified the route toward an answer, not the answer itself.

Sources#