The shape
of a singularity.
Water flows. Air swirls. The equations that describe them can hide an extraordinary possibility: a smooth flow that stops being smooth.
A guide to Navier–Stokes, the newly announced proof, and the mathematics behind the headlines.
Follow the story About 15 minutes · No advanced math requiredas the tube narrows.z
↑
Can a smooth beginning
have a singular ending?
The million-dollar questionis about a guarantee.
and extraordinarily hard
to understand.
Picture a drop of ink falling into water. It stretches into filaments, curls into eddies, and disperses. Replace the ink with air flowing around a wing and you meet the same family of equations: Navier–Stokes.
They describe a fluid through a velocity field: an arrow at every point, telling you how fast the fluid moves and in which direction. Pressure and viscosity influence how those arrows change. The mathematics treats the fluid as continuous, rather than tracking individual molecules.
The prize question is whether a three-dimensional, incompressible flow can lose its smoothness in finite time under the specified assumptions. Here, smooth means the fields have derivatives of every order. Blowup means the required smooth description cannot continue through a particular time.
A single admissible counterexample can establish breakdown. It does not need to describe a common flow. And the prize never asked for a handy formula that calculates every possible swirl. [1]
What was known before this announcement?
Generalized, or “weak,” solutions have been known since Leray’s work in 1934. They satisfy the equations in an averaged mathematical sense. The difficult question is whether the required smoothness persists in three dimensions. The corresponding two-dimensional smoothness theory was already established. A weak solution and a smooth solution are different promises. [1]
Does the mathematical solution remain well behaved?
Can we forecast the flow accurately from imperfect information?
Newton’s law.
For every drop.
Select a term to read the equation
as a sentence.
Watch one point in the water.
How velocity changes here.
Stand on a bridge and watch the water directly below you. This term measures how its velocity changes over time at that fixed location.
Together with transport, it gives the acceleration of a moving fluid parcel.
Constant-density, incompressible form; f is force per unit mass. The terms interact in a single equation. [2]
Why the third
dimension matters.
A vortex can do something in 3D that a strictly planar flow cannot: stretch along its own axis.
Try stretching the vortex aboveVorticity measures local spinning motion. In a three-dimensional flow, stretching a vortex tube can make it thinner and intensify its rotation. The opening diagram illustrates that mechanism using a volume-preserving tube and an idealized inviscid picture.
Viscosity tends to smooth velocity differences. The deep difficulty is controlling the competition between nonlinear amplification and viscous smoothing, at every scale. Stretching alone is not a proof of a singularity. [3]
Does “high turbulence” mean “about to blow up”?
No. Reynolds number, Re = UL/ν, compares inertia and viscosity at a chosen speed U and size L. Large Re can make many scales important, but it is not a singularity detector. No universal Reynolds number marks the onset of turbulence for all geometries and disturbances. [4]
Is air really incompressible?
Incompressibility is a useful approximation in some flow regimes. Compressible aerodynamics allows changing density and couples momentum to mass and energy equations. Sound, heating, and high-speed flight require more physics than the particular model in the prize question. Euler is another related model: it removes viscosity. Proving something for Euler does not automatically prove it for Navier–Stokes. [5]
One headline.
Two different results.
The distinction that makesthe news make sense.
Two research efforts sit behind the headlines: Buckmaster and Alpöge’s collaboration, and OpenAI’s separate announcement about the prize problem.
| What to compare | Buckmaster & Alpöge | OpenAI’s announcement |
|---|---|---|
| The people | Tristan Buckmaster, an NYU mathematician, and Levent Alpöge, an Anthropic researcher; a personal collaboration. | An OpenAI research effort, with humans coordinating a large group of AI agents. |
| The central result here | Blowup for smoothly forced Euler, alongside related equations. | Blowup for smoothly forced Navier–Stokes, with positive viscosity. |
| Which AI? | Claude and Codex, especially GPT‑5.6 Sol; Astra later helped with exposition and audits. | An unreleased internal model for discovery; GPT‑6 Astra for subsequent Lean work. |
| Connection to the prize | Related foundational progress. Euler omits the viscous term. | Claims the permitted breakdown routes C and D in Clay’s formulation. |
Attributions follow the participants’ own accounts and published artifacts. [6] [7] [8]
Start at rest.
Apply a smooth force.
Lose smoothness in finite time.
For every positive viscosity, the paper constructs a 3D incompressible flow with zero initial velocity and smooth forcing confined in space and time. The velocity stays smooth before the singular time; its maximum becomes unbounded, while its total energy remains bounded.
The construction works in all of space and transfers to a periodic domain—a box with opposite faces identified. It does not settle the case with no external force. [9]
How can a forced result qualify for the Millennium Prize?
The official statement offers four acceptable routes. A and B ask for global smoothness with no forcing. C and D allow a counterexample with suitable smooth initial data and smooth forcing. A proof of C or D can therefore qualify without resolving unforced regularity. Injecting an infinite force would fail those requirements; keeping the force smooth is essential. [1]
The architecture of the announced proof
Explore the logic. These are explanatory sketches, not the full argument.
Concentrate motion into a shrinking core.
The leading vortex narrows as the singular time approaches. Both its radius and axial length shrink, with the radius shrinking faster. “More elongated” means relatively slender, not literally growing longer.
The opening stretch experiment teaches a general mechanism. This construction has different geometry.
A reading of the paper’s overview, not a certification of its 166-page argument. [9]
Infinite speed.
Finite energy.
A maximum describes the fastest point. Energy adds contributions from the whole volume. An increasingly intense motion can occupy an increasingly tiny region.
Move closer to the limiting time to see these quantities go in opposite directions.
Normalized leading power laws from the paper, with h = 0.005 and τ = 1 − t. Illustrative ratios, not measured values or a numerical solution. The slider never reaches τ = 0.
Show the mathematics behind the chart
Radius scales as τ1/2, axial length as τ1/2−h, and speed as τ−1/2−h. Thus core energy scales like speed² × radius² × axial length = τ1/2−3h. For 0 < h < 0.01, that last power is positive: core energy decreases even as speed grows. The whole flow has bounded energy. [9]
A limit of the model.
A beginning for new ideas.
What this could changeoutside mathematics.
is not a glass of water
exploding.
Real fluids consist of molecules. Infinite speed belongs to a mathematical construction in a continuum model. Whether a related mechanism can occur in a physical setting requires checking the force, scales, geometry, and omitted physics.
If the proof withstands scrutiny, the firm conclusion is a failure of an unconditional smoothness guarantee in its allowed setting. Its value to engineering would come through further work: new analytical techniques, sharper estimates, and better understanding of difficult flows.
The consequences below are interpretations of the theorem’s scope, not announced engineering results.
Better understanding.
Not an instant better wing.
AERODYNAMICSWhat could follow
New mathematical techniques could inform analysis of intense, small-scale flow and provide demanding test cases for future methods.
What still needs work
Predicting stall and separated turbulent flow still depends on geometry, grids, turbulence models, and validation against experiments. This proof does not supply a faster CFD solver or certify an aircraft.
Read NASA on turbulence and aircraft certification ↗Would proving smoothness instead have made weather perfectly predictable?
No. A smooth, deterministic system can still amplify tiny errors in its initial conditions. Weather forecasts begin with incomplete observations and imperfect models. Ensembles help quantify the resulting uncertainty. Mathematical existence, numerical accuracy, and forecast predictability are separate questions. [10]
A new kind of
mathematical collaborator.
Discovery. Verification.Understanding.
The strongest signal is the ability to produce a checkable research artifact.
OpenAI reports roughly 88 hours of discovery work, followed by 17 hours of Lean work, using about 10,000 concurrent agents in its successful group. These are company-reported figures, not an independent efficiency benchmark. [7]
Buckmaster describes an iterative collaboration using Claude and Codex, with people providing ideas and working to understand the resulting arguments. Their Euler repository attributes its Lean code to Claude under Alpöge’s direction. Earlier ideas from Córdoba and Martínez-Zoroa matter to that research lineage. [6] [8] [11]
Find an argument that could establish something new.
Translate the claim and its reasoning into precise statements a proof assistant can check.
Check the code, assumptions, definitions, and correspondence to the intended theorem.
Explain the mechanism, simplify the reasoning, and discover what it lets us do next.
Is this mathematical
superintelligence?
If validated, solving a frontier research problem is compelling evidence of advanced mathematical capability. But one achievement does not establish reliable superiority across all mathematics, let alone every scientific or human task.
Formal verification is valuable precisely because a model’s confidence is insufficient. A successful Lean check certifies an encoded statement under its definitions and axioms. People must still assess whether that statement captures the intended problem and what the result means. [12]
What about the controversy over credit?
Buckmaster disputes aspects of OpenAI’s presentation and describes disagreements about timing and authorship. OpenAI offers a different account and says its investigation found his Codex interactions did not influence its system. Those accounts do not establish that research was stolen. Read both primary statements; the dispute should not be confused with the validity of the mathematical arguments. [6] [7]
What “solved”
needs to mean.
As of 11 September 2026
OpenAI has released an analytical writeup and a Lean repository. That is substantially more concrete than an unsupported chatbot answer. [9] [12]
This explainer examined the theorem, overview, public statements, and project documentation. It did not check all 166 pages or rebuild the Lean projects.
Clay’s page still lists the problem as active. Its rules require publication in a qualifying outlet, at least two years since that publication, and general acceptance before consideration. OpenAI says it will not seek the prize. [13] [7]
The careful headline: OpenAI has announced a formally supported forced Navier–Stokes blowup proof. Its precise scope matters, and public mathematical scrutiny matters.
Follow the evidence.
Primary sources, so you can go deeper.
News claims are dated; the equations are not.
- The official Millennium problemCharles Fefferman · Clay Mathematics Institute · PDF
The exact hypotheses and acceptable statements A–D.
- Where the equations come fromTerence Tao · Lecture notes · 2018
The continuum approximation and incompressible fluid equations.
- Viscosity and vorticityDavid Tong · Cambridge lecture notes · PDF
Vortex stretching, viscous effects, and the Burgers vortex.
- Reynolds numberNASA Glenn · Educational reference
How inertia and viscosity compare.
- The aerodynamic equationsNASA Glenn · Beginner’s guide
Why practical fluid models include additional equations and physics.
- Buckmaster’s accountTristan Buckmaster · September 2026 · PDF
Collaboration, model use, chronology, and the attribution dispute.
- OpenAI’s announcementOpenAI · 8 September 2026; updated 10 September
The company’s result, reported process, and response to the controversy.
- The forced Euler formalizationAlpöge–Buckmaster · Public repository
The theorem statement, formal proof, and AI contribution statement.
- The Navier–Stokes paperOpenAI · Theorem 1.1 and §2 · PDF
The claimed result and the leading scales behind the interactive chart.
- Why forecasts remain uncertainECMWF · Research overview
Initial conditions, model error, and ensemble prediction.
- An expert’s reading of the earlier breakthroughTerence Tao · 7 September 2026
The Córdoba–Martínez-Zoroa strategy and Alpöge–Buckmaster extensions.
- OpenAI’s Lean proof artifactsOpenAI · Public repository
Formal statements, code, and instructions for reproducible checking.
- How a Millennium Prize is recognizedClay Mathematics Institute · Official rules
Publication, the two-year minimum, and community acceptance. Current problem status.
- Turbulence and aircraft certificationNASA · Research overview
The practical challenges of high-lift flow simulation.
- Ocean and ice processesNOAA GFDL · Research overview
Mixing, unresolved processes, and coupled climate models.
Reading guide: begin with Clay’s statement for the question, the paper’s opening pages for the claimed answer, and the repositories for the checkable artifacts. Interpretive passages on engineering and AI are explicitly separated from reported findings.