Navier–Stokes

When a swirl
meets its limit.

The Navier–Stokes equations describe fluid motion through a balance of acceleration, pressure, viscosity and external forces. The incompressibility condition says the fluid does not locally expand or compress. [1] ↗

The mathematical question is whether a smooth three-dimensional flow can lose that smoothness in finite time. A numerical image cannot settle it: a proof must control what happens at arbitrarily small scales. [1] ↗

OpenAI reports a vortex whose core narrows as it stretches, with unbounded speed but finite total energy under smooth external forcing. [2] ↗

How can the energy stay finite?

A peak value and a total are different things. A taller and taller spike can occupy a smaller and smaller region. As a simple mathematical analogy, increasing a quantity by a factor of n inside a volume of n⁻³ makes its squared-volume contribution scale as n² × n⁻³ = n⁻¹. This is an illustration of the distinction, not the research construction.

The official Clay formulation includes breakdown statements that allow smooth forcing. A forced counterexample does not, by itself, resolve the unforced problem. [1] ↗

Let the equations
move the particles.

The Animation mode tells the original three-stage story. Fluid simulation is a separate experiment: its velocity field evolves on a three-dimensional grid, and passive markers follow that field rather than a prescribed curve.

The solver uses a staggered grid, semi-Lagrangian advection and an iterative pressure projection. The pressure solve is approximate, so residual divergence remains. Tracers use midpoint integration. These are standard building blocks for interactive fluid simulation. GPU Gems ↗

A finite experiment

The domain is a periodic cube: flow leaving one side re-enters from the opposite side. This is a low-resolution numerical approximation with numerical diffusion, not a recreation of the reported research construction or evidence of a singularity. The three narrative stages are not imposed on this mode.

Choose 16³, 24³ or 32³ cells. Higher resolution costs more graphics work. If your device cannot keep up, simulation time advances more slowly rather than taking unstable catch-up steps. In VR, the same field is computed once and rendered for both eyes.

Colored points and trails are passive markers, not a water surface. Palette changes affect appearance only. Physics parameters use simulation units, not measured properties of a particular liquid.

Method reference: Real-Time Simulation and Rendering of 3D Fluids ↗

A research team.
At machine scale.

On September 8, 2026, OpenAI announced an analytical proof and Lean formalization for the forced Navier–Stokes problem. It says the work used an internal research model, not an ordinary ChatGPT session. [2] ↗

Its account describes parallel groups exploring different formulations, running code and sharing intermediate findings. An earlier unforced Euler result redirected the effort; humans selected problems and consolidated insights. OpenAI does not intend to claim the prize. These are company-reported figures and claims. [2] ↗

What does Lean add?

Lean is a programming language and proof assistant. A formal proof expresses the reasoning in a form its kernel can check against precise definitions and assumptions. This is different from a chatbot simply saying its answer is correct. [3] ↗

Our takeaway: the interesting model is not “ask once, trust the answer.” It is a research workflow in which exploration, criticism and checking are separate tasks.

The mathematics.
And the disagreement.

This is not simply “Anthropic versus ChatGPT.” Tristan Buckmaster describes his collaboration with Levent Alpöge, an Anthropic employee, as personal rather than institutional. They used both Claude and OpenAI tools. [4] ↗

Buckmaster’s account

He credits Diego Córdoba and Luis Martínez-Zoroa with the underlying research program. He reports forced Euler progress in August, questions the timing of OpenAI’s effort and raises concerns about unpublished work entered into Codex. He also alleges discussions about authorship that would have excluded Alpöge. He explicitly says he does not know whether their data was used. [4] ↗

OpenAI’s response

OpenAI denies accessing their unpublished work or specific user data, recognizes their priority on forced Euler, and says the proofs differ. It does not entirely exclude de-identified usage data influencing training. [2] ↗

Our question for the future: when a research platform also competes in research, how should it demonstrate that confidential work, priority and attribution are protected?

Beyond
the beautiful swirl.

The equations already matter in aircraft design, weather forecasting and blood-flow research. These are applications of fluid modeling, not products made possible overnight by a blowup proof. [2] ↗

Made by
OHZI.IO

We turn complex ideas into interactive experiences. OHZI.IO is an independent creative technology studio building immersive websites, real-time 3D and experiences for the spatial web.

From a first visual idea to a polished experience in the browser, we bring design and engineering together.