A needle of water: how 10,000 agents made the Navier–Stokes equations blow up ✛ ✎ ✦ OpenAI theoretical made by OpenAI AI agents (≈10,000 agents, 88 h) · checked in Lean
In 1934 Jean Leray proved that the Navier–Stokes equations always have solutions "in a generalized sense". Whether a solution stays smooth or can break down into a singularity he could not tell, and nobody could for ninety years. In 2000 the Clay Institute made the question one of its seven Millennium Prize problems. The statement written by Charles Fefferman is honest: he asks for a proof of any one of four claims. (A) and (B): without an external force smoothness always holds, in space and on the torus. (C) and (D): there exist a smooth force and smooth initial data for which no smooth finite-energy solution exists.
OpenAI's paper proves (C) and (D). For every viscosity \(\nu\gt 0\) it constructs a smooth force, compactly supported in space and time, and a flow that starts from complete rest, keeps its energy bounded and reaches infinite speed at \(t=1\). How this is done is worth understanding, because the method turns the problem inside out.
The problem is solved backwards. The Navier–Stokes equation can be read as a definition of the force. Write down any velocity u and pressure p, substitute them, and declare whatever fails to balance to be the external force f. The equation holds identically. A blowing-up flow is easy to invent. The difficulty moves to one place: the residual must be smooth. Individual terms in it diverge, and one has to arrange cancellations so that the sum and all its derivatives pass calmly through the blowup time. That is what the 166 pages are spent on.
The needle vortex. A vortex on the vertical axis collapses toward the origin. Inside, fluid spirals toward the axis and flows out along it, up and down. Conservation of angular momentum drives the spin-up: for a parcel with no torque \(r\,u_\theta\) is constant, so smaller r means larger \(u_\theta\). Incompressibility keeps the inflow going: the incoming fluid is carried away by the axial flow, otherwise it would clog the axis.
The core contracts self-similarly: the profile keeps its shape and only the scales change. Let \(\tau=1-t\) be the time left to blowup. The core radius is \(\sim\tau^{1/2}\), the height \(\sim\tau^{1/2-h}\) with a tiny \(h\lt 1/100\), the speeds \(\sim\tau^{-1/2-h}\). The radius shrinks faster than the height, so the core stretches into a thin needle of volume \(\sim\tau^{3/2-h}\). From here comes the most elegant point of the paper: the core's kinetic energy \(\sim\tau^{1/2-3h}\) tends to zero. The speed is infinite, the energy vanishes, the volume falls faster than the square of the speed grows.
Viscosity is not defeated but built in. Viscosity is what separates Navier–Stokes from Euler, and it is why the equation was thought perhaps to be safe. The paper computes two Reynolds numbers. In rotation \(\mathrm{Re}_\theta\sim\tau^{-h}\to\infty\): the fluid makes ever more turns within a viscous diffusion time, rotation wins. In the radial inflow \(\mathrm{Re}_r=O(1)\): viscosity remains an equal partner. The rates of all processes, axial and radial transport and diffusion across the core, grow alike as \(\tau^{-1}\) and stay in the leading balance to the end. Along the axis diffusion is weaker by a factor \(\tau^{2h}\), and the construction uses that gap.
Where it breaks. The core has to be stitched to calm fluid outside. In the annulus between core and exterior the stitching does not balance, and the residual grows without bound as \(t\to1\). The force must be smooth, yet here it would have to be infinite. The background vortex does not sustain itself.
The fluid produces the force itself. Into the annulus they plant small oscillating disturbances, each a complete ring around the axis, localized in radius, height and time and started by an exponentially small seed. How does a zero-mean oscillation yield a nonzero force? Moving outward, fluid carries a surplus of azimuthal velocity; moving inward, a deficit. The mean velocity is zero, but the flux of angular momentum, the product of the two, keeps its sign in both half-cycles. Flip both signs and the product is unchanged. The result is a steady mean momentum flux whose spatial variation acts on the background like a genuine force. Physicists call this Reynolds stress and usually treat it as a nuisance; here it is a tool, tuned to cancel exactly the divergent part of the residual.
The pulses flare and fade on their own. They grow on the background shear by a centrifugal mechanism: a parcel with surplus rotation is pushed outward, and if angular velocity falls steeply enough with radius, its surplus relative to its neighbours grows. The two effects feed each other, giving exponential growth while it beats viscous damping. They are killed by shear as well: it carries different radii at different speeds and shortens the radial wavelength, and viscous damping grows with the square of the wavenumber. The initial wavelength is chosen so that amplification wins first and damping inevitably later. Each pulse leaves a tail so small that all its derivatives vanish at the blowup time.
The rest of the assembly. There are two families of pulses: both the flux of angular momentum and the flux of axial momentum must be supplied, and the families differ in the ratio of the two. Toward the blowup the pulses are laid on ever finer grids in space and time, as in convex integration: an infinite ladder of corrections, each finer than the last. There is a deliberate asymmetry: with exact up-down symmetry the amplification mechanism vanishes in the plane \(z=0\), so the axial profile is slightly tilted. Outside the annulus the flow is purely azimuthal, independent of height, and solves the radial heat equation exactly: its residual is zero and no force is needed there. This allows a smooth spatial cutoff, and compactness carries the result to the torus, hence (D).
What is new compared with predecessors. The forcing-and-amplification-across-scales strategy belongs to Córdoba and Martínez-Zoroa: first for 3D Euler, then with Zheng for Navier–Stokes with weakened dissipation. Tao in 2016 built blowup for an averaged Navier–Stokes that keeps the energy cancellation. Two steps are new here: true viscosity and the full equation for every \(\nu\gt 0\), and pulses that do not merely amplify but act as a source of the required stress, fitted exactly under the hole in the background. The bricks are all known: centrifugal instability, wavevector transport by shear, exact viscous shearing waves, the realization of stress by oscillations from the Euler constructions of Daneri and Székelyhidi. Only the assembly is new.
Who did it, and how. By OpenAI's account, the proof was found by a group of about ten thousand coordinating agents on an internal model, in 88 hours, sending 2.7 million messages and spending about 130 billion output tokens; formalization and checking in Lean took another 17 hours. No human signs the paper: the author is listed as OpenAI. A day earlier Tristan Buckmaster and Levent Alpöge posted three preprints using the same smooth-forcing approach for the porous medium equation, Boussinesq and 3D Euler, also with Lean. A priority dispute arose; by mutual acknowledgement the proofs differ, and the Euler results themselves differ: forced on one side, unforced on the other.
What remains open. Starting from rest sounds dramatic but means the opposite of how it reads: since the fluid was still, all the energy came from the external force. The blowup here is not "order out of nothing" but a very precisely computed pumping. The unforced problem, alternatives (A) and (B), is unsolved, the Clay Institute lists it as open, and OpenAI says it will not claim the prize. But the pattern in which old tools are carried to the end by one machine and checked by another has now been tried on a problem that resisted people for ninety years.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
A flow is constructed whose speed goes to infinity in finite time under a smooth external force, from rest and with finite energy. The singularity lives in a needle that shrinks faster than it speeds up, and the missing force is supplied by the fluid itself through the mean stress of oscillatory pulses. For us this is the first rigorously built example of a local singularity of a continuous medium at finite energy, an object to keep working with.
Where its strength is
The mechanism is explicit: needle vortex, annulus, two families of pulses, heat exterior. So it can be transferred: compute its effective metric, ask what vanishing viscosity does to it, try to reproduce it in a laboratory.
Where it could go next
- The link to general relativity. Since 1981 it has been known that disturbances in a moving fluid run on an effective metric, and horizons and Hawking radiation are reproduced in water and in atomic clouds. The next step suggests itself: compute the acoustic metric of the constructed flow. A needle with divergent speed and vanishing energy is a candidate for a point where the effective metric loses smoothness, a laboratory analogue of a spacetime singularity at finite energy. The boundary is honest: analogue gravity transfers the behaviour of fields on a curved background, not Einstein's equations, and that is exactly why the question is solvable.based on: Black Holes in a Sink of Water · Quantum Analog of a Black Hole in an Atomic Cloud · Black Holes from Light and Vortices · Hawking Radiation on a Tabletop arXiv:2604.24861 arXiv:2602.20508 arXiv:2603.01664 arXiv:2602.04593
- The singularity as a centre of anti-entropy. Working hypothesis: a local singularity is a window where the second law stops being absolute and order can be born from chaos. The constructed flow admits a direct test: the entropy budget separately in the core, in the annulus and outside. Energy is injected by the force, the global balance is positive; the question is about a region where local entropy production is negative and held by the pulses. If it exists and its size follows the scalings of the paper, a 'centre of anti-entropy' becomes a measurable quantity.based on: Twilight of Time: Why the Arrow Fades but Time Endures · Hourglass of Entropy: How Cold Atoms Keep Time arXiv:2508.14312 arXiv:2509.07745 arXiv:2506.01351
- Superfluids and Bose–Einstein condensates. In a classical fluid viscosity eats the pulses, and that is part of the construction; in superfluid helium and in a condensate viscosity vanishes, vortices are quantized, they self-assemble into lattices, quantum turbulence is observed in the laboratory. Hence the thought: where there is no dissipation, the singularity may not disperse but hold. Programme: carry the annulus-pulse mechanism into the Gross–Pitaevskii equation and find what replaces viscous damping.based on: Friction Vanishes Before Your Eyes · Quantum Leapfrog: How Vortices Play in Liquid Light · Vortex Rings Without a Grid arXiv:2601.00190 arXiv:2512.07935 arXiv:2601.06942 arXiv:2512.21416
- A macroscopic resonator. Two counter-rotating wheels in antiphase are a driven rotating shear flow, the very background on which the pulses grow by the centrifugal mechanism. It can be measured directly: the Reynolds stress in the annulus and its dependence on rotation speed. Agreement of sign and growth with the construction would be the first laboratory point; the analogue 'bomb' in a cup with a whirlpool shows such measurements are done.based on: How a Whirlpool in a Cup Becomes a Black Hole Bomb arXiv:2511.05351 arXiv:2601.01258 arXiv:2501.00037
- Two paths to one door. A resonator in the metric and a resonator in a quantum fluid are described by one mathematics: a sheared background, an annulus, oscillations supplying stress. It is worth building a correspondence table: what in the needle answers to the horizon, to a quantized vortex, to the pumping. Then one sees which door is closer to experiment and which to theory.based on: Black Holes from Light and Vortices · Quantum Analog of a Black Hole in an Atomic Cloud arXiv:2603.01664 arXiv:2602.20508
Worth trying
- Derive the acoustic metric of the constructed flow and check whether its curvature diverges as t → 1.
- Compute local entropy production over the three zones of the construction and look for a region with negative balance.
- Repeat the annulus-pulse mechanism in a Gross–Pitaevskii model at zero viscosity.
- Build a rotating tank with counter-rotating drives and measure the Reynolds stress in the annulus.
Why this may matter
If a flow singularity transfers to an effective metric, ordinary and quantum fluids become a testbed for questions that cannot be tested in gravity itself: how entropy behaves next to a singularity and whether order can hold where the continuous description breaks. OpenAI's paper gives such a testbed, for the first time, an explicit, rigorously built singularity at finite energy.
Nearby works in our archive
- Black Holes in a Sink of Watersimilarity 0.00
- Quantum Analog of a Black Hole in an Atomic Cloudsimilarity 0.00
- Black Holes from Light and Vorticessimilarity 0.00
- Hawking Radiation on a Tabletopsimilarity 0.00
- How a Whirlpool in a Cup Becomes a Black Hole Bombsimilarity 0.00
- Friction Vanishes Before Your Eyessimilarity 0.00
- Quantum Leapfrog: How Vortices Play in Liquid Lightsimilarity 0.00
- Vortex Rings Without a Gridsimilarity 0.00
- Twilight of Time: Why the Arrow Fades but Time Enduressimilarity 0.00
- Hourglass of Entropy: How Cold Atoms Keep Timesimilarity 0.00
What we make of it
This work has not been peer-reviewed. What follows is our reading of it.
What works
Inverting the problem: the flow is written down, the force is defined as the residual, and the whole difficulty is honestly moved into the smoothness of the residual. The self-limiting pulses that flare up on the shear and die out by viscosity, with their mean stress fitted to the divergence of the background, form a complete, checkable mechanism rather than an upper bound.
What would make it stronger
- Give a human exposition of Sections 4–9 shorter than a hundred pages: the physics is clear in six pages, but the road from it to the estimates is open only formally.
- State explicitly what the Lean formalization covers (Theorem 1.1 in full or particular estimates) and how text and code correspond: without this 'checked in Lean' reads broader than it may be.
- A numerical illustration of the constructed flow on a finite τ horizon would show the scales and pulses visibly and give an independent check of the choice of h and wavelengths.
- Discuss which part of the construction is obstructed by the absence of a force: that is the only road from (C) to (A).
Questions for the author
- How stable is the construction under perturbations of the background and pulses: does blowup persist in an open neighbourhood of the constructed data?
- Does the self-feeding mechanism of the pulses give any hint for the unforced case, where the stress would have to arise from the initial data?
- What exactly is formalized in Lean: the full Theorem 1.1 with supports and estimates, or its nonexistence corollary?
No human signed the paper: the author line reads OpenAI. Around ten thousand agents found the proof in 88 hours, and it was checked for 17 more hours not by a person but by Lean.
Take the whole work
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.