A few weeks ago, I published an experiment in which I asked frontier models to reason from the three-dimensional incompressible Navier–Stokes equations themselves, without searching for an existing solution.
It did not solve the problem. What it produced was a set of constraints on any possible singularity: it would need sustained transfer to smaller scales, geometry far from special non-stretching configurations, and—most importantly—coherent nonlinear alignment. Large norms alone were not enough.
Now OpenAI has released a claimed construction of finite-time blow-up for the forced three-dimensional Navier–Stokes equations. It is a much more specific and technically formidable result than anything in my post. But reading it, I was struck by how much of its physical structure fits the picture that emerged from that earlier analysis.
The overlap is real. The differences are even more instructive.
The important distinction: this is forced Navier–Stokes
My post considered the usual unforced equation:
$$ \partial_t u + (u\cdot\nabla)u = -\nabla p + \nu\Delta u. $$That equation has the familiar energy identity: kinetic energy decreases, and viscosity dissipates it.
OpenAI’s construction instead solves
$$ \partial_t u + (u\cdot\nabla)u = -\nabla p + \nu\Delta u + f, $$where \(f\) is smooth and compactly supported in both space and time.
That is not a cosmetic difference. It means the unforced energy and enstrophy identities I used do not directly constrain this solution: forcing adds extra terms to both balances. The construction is nevertheless highly nontrivial because the force remains perfectly smooth even while the velocity becomes unbounded. You are not simply inserting an infinite force by hand.
The whole game is to arrange a flow whose acceleration, pressure, viscosity, and nonlinear transport individually become enormous, but cancel so precisely that their residual—the external force—extends smoothly through the singular time.
That is the central new degree of freedom the forced problem gives you.
The singularity is a collapsing, spinning, stretched vortex
The construction’s core is not an arbitrary high-frequency cascade. It is an axisymmetric vortex that spirals inward toward an axis while flowing outward along that axis.
As the core contracts, angular momentum is carried to smaller radius and the azimuthal velocity rises. Incompressibility prevents matter simply accumulating on the axis: it is expelled axially. The velocity diverges in a shrinking region, while the total kinetic energy remains bounded—in fact, the energy of the core itself goes to zero.
The core is also anisotropic. Its radial scale shrinks like roughly \((T-t)^{1/2}\), while its axial scale shrinks a little more slowly. So it becomes an increasingly slender column.
That is already close to the broad picture in my post: a singularity cannot just be “large vorticity.” It needs an organised geometry that keeps moving activity to smaller scales while respecting incompressibility and the energy budget.
Alignment really is part of the mechanism
The strongest part of my earlier post was probably the claim that the missing dynamical variable is alignment.
I wrote the rate of change of a characteristic squared wavenumber as a competition between nonlinear vortex-stretching production and a viscous penalty associated with spectral spread. The point was not merely that nonlinearity must be large. It must point in the right direction.
OpenAI’s construction makes that idea concrete.
Around the collapsing vortex, it places spatially oscillatory pulses. Their velocities have zero angular mean, but their quadratic momentum fluxes do not. By selecting the orientation, placement, and timing of those pulses, their nonlinear interactions generate a mean stress in exactly the directions needed to cancel the singular residual of the background vortex.
This is phase-sensitive nonlinear engineering. The pulses cannot merely be energetic. Their correlations have to have the right sign and tensor structure.
That is very close to the intuition behind saying that a cascade needs coherent alignment, not just a worst-case norm estimate.
Viscosity is not defeated; it is managed
My post described a possible singularity as becoming increasingly Euler-like at its strongest bursts. That needs qualification after reading this construction.
The azimuthal Reynolds number of the vortex does diverge: in that sense, angular motion becomes increasingly inertia-dominated. But the radial Reynolds number stays order one. Radial diffusion remains in the leading-order balance all the way to blow-up.
This is therefore not a simple story in which viscosity eventually becomes irrelevant. The construction lives in a mixed regime:
- angular spin-up becomes increasingly high-Reynolds-number;
- radial inflow and radial diffusion continue to compete;
- viscosity actively shapes which oscillatory pulses can grow.
In fact, the pulses are designed so that shear initially amplifies them, but their changing orientation shortens their radial wavelength. That increases their wavenumber and therefore their viscous damping. Eventually damping overtakes amplification.
So viscosity is not merely an obstacle that the singularity outruns. It is one of the components being balanced and exploited.
A useful correction to “no self-similarity”
My post argued that an unforced singularity could not simply converge to a bounded, fixed self-similar profile at critical scale. The new construction has a leading self-similar vortex profile, which might sound like a contradiction.
It is not.
The construction is anisotropically self-similar, and its critical \(L^3\) size diverges as the singularity is approached. It is not a bounded, precompact critical similarity orbit of the kind ruled out by the standard continuation theorem. Its shape is organised, but it escapes at the critical level.
That feels like the right refinement: a singularity may have a leading similarity structure, but not a tame universal profile that remains bounded in the critical topology.
Where this fits in the larger story
Tristan Buckmaster and Levent Alpöge’s accompanying statement places this in a longer programme associated with Diego Córdoba and Luis Martínez-Zoroa: construct forced blow-ups while keeping the forcing progressively more regular. Buckmaster and Alpöge report smooth-forced blow-up results for Euler and related equations; OpenAI’s claimed result extends the forced construction to full-viscosity Navier–Stokes.
The technical routes differ, but the family resemblance matters. This is not the discovery of a generic spontaneous singularity in freely evolving turbulence. It is the discovery—or at least the claimed construction—of an exquisitely designed flow in which the fluid’s own nonlinear stresses cancel what would otherwise require singular forcing.
Was the earlier analysis close?
Not close to a proof. Not close to this construction’s detailed machinery of profiles, oscillatory corrections, stress cones, and all-order cancellation.
But I think it was directionally close to the mechanism.
The construction supports three claims from the earlier piece:
A singularity needs more than large nonlinear terms; it needs coherent, correctly directed nonlinear transfer.
The relevant object is not a single shrinking vortex of fixed shape, but a multiscale and geometrically organised structure.
Viscosity’s role is subtler than “energy goes down, therefore no blow-up.” It constrains the route to concentration, but it can coexist with—and help organise—an extreme anisotropic collapse.
The striking lesson is that the route to singularity is not simply vortex stretching overpowering viscosity. It is a much more precise act of dynamical coordination: spiral inward, stretch outward, create oscillations, tune their phase, use their averaged stresses, and arrange cancellation so exact that the remaining external force stays smooth while the velocity diverges.
That is a far stranger—and more interesting—answer than “turbulence gets too violent.”
Links: my original post, OpenAI’s announcement, the claimed proof, and Tristan Buckmaster’s statement. As with any newly released major mathematical claim, the work still needs sustained independent scrutiny by the mathematical community.