OpenAI Reports Navier-Stokes Singularity Construction Built With 10,000 Agents

According to OpenAI's result as reported by Leiphone, the construction allows a three-dimensional fluid, under smooth external force and finite kinetic energy, to evolve until local velocity becomes unbounded. After publication, debate followed. Chris Combs, an aerospace engineering professor, cautioned that the result does not amount to a general method for computing arbitrary fluid motion and would not immediately change existing computational fluid dynamics work, Leiphone reported. On the mathematics side, the article said, the discussion concerns whether smooth three-dimensional solutions remain regular; a finite-time blow-up directly touches that proposition. Similar research by New York University mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge was also drawn into the discussion, raising questions about research routes and attribution, according to Leiphone.

Leiphone also reported claims circulating among academics and industry figures on social media that OpenAI may have learned Buckmaster's team's core framework through channels such as academic exchanges or an unpublished preprint, and then used internal compute to scoop the result. The article described a further rumor that OpenAI offered Buckmaster first authorship on condition that Anthropic collaborator Levent Alpoge be excluded. These claims were presented as rumor, not as confirmed facts.

The Leiphone article argued that AI did not independently discover the fluid mechanics; human mathematicians supplied the central intuition. Buckmaster and others, it said, contributed ideas such as deliberately departing from a critical scale, high-frequency oscillatory pulses, and separating local velocity from overall kinetic energy. The agents' role was to search many mutually constraining, high-dimensional nonconvex conditions and fill in a closed construction.

To create the singularity, OpenAI designed an anisotropic self-similar scaling from an axisymmetric swirling vortex core, according to the article. Near the singularity, the core contracts rapidly in the radial direction and more slowly in the axial direction, becoming a thin rotating column. A small exponent bias lets pointwise velocity grow without making total kinetic energy unbounded, because the high-speed region shrinks faster. The bias also rearranges viscosity: radial transport keeps its leading competition with diffusion, angular inertia grows stronger, and axial diffusion becomes a lower-order effect.

Connecting the inner vortex core to an outer, viscous-diffusion-controlled flow leaves a ring-shaped transition zone and a momentum residual. Rather than defining that residual as an external force, which could itself diverge, the construction converts it into momentum transport by fine-scale motion. High-frequency oscillations have small mean velocity but, through the nonlinear quadratic terms of Navier-Stokes, produce a steady mean momentum flux known as Reynolds stress. Different wave modes cover only a limited stress cone, so the background residual must be moved into that cone. Small-amplitude, high-frequency radial perturbations adjust local shear; low-frequency corrections restore global moment conditions. The same shear amplifies the pulses and shortens their wavelength, after which viscosity dissipates them.

According to Leiphone, the search problem is ill-suited to single-thread reasoning because self-similar scales, profiles, boundary extensions, external flow, local shear, stress-cone constraints, radial moments, oscillation modes and higher-order corrections must all close together. About 10,000 agents maintained many candidate structures in parallel. OpenAI also put nearly 100 agents on the adjacent Euler problem, removing viscosity to expose nonlinear singularity structure, then moved useful geometry and construction ideas back to the Navier-Stokes route. The route generated about 2.7 million agent messages and 130 billion output tokens, much of it parameter tests, failed constructions and intermediate lemma searches.

Lean was used to check the final proof chain, with every variable type, lemma premise and dependency made explicit, Leiphone reported. OpenAI added Comparator, an independently formalized Navier-Stokes proposition, to check that the target statement had not been quietly weakened by changing function spaces, lowering smoothness requirements for forcing, or altering the definition of a global solution.

Leiphone concluded that the technical weight lies less in a single trick than in the interlocking of self-similar scaling, anisotropic structure, residual absorption through Reynolds stress, multiscale corrections, large-scale agent search and formal verification. In this account, OpenAI demonstrated a system that turns broad search into a strict proof of the original proposition.