From Elevator to Vortex: Why the Navier-Stokes Equation Jammed and What OpenAI Claims

The video opens with Newton's second law, the compact F=ma relation most of us last met in a school classroom, and connects it to OpenAI's claim about the Navier-Stokes smoothness problem. A worked elevator example carries the intuition: a 1,000 kg cabin under 9.8 m/s2 of gravity feels about 9,800 N of pull, so the cable must match it, and adding 1,000 N more accelerates the whole cabin at 1 m/s2.

That arithmetic stays easy because a solid body moves as one piece, and any two points on the cabin share the same acceleration. Fluids refuse that convenience, since neighboring parcels of water can head in different directions at different speeds and different accelerations at once. Every one of those local differences has to be represented in the mathematics, which is where the trouble starts.

Applying Newton's law to a fluid yields the Navier-Stokes equation, assembled gradually by many contributors across history. One side of the relation gathers the forces on a small region of fluid, including pressure, viscosity, and usually gravity. The other side tracks how the fluid answers those forces through its acceleration and its density.

A garden hose makes the force side concrete: water near the tap sits at higher pressure than water leaving the nozzle, so a pressure gap pushes the flow along. Viscosity slows the layer hugging the hose wall through friction while the central stream runs faster. Gravity or other imposed forces can tilt the balance further, and the equation sums all of them.

Acceleration in a fluid splits into two channels worth keeping apart. Turning the tap changes the speed at a fixed spot over time, while a funnel speeds water up simply because the channel narrows downstream. Density completes the picture by recording how much mass sits in a given volume, the reason salty water and fresh water behave differently.

The equation earns its keep: for roughly two centuries it has guided pipe design, weather forecasting, and ocean current modeling with genuine engineering success. Yet the mathematics was never shown to stay well-behaved under every fair condition, and that gap is the famous million-dollar question. Either a proof shows smooth solutions always persist in three dimensions, or a counterexample shows the equations can blow up into a singularity.

OpenAI asserts the second route, presenting a configuration where the equations break down, found with about 10,000 agents running together and 130 billion tokens spent. If independent checking sustains that claim, it would be a landmark, and it would hint at what frontier mathematical work looks like when massive agent orchestration joins the effort.

The alleged counterexample is pictured as a coiling swirl: fluid circles a central line while being drawn inward and stretched upward at the same time. The nearest everyday cousin is a tornado, though the configuration is far stranger than anything routinely observed. The Millennium formulation demands exactly this stringency, covering every valid three-dimensional starting state or exhibiting finite-time breakdown.

To see why rival researchers cry foul, the video rewinds to Euler, whose older flow equation omits viscosity. Tristan Buckmaster and Levent Alpoege of Anthropic spent about a year hunting a blowup mechanism there, building on earlier work, and they announced a forced solution. Forced means an added external push helped trigger the breakdown, so the result was progress rather than a finished proof, with viscosity still unaccounted for.

OpenAI then moved fast on the same trail: an unforced Euler blowup in about 50 hours with 100 agents, followed by a full Navier-Stokes case in roughly 88 hours. The dispute turned personal when Buckmaster was reportedly offered shared credit on condition that Alpoege's name be left out, and when message logs surfaced alongside denials. The hardest allegation, denied repeatedly by OpenAI, is that inspiration was lifted from researcher chat histories inside Codex. The video closes on the deeper unease: who gets to make discoveries when 10,000 agents require a frontier lab's budget, and award rules written for people have no line for machine authors.