SOLVED: Navier-Stokes cracked by AI - Numberphile

Summary
Physicist Tony Padilla discusses OpenAI’s reported resolution of the Navier–Stokes Millennium Prize problem regarding finite-time blowup with forcing, filmed and interviewed by Brady Haran for Numberphile. Padilla breaks down the mathematics of the Navier–Stokes equations, how the AI-constructed blowup mechanism operates, and the mathematical history and priority controversy surrounding the discovery.

What is shown

• [00:00] Padilla uses a mug of stirred tea to illustrate fluid vortex flow and introduces the Navier–Stokes equations written on brown paper.
• [01:05] Display of OpenAI’s preprint cover (Finite Time Blowup for Navier-Stokes) and blog post excerpts detailing agent compute statistics.
• [01:45] Mathematical breakdown of the incompressible Navier–Stokes equations, highlighting velocity ($u$), pressure ($p$), viscosity ($\nu$), external force ($f$), and the incompressibility condition ($\nabla \cdot u = 0$).
• [03:40] Overview of the Clay Mathematics Institute’s Millennium Prize Problems and Charles Fefferman’s four official problem formulations (A, B, C, D) on paper [10:50].
• [07:55] Dimensional scaling analysis deriving the Reynolds number ($Re$) comparing the nonlinear advective term $(u \cdot \nabla)u$ to the viscous dissipation term $\nu \nabla^2 u$.
• [13:55] Detailed sketch and OpenAI diagram of the blowup vortex core featuring inward radial flow, axial stretching jets, anisotropic length scales ($l_r, l_\theta$), and velocity scaling exponents with parameter $0 < h < 0.01$.
• [19:15] Diagram explaining how the “wild core” is glued to a “boring exterior” across an annular region using high-frequency oscillatory fluctuations whose linear terms average to zero while quadratic nonlinearities produce the required driving force.
• [22:15] Historical survey of milestones, displaying papers and portraits of Jean Leray, Caffarelli–Kohn–Nirenberg (CKN), Terence Tao, and Córdoba–Martínez-Zoroa–Zheng (CMZ).
• [26:07] Discussion of recent work by Tristan Buckmaster and Levent Alpöge, displaying statements regarding the authorship and priority dispute between OpenAI and Anthropic-affiliated researchers.

Claims & numbers

• Padilla states that the Navier–Stokes equations describe non-relativistic, incompressible fluid flow [04:08].
• Padilla cites OpenAI’s release stating they deployed on the order of 10,000 concurrent agents, which sent approximately 2.7 million messages and used roughly 130 billion output tokens [01:14].
• Padilla reports estimated compute costs for OpenAI’s Navier–Stokes proof run ranging from $6 million to $15 million, completing the core work in about 88 hours [01:23, 01:37].
• Padilla explains that the Clay Millennium Prize offers a $1 million prize for resolving the problem [00:44].
• In the AI blowup construction, Padilla states the radial length scale scales as $l_r \sim (t_b - t)^{1/2}$ and angular length scale as $l_\theta \sim (t_b - t)^{1/2 - h}$, with $0 < h < 0.01$, while radial velocity scales as $u_r \sim (t_b - t)^{-1/2}$ and angular velocity as $u_\theta \sim (t_b - t)^{-(1/2 + h)}$ [15:13–16:16].
• As a result of this scaling, Padilla shows the angular Reynolds number $Re_\theta$ blows up proportionally to $(t_b - t)^{-h}$, enabling the nonlinear term to dominate viscosity [17:35].

Notable quotes

• [01:04] “Not by a person, but by AI... by OpenAI so to speak.”
• [17:41] “...the Reynolds number along the angular direction blows up, which means there’s a component of the nonlinear term which is winning.”
• [21:09] “They fake the fake force that way. That’s how it’s done—by these pulses that average to nothing, but their squares don’t average to nothing.”

Assessment
This is an educational mathematics review and commentary video from Numberphile. Padilla explains the mathematical mechanics and community reception of OpenAI’s preprint on paper without independently re-deriving the full machine-generated proof during the recording.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.