OpenAI Solves Forced Navier–Stokes Blowup as Core Physical Case Looms

How OpenAI's AI "Broke" Navier–Stokes: The Finite-Time Blowup Proof, Explained | Loop State EP.03 — Xiaol.x

Xiaol.x

OpenAI's counterexample satisfies the Clay Millennium conditions by reverse-engineering a collapsing vortex whose mathematical blowup is canceled by an artificial external force, leaving the question of whether natural fluids break down completely untouched.

AI-assisted summary of How OpenAI's AI "Broke" Navier–Stokes: The Finite-Time Blowup Proof, Explained | Loop State EP.03 — original source. Check the original for full context.

Source published:

Key ideas

  • OpenAI released a 166-page paper claiming to resolve alternatives (C) and (D) of the Clay Millennium Prize problem by proving finite-time blowup for the 3D forced Navier–Stokes equations. (0:01)
  • The proof operates 'backwards' by constructing an engineered, diverging flow and spending the analytical bulk ensuring the residual forcing function and all its derivatives remain smooth through the singularity. (4:19)
  • Although the peak fluid speed reaches infinity at t = 1, total kinetic energy remains finite—and drops toward zero in the core—because the singular volume contracts faster than speed diverges. (11:47)
  • The paper does not resolve alternatives (A) and (B): whether an unforced fluid with f = 0 can develop a singularity remains entirely open, and the Clay Institute has not certified the result. (13:24)

Detailed notes

## Background Published on YouTube on September 12, 2026, by the channel Xiaol.x, this episode of the podcast Loop State (EP.03) breaks down OpenAI's 166-page paper, 'Finite Time Blowup for Navier–Stokes.' Hosts Mara and Theo walk through how the paper claims to resolve alternatives (C) and (D) of the Clay Millennium Prize problem by constructing a smooth fluid driven by a smooth force that hits infinite speed in finite time while maintaining finite total energy.

## TL;DR The hosts explain how the proof bypasses forward numerical integration by engineering an explicit blowup velocity field backwards, shifting the mathematical burden into showing that the residual external force remains completely smooth through the singularity. By introducing an anisotropic exponent h, the construction breaks collinear scaling and allows an iterative cycle of oscillating wave pulses to cancel singular residual terms order by order until the force is flat. Total kinetic energy remains finite—and actually drops to zero inside the collapsing vortex core—because the core's volume contracts faster than its peak speed squares grow. Finally, the breakdown is converted into a Millennium Prize counterexample via uniqueness arguments, though the hosts stress that the proof only resolves the artificially forced problem and leaves the classic unforced Navier–Stokes question completely open.

## Atmosphere & style The discussion carries a lucid, intellectually rigorous, and pedagogical energy driven by Mara and Theo's collaborative back-and-forth rhythm. Moving seamlessly from vivid mechanical intuitions—such as 'a vortex that eats itself'—into fine-grained mathematical calculus, the conversation feels like an advanced chalkboard seminar brought to life. Crucially, the tone avoids uncritical tech hype, maintaining a sober, disciplined scientific skepticism that dedicates substantial care to outlining boundaries, computational caveats, and the lack of external peer review.

## Bulletin - [00:01](https://www.youtube.com/watch?v=RzJokrUsnbM&t=1s) OpenAI released a 166-page paper claiming to resolve alternatives (C) and (D) of the Clay Millennium Prize problem by proving finite-time blowup for the 3D forced Navier–Stokes equations. - [04:19](https://www.youtube.com/watch?v=RzJokrUsnbM&t=259s) The proof operates 'backwards' by constructing an engineered, diverging flow and spending the analytical bulk ensuring the residual forcing function and all its derivatives remain smooth through the singularity. - [11:47](https://www.youtube.com/watch?v=RzJokrUsnbM&t=707s) Although the peak fluid speed reaches infinity at t = 1, total kinetic energy remains finite—and drops toward zero in the core—because the singular volume contracts faster than speed diverges. - [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) The paper does not resolve alternatives (A) and (B): whether an unforced fluid with f = 0 can develop a singularity remains entirely open, and the Clay Institute has not certified the result.

## Key takeaways - [00:01](https://www.youtube.com/watch?v=RzJokrUsnbM&t=1s) The paper targets the forced Navier–Stokes problem, constructing a smooth fluid driven by a smooth force that hits infinite speed at t = 1. - [01:34](https://www.youtube.com/watch?v=RzJokrUsnbM&t=94s) The Millennium Prize problem formulation splits into four cases: alternatives (A) and (B) examine unforced fluids (f = 0), while (C) and (D) involve smooth external forcing. - [04:19](https://www.youtube.com/watch?v=RzJokrUsnbM&t=259s) Instead of solving the PDE forward in time, the authors engineer a diverging velocity candidate u and evaluate the residual force f, shifting the entire proof burden into proving f remains smooth. - [05:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=329s) A naive self-similar profile fails because all terms in the residual lock into the same temporal power, making differential cancellations impossible across the domain. - [06:53](https://www.youtube.com/watch?v=RzJokrUsnbM&t=413s) Introducing an anisotropic scaling parameter h decouples radial contraction (tau^(1/2)) from vertical contraction (tau^(1/2 - h)), splitting residual terms into distinct temporal orders. - [08:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=509s) Oscillating wave pulses with zero-mean velocity generate non-zero mean momentum fluxes, cancelling leading-order residual stresses in an annular region around the core. - [10:13](https://www.youtube.com/watch?v=RzJokrUsnbM&t=613s) An iterative four-step correction cycle improves residual decay order-by-order, yielding a 'flat' force whose space-time derivatives vanish faster than any power near the blowup. - [11:47](https://www.youtube.com/watch?v=RzJokrUsnbM&t=707s) Due to core volume scaling of tau^(3/2 - h) versus velocity squared scaling of tau^(-1 - 2h), the kinetic energy in the collapsing core drops to zero as t approaches 1. - [12:26](https://www.youtube.com/watch?v=RzJokrUsnbM&t=746s) Outside the core annulus, the flow reduces to a pure swirl obeying the radial heat equation, allowing the solution to be smoothly cut off without disturbing the blowup column. - [12:26](https://www.youtube.com/watch?v=RzJokrUsnbM&t=746s) Lemma 10.5 uses weak-strong uniqueness to show any smooth candidate solution must coincide with the constructed blowup before t = 1, ruling out the existence of global smooth solutions. - [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) The forcing function is meticulously engineered and compactly supported rather than physically occurring in nature, settling the Millennium Prize statement strictly as written. - [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) OpenAI reported verifying the proof in Lean using internal AI models, but the proof lacks external peer review and has not been certified by the Clay Mathematics Institute. - [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) Priority for resolving finite-time blowup for the forced Euler equations belongs to Levent Alpöge and Tristan Buckmaster, which OpenAI explicitly acknowledged.

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=RzJokrUsnbM&t=1s) The Counterexample: A Vortex That Eats Itself

On September 8, 2026, OpenAI released a 166-page paper titled *"Finite Time Blowup for Navier–Stokes"*. Hosts Mara and Theo open the discussion by examining what this counterexample flow actually looks like before jumping into the formal mathematics. Theo describes the core mechanism as "a vortex that eats itself." In this configuration, fluid spirals inward toward a vertical axis and is ejected along that axis—shooting upward on one side and downward on the other. As the system evolves toward time $t = 1$, the vortex core contracts, thinning out faster than it shortens in height. Consequently, the fluid velocity inside this shrinking column climbs without bound, hitting infinite speed precisely at time $t = 1$.

```mermaid flowchart TD accTitle: Fluid dynamics of the self-collapsing vortex accDescr: Fluid spirals radially inward into a narrowing column, which accelerates vertical jets and drives velocity to infinity within a shrinking core volume.

subgraph INFLOW["Inflow & Contraction"] A["Inward Radial Spiral"] --> B["Vortex Core Thinning"] end

subgraph BLOWUP["Core Dynamics"] B --> C["Axial Ejection (Up/Down Jets)"] C --> D["Speed Climbs Unbounded"] D --> E["Blowup at t = 1 (Infinite Speed)"] end

subgraph ENERGY["Volume & Energy"] B --> F["Core Volume Shrinks Rapidly"] F --> G["Kinetic Energy in Core -> 0"] end ```

The construction yields a smooth fluid driven by a smooth force that reaches an infinite speed within finite time, all while maintaining finite total energy. Crucially, the total kinetic energy contained within the vortex core does not explode; it actually drops toward zero as the core collapses. Because an enormous peak speed is concentrated inside an vanishingly small spatial volume, the integral of energy remains well-behaved. Theo and Mara emphasize an important contextual caveat: while OpenAI reported that the entire proof is formalized in Lean, the Clay Mathematics Institute has not certified the result, and no external peer review has taken place yet.

---

[01:34](https://www.youtube.com/watch?v=RzJokrUsnbM&t=94s) Navier–Stokes Fundamentals and the Four Clay Millennium Problems

Navier–Stokes is simply Newton's second law ($F = ma$) applied to a microscopic parcel of fluid. Dividing force by mass yields acceleration, which decomposes into the unsteady time derivative $\partial_t u$ and the non-linear convective transport term $(u \cdot \nabla)u$, accounting for the parcel changing position within a varying velocity field. The balance of forces acting on the parcel includes:

1. **Pressure gradient** ($-\nabla p$): Pushes fluid parcels from high-pressure regions to low-pressure regions. 2. **Viscous dissipation** ($\nu \Delta u$): Internal friction between neighboring fluid layers that smooths out sharp velocity gradients. 3. **External forcing** ($f$): An applied body force driving the fluid.

These forces are bound together under the incompressibility constraint $\nabla \cdot u = 0$, meaning fluid cannot be squeezed along one axis without stretching along another. This divergence-free condition turns pressure into a global enforcer across the domain rather than an isolated local effect. The central mathematical battle is between viscosity (a linear term that smooths variations) and transport (a quadratic term $(u \cdot \nabla)u$ that steepens gradients). The 90-year-old open question asks whether viscous smoothing always wins in three dimensions.

The Clay Millennium Prize problem formulated by Charles Fefferman offers four distinct paths:

| Alternative | Domain | Forcing ($f$) | Problem Direction | |:--- |:--- |:--- |:--- | | **(A)** | Whole space $\mathbb{R}^3$ | Zero ($f = 0$) | Regularity (Smooth forever) | | **(B)** | Periodic torus $\mathbb{T}^3$ | Zero ($f = 0$) | Regularity (Smooth forever) | | **(C)** | Whole space $\mathbb{R}^3$ | Smooth external $f$ | Breakdown / Finite-time blowup | | **(D)** | Periodic torus $\mathbb{T}^3$ | Smooth external $f$ | Breakdown / Finite-time blowup |

Theorem 1.1 of the OpenAI paper directly targets alternative (C): starting from rest ($u|_{t=0} = 0$), driven by a smooth and compactly supported force, the flow develops infinite speed at $t = 1$ while keeping total energy bounded. Because the construction is compactly supported in space and time, Corollary 10.6 wraps the solution onto the torus, establishing alternative (D). Theo clarifies that this does not resolve whether an unforced fluid can blow up—alternatives (A) and (B) remain wide open, meaning the paper does not show that ordinary "water in a glass can blow up."

---

[04:19](https://www.youtube.com/watch?v=RzJokrUsnbM&t=259s) Designing a Breakdown by Running the Equation Backwards

Directly solving the non-linear partial differential equation forward in time to hit a singularity is notoriously difficult. Instead, the paper bypasses forward integration entirely by running the equation backwards. One picks an explicit, smooth velocity candidate $u$ and pressure $p$ whose peak velocity blows up at $t = 1$. The residual is then evaluated directly:

$$f \coloneqq \partial_t u + (u \cdot \nabla)u - \nu \Delta u + \nabla p$$

By defining $f$ as this exact residual, Navier–Stokes is satisfied by construction.

This reverses the mathematical hurdle. Navier–Stokes is never in danger of failing; rather, the entire analytical burden shifts into the regularity of $f$. The challenge is finding a velocity field $u$ that diverges at $t = 1$ while ensuring that $f$ and all of its derivatives remain completely smooth, bounded, and well-behaved through $t = 1$. Individual terms in the residual may blow up individually, but their divergences must cancel out completely when summed together.

```mermaid flowchart LR accTitle: The residual cancellation architecture accDescr: Diverging velocity components generate diverging individual terms whose singularities must cancel out to yield a smooth residual force f.

subgraph VELOCITY["Engineered Flow u"] U["Choose Diverging Flow u<br/>(Peak speed -> inf at t=1)"] end

subgraph RESIDUAL["Residual Computation"] U --> T1["Time Derivative: dt u (Diverges)"] U --> T2["Transport: (u. grad)u (Diverges)"] U --> T3["Viscosity: - nu Delta u (Diverges)"] U --> T4["Pressure Gradient: grad p"] end

subgraph CANCELLATION["Target Property"] T1 & T2 & T3 & T4 --> RES["Sum = Residual Force f"] RES --> CANCEL{"Engineered Cancellation"} CANCEL --> OK["f and all derivatives remain SMOOTH through t=1"] end ```

---

[05:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=329s) Why the Naive Self-Similar Guess Fails

To illustrate why this cancellation is difficult, Theo and Mara explore a simplified 1D toy cartoon: 1D space, $\nu = 1$, zero pressure, setting $f = u_t + u u_x - u_{xx}$. Testing a standard self-similar Gaussian bump scaling with $\tau = 1 - t$:

$$u_0(x, t) = \tau^{-1/2} \exp\left(-\frac{x^2}{\tau}\right)$$

The peak amplitude grows as $\tau^{-1/2}$, and the spatial width contracts as $\sqrt{\tau}$, creating infinite velocity at $t = 1$. However, evaluating the force at $x = 0$ exposes a fatal breakdown: - By symmetry, the convective transport term vanishes: $u u_x|_{x=0} = 0$. - The time derivative yields: $\partial_t u_0|_{x=0} = \frac{1}{2}\tau^{-3/2}$. - The viscous term contributes: $-u_{xx}|_{x=0} = 2\tau^{-3/2}$. - Summing these yields a total force of $\frac{5}{2}\tau^{-3/2}$, which blows up violently as $\tau \to 0$.

Because every term carries the exact same power of $\tau$, they add together rather than canceling. In any single-scale self-similar Ansatz, the differential operators preserve collinear scaling, making cancellation impossible across the domain.

---

[06:53](https://www.youtube.com/watch?v=RzJokrUsnbM&t=413s) Breaking the Tie with the Anisotropic Exponent $h$

To overcome collinear scaling, the paper introduces a tiny parameter $h \in (0, 1/100)$. This exponent breaks the single-scale symmetry by decoupling spatial and temporal contractions: - The core radius contracts at rate $\tau^{1/2}$. - The core height contracts at rate $\tau^{1/2 - h}$.

Because the height shrinks slower than the radius, the vortex core deforms into an increasingly slender column. The velocities scale anisotropically as well: swirl and axial velocities scale as $\tau^{-1/2 - h}$, while the radial inflow scales as $\tau^{-1/2}$. This means the swirl Reynolds number diverges to infinity, causing fluid parcels to complete ever more revolutions per diffusion time, while the radial Reynolds number remains bounded so viscosity can still compete with inflow.

In the 1D cartoon, this corresponds to an anisotropic profile $\tau^{-a} \exp(-\tilde{x}^2)$ where $\tilde{x} = x / \tau^b$, with $a$ and $b$ offset by $\pm h/3$. The force at the origin splits into terms of different orders:

$$a \tau^{-a - 1} + 2 \tau^{-a - 2b}$$

Because the exponents differ by $\frac{2}{3}h$, the terms no longer lock together. While this split initially seems worse because both terms diverge, it makes the diverging pieces distinct. The exponent $h$ provides separate "handles" that allow corrections to target individual divergent components without disrupting the others.

---

[08:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=509s) Canceling Residuals via Oscillating Wave Pulses

With separate handles available, corrections can be applied to cancel residual terms. In the 1D cartoon, this is demonstrated by adding an odd pulse:

$$w = -c(\tau) \tau^{-a} x \exp(-\tilde{x}^2)$$

Although $w(0) = 0$, its non-zero spatial slope at the origin interacts with the non-zero background velocity via the transport term $u w_x$. Setting $c(\tau) = a + 2\tau^{2h/3}$ forces the total residual at $x = 0$ to cancel out identically.

In the full 3D paper, the background vortex leaves an uncanceled residual within an annular region surrounding the core, which matches the negative divergence of a Reynolds stress tensor: - Pulses are introduced as oscillating waves whose velocity averages to zero around the azimuth, but whose momentum flux does not vanish. - Outward fluid motion carrying excess swirl and inward fluid motion carrying a swirl deficit both transport angular momentum radially outward. - Just as in the 1D product $u w_x$, zero-mean fluctuations yield a non-zero mean momentum flux. - Proposition 7.5 establishes that two wave families provide two independent flux directions, and the required residual stress lies strictly inside the cone they span, ensuring strictly positive weights. - Each wave pulse is seeded by an exponentially small forcing, amplified via background shear, and eventually dissipated by viscosity as shear cascades its wavelength down to dissipative scales.

---

[10:13](https://www.youtube.com/watch?v=RzJokrUsnbM&t=613s) The Iterative Cycle and the Flat Force

Canceling the leading-order residual creates secondary errors: the correction pulse $w$ generates a linear term (the pulse interacting with the background) and a quadratic term (the pulse interacting with itself). The paper formalizes this correction cycle using four distinct operations:

```mermaid flowchart TD accTitle: The four-step correction cycle accDescr: Iterative four-step cycle resolving wave amplitudes, pulse amplitudes, mean flow/pressure, and radial moments to eliminate errors order by order.

subgraph CYCLE["Four-Step Residual Correction Cycle"] STEP1["1. Solve wave amplitude equations<br/>(Leftover harmonics)"] STEP2["2. Adjust pulse amplitudes<br/>(Cross-covariance fixes average stress)"] STEP3["3. Correct angular mean flow and pressure"] STEP4["4. Solve 5 radial moment equations<br/>(Maintains compact support & incompressibility)"]

STEP1 --> STEP2 STEP2 --> STEP3 STEP3 --> STEP4 end

STEP4 -->|"Gains 1/10 per pass"| STEP1 ```

- Each iteration improves residual decay by $q^{h \sigma}$, where $\sigma$ begins at $1/5$ and gains $1/10$ with each pass, driving $\sigma \to \infty$. - Summing these corrections with spatial cutoff functions shrinking toward the singularity produces a **flat** residual. - "Flat" means every space-time derivative of the residual vanishes faster than any power of the distance to the singularity: bounded by $C_N q^N$ for all $N \ge 1$. - Lemma 10.3 extends this flat residual smoothly past $t = 1$, yielding an external forcing supported up to $t = 2$.

---

[11:47](https://www.youtube.com/watch?v=RzJokrUsnbM&t=707s) Finite Kinetic Energy in an Infinite-Speed Flow

A fluid reaching infinite speed while keeping finite total kinetic energy seems paradoxical, but Theo explains that it is entirely resolved by volume scaling:

- **Radius:** $r \sim \tau^{1/2}$ - **Height:** $z \sim \tau^{1/2 - h}$ - **Core Volume:** $V \sim r^2 z \sim (\tau^{1/2})^2 \tau^{1/2 - h} = \tau^{3/2 - h}$ - **Peak Speed Squared:** $|u|^2 \sim (\tau^{-1/2 - h})^2 = \tau^{-1 - 2h}$ - **Core Kinetic Energy:**

$$E_{\text{core}} \sim \int_{\text{core}} |u|^2 \, dV \sim \tau^{3/2 - h} \cdot \tau^{-1 - 2h} = \tau^{1/2 - 3h}$$

Because $h < 1/100$ (and strictly less than $1/6$), the exponent $1/2 - 3h$ remains strictly positive. Consequently, as $\tau \to 0$, the kinetic energy trapped inside the core goes to zero. The same volumetric bookkeeping applies to the dissipation rate, keeping integrated energy dissipation finite. Lemma 10.4 establishes these rigorous localized bounds directly from the governing equations, proving that infinite velocity can occur inside a localized singularity without exceeding finite energy bounds.

[12:26](https://www.youtube.com/watch?v=RzJokrUsnbM&t=746s) The Spatial Cutoff and Lemma 10.5: Turning Construction into Contradiction

To wrap up the blowup construction, the spatial boundaries of the flow have to be tamed. Outside the core and beyond the outer annulus, the fluid flow behaves as a pure swirl that solves the radial heat equation exactly. Because it satisfies the heat equation on its own, it requires absolutely no external forcing in that region. This convenient property allows the authors to cut everything off in space using smooth cutoff functions, leaving the collapsing core completely untouched. Furthermore, a standard scaling argument takes the solution derived for viscosity $\nu = 1$ and extends it to hold for every positive viscosity $\nu > 0$.

However, an important logical gap remains before concluding the proof: the Millennium Prize statement requires showing that *no* smooth, bounded-energy global solution can exist for this setup, whereas the authors have merely explicitly constructed one particular blow-up solution. The hosts explain that this gap is bridged by Lemma 10.5. By standard uniqueness and weak-strong uniqueness principles, any hypothetical smooth solution sharing the exact same forcing, zero initial data, and bounded energy must coincide identically with the constructed solution before time $t = 1$. Consequently, any smooth candidate inherits the exact same finite-time singularity, meaning it cannot be extended globally in time. That contradiction establishes Theorem 1.1.

```mermaid flowchart TD accTitle: Navier-Stokes Blowup Contradiction Flow accDescr: Diagram showing how the constructed blowup solution forces any hypothetical smooth solution to blow up via Lemma 10.5.

subgraph Construction ["Constructed Solution"] A["Core Swirl & Annulus"] --> B["Spatial Cutoffs"] B --> C["Blowup Solution (t < 1)"] end

subgraph Contradiction ["Uniqueness Contradiction"] D["Hypothetical Global Smooth Solution\n(Zero initial data, same smooth force)"] C --> E["Lemma 10.5: Must agree before t = 1"] D --> E E --> F["Smooth solution inherits blowup"] F --> G["No global smooth solution can exist\n(Theorem 1.1 Proven)"] end ```

- Outside the annulus, the fluid is a pure swirl governed entirely by the heat equation, requiring zero external forcing. - Smooth spatial cutoffs cleanly isolate the active collapsing column without disturbing the singular dynamics. - Scaling extends the viscosity $\nu = 1$ solution to arbitrary positive viscosity parameters. - Lemma 10.5 forces any candidate smooth solution with matching initial data and forcing to agree with the blowup profile, ruling out global smoothness.

---

[13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) Caveats and Boundaries: What the Paper Does Not Settle

With the proof explained, the hosts pivot directly to the caveats: "Now the part where we say what this does not settle. Boundaries—give me the list." They run through six essential points to prevent misunderstandings:

1. **Forced vs. Unforced**: The paper addresses only alternatives (C) and (D) of the Clay Millennium Prize formulation—the forced Navier–Stokes problem. Alternatives (A) and (B), which ask whether an unforced, smooth fluid left entirely alone can blow up from smooth initial data, remain completely untouched. 2. **Artificial, Tuned Forcing**: The forcing function is engineered, smooth, and compactly supported, meticulously tuned so that the corrective pulses arrive exactly when and where needed. "Nobody claims this flow arises in nature. It settles the Millennium statement as written, not a physical question." 3. **Verification Status**: While OpenAI claims the paper's Lean formalization was machine-checked in approximately 17 hours using GPT-6 Astra, this is purely internal self-reporting. As of September 11, 2026, there is no formal external peer review, and the Clay Mathematics Institute has stated it is not treating the Millennium Prize problem as resolved. 4. **Compute and Resource Claims**: OpenAI reported utilizing an unreleased internal model more capable than GPT-6 Astra, coordinating roughly 10,000 concurrent agents over 88 hours to produce the result, consuming roughly 130 billion output tokens. OpenAI has explicitly stated it does not plan to claim the $1 million prize money. 5. **Priority and Prior Art**: Priority for the forced Euler problem belongs to mathematicians Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU), who independently resolved forced Euler blowup using an Anthropic AI model. OpenAI explicitly acknowledged their priority on Euler; OpenAI's paper focuses strictly on the viscous Navier–Stokes equation. 6. **The Podcast's Explanatory Toy Model**: The hosts clarify that the one-dimensional toy model ($5/2 \, \tau^{-3/2}$) used earlier in the episode to illustrate the pulse cancellation strategy was created purely by the podcast hosts for pedagogy. It is not an actual formula from the 166-page paper, and anyone quoting it as such is misattributing the show's heuristic.

---

[15:17](https://www.youtube.com/watch?v=RzJokrUsnbM&t=917s) Final Takeaways and the Unanswered Millennium Problem

In wrapping up the episode, the hosts emphasize the central conceptual insight of the entire proof: the difficulty was never about magically inverting a differential equation, but rather engineering a candidate flow from scratch and "spending 160 pages making the leftover force smooth." The flow shoots fluid out at infinite speed along an ever-narrowing column, all while total kinetic energy remains bounded and actually vanishes in the core.

Yet the core physical intuition of fluid mechanics remains unresolved. Strip away the carefully sculpted external force, leave a smooth parcel of fluid to evolve purely under its own inertia and viscosity, and mathematics still cannot say whether it stays smooth forever or tears itself apart. Alternatives (A) and (B) of the Clay Millennium Prize problem remain just as open and mysterious as they were before this proof was published.

## Notable quotes > [00:01](https://www.youtube.com/watch?v=RzJokrUsnbM&t=1s) a vortex that eats itself. — Theo

> [01:34](https://www.youtube.com/watch?v=RzJokrUsnbM&t=94s) Navier–Stokes is simply Newton's second law ($F = ma$) applied to a microscopic parcel of fluid.

> [01:34](https://www.youtube.com/watch?v=RzJokrUsnbM&t=94s) water in a glass can blow up. — Theo

> [04:19](https://www.youtube.com/watch?v=RzJokrUsnbM&t=259s) By defining $f$ as this exact residual, Navier–Stokes is satisfied by construction.

> [05:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=329s) Because every term carries the exact same power of $ au$, they add together rather than canceling.

> [06:53](https://www.youtube.com/watch?v=RzJokrUsnbM&t=413s) The exponent $h$ provides separate "handles" that allow corrections to target individual divergent components without disrupting the others.

> [08:29](https://www.youtube.com/watch?v=RzJokrUsnbM&t=509s) Just as in the 1D product $u w_x$, zero-mean fluctuations yield a non-zero mean momentum flux.

> [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) Now the part where we say what this does not settle. Boundaries—give me the list.

> [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) Nobody claims this flow arises in nature. It settles the Millennium statement as written, not a physical question.

> [13:24](https://www.youtube.com/watch?v=RzJokrUsnbM&t=804s) The one-dimensional toy model ($5/2 \, \tau^{-3/2}$) used earlier in the episode to illustrate the pulse cancellation strategy was created purely by the podcast hosts for pedagogy.

> [15:17](https://www.youtube.com/watch?v=RzJokrUsnbM&t=917s) spending 160 pages making the leftover force smooth.

> [15:17](https://www.youtube.com/watch?v=RzJokrUsnbM&t=917s) Strip away the carefully sculpted external force, leave a smooth parcel of fluid to evolve purely under its own inertia and viscosity, and mathematics still cannot say whether it stays smooth forever or tears itself apart.

## Mentioned Anthropic, Charles Fefferman, Clay Mathematics Institute, Clay Millennium Prize, Euler equations, Finite Time Blowup for Navier–Stokes, GPT-6 Astra, Lean, Lean 4, Leanim, Levent Alpöge, Loop State, Mara, Navier–Stokes equations, Newton's second law, OpenAI, Theo, Tristan Buckmaster, Xiaol.x

## Video description On 8 September 2026, OpenAI released a 166-page paper, "Finite Time Blowup for Navier–Stokes," claiming to settle alternatives (C) and (D) of the Clay Millennium Prize problem: a smooth fluid, pushed by a smooth force, that reaches infinite speed in finite time while its energy stays bounded.

In this episode of Loop State, Mara and Theo walk through what the paper actually proves and how the proof works, from the equation itself to the last lemma.

We open with a particle render of the counterexample flow: a vortex whose core spirals inward, shoots fluid out along the axis, and collapses into an ever-thinner column as the speed inside climbs without bound.

CHAPTERS 0:00 What does a fluid that breaks Navier–Stokes lo 1:35 What is the Navier–Stokes equation actually sa 2:50 What exactly does the Millennium problem ask? 4:19 How do you build a blow-up instead of searchinhe-force trick) 5:29 Why does the naive self-similar guess fail? (o 6:53 What does the exponent h buy you? 8:29 How do oscillating pulses cancel a force they 10:13 How do you make the error vanish to every ord 11:47 Why is the energy finite while the speed is i 13:24 What does this paper not settle? 15:17 Sign-off

WHAT THIS EPISODE IS CAREFUL ABOUT • The theorem is the FORCED problem (alternatives C (A and B), whether a fluid left alone can blow up, is untouched. • The force is designed, smooth, and compactly suppw arises in nature. • The Lean formalization is public (github.com/openwe did not build it. As of 11 September 2026 there is no external peer review, and the Clay Mathematics Ie result. • Agent counts, hours, tokens, and the model descris. • Alpöge and Buckmaster's priority on the forced Euy OpenAI. • The one-dimensional toy model (f = u_t + u·u_x − he shape of the strategy. It is not in the paper.

SOURCES Paper (PDF): https://cdn.openai.com/pdf/32d9f210-8bier-stokes.pdf OpenAI blog post: https://openai.com/index/navier-s Lean formalization: https://github.com/openai/Navie Clay problem statement (Fefferman): https://www.cla2022/06/navierstokes.pdf

HOW IT WAS MADE Two-host podcast rendered as a Leanim board (Lean 4-hd. Pixel-art avatars, SVG diagrams (redrawn afterthe paper's Figures 1–6), and the tracer render of this episode. The tracer animation advects 24kparticles in the paper's similarity variables with ial stretch is exaggerated for visibility, as in thepaper's own Fig. 1.

#NavierStokes #MillenniumPrize #OpenAI #Mathematicsean4

[Source](https://www.youtube.com/watch?v=RzJokrUsnbM)

中文简报

OpenAI攻克受迫Navier–Stokes方程爆破问题，核心物理情形仍悬而未决 OpenAI的反例通过反向设计一个塌缩涡旋满足了克雷千禧年大奖问题的条件；该涡旋的数学爆破被人为外力抵消，但自然流体是否会发生崩溃的问题依然完全未被触及。 主播们解释了该证明如何通过逆向构建显式爆破速度场来绕过正向数值积分，从而将数学论证的重点转移至证明残差外力在经历奇点时始终保持完全光滑。通过引入各向异性指数 h，该构造打破了共线尺度缩放，并允许振荡波脉冲的迭代循环逐阶抵消奇异残差项，直至外力平滑。由于涡核体积收缩的速度快于峰值速度平方的增长速度，总动能始终保持有限，并且在塌缩涡核内部实际上降至零。最终，这一崩溃过程通过唯一性论证转化为千禧年大奖问题的反例，但主播们强调，该证明仅解决了人为受迫的情形，经典的无外力Navier–Stokes方程问题依然完全悬而未决。 OpenAI发布了一篇长达166页的论文，声称通过证明三维受迫Navier–Stokes方程在有限时间内发生爆破，解决了克雷千禧年大奖问题的备选分支 (C) 和 (D)。 (0:01) 该证明采取“逆向”思路，构建了一个人工设计的发散流动，并将绝大部分分析工作用于确保残差外力函数及其所有导数在穿过奇点时均保持光滑。 (4:19) 尽管流体峰值速度在 t = 1 时达到无穷大，但总动能依然保持有限——且在涡核内部趋近于零——因为奇异区域的体积收缩速度快于速度发散的速度。 (11:47) 该论文并未解决备选分支 (A) 和 (B)：外力 f = 0 的流体是否会产生奇点依然完全悬而未决，且克雷数学研究所尚未认证该研究结果。 (13:24)

Reading summary

In a September 2026 paper, OpenAI researchers constructed an explicit smooth fluid flow that achieves infinite velocity in finite time under an engineered smooth force, targeting alternatives (C) and (D) of the Clay Millennium Prize problem for the Navier–Stokes equations. The Self-Collapsing Vortex and Energy Bookkeeping (0:01) The proposed counterexample models an inward-spiraling vortex that ejects fluid along its vertical axis in opposed jets. As time approaches $t = 1$, the core thins faster than it shortens, driving the fluid velocity to infinity at $t = 1$. Despite the infinite peak speed, the total kinetic energy inside the core does not diverge. The core radius contracts at $r \sim \tau^{1/2}$ while its height contracts at $z \sim \tau^{1/2 - h}$ (with $\tau = 1 - t$), yielding a shrinking volume of $V \sim \tau^{3/2 - h}$. Because peak velocity squared scales as $\tau^{-1 - 2h}$, the localized kinetic energy integral scales as $\tau^{1/2 - 3h}$. Constraining the parameter $h \in (0, 1/100)$ keeps the exponent strictly positive, meaning core kinetic energy actually drops toward zero as the singularity forms. Reverse Engineering Residuals Through Anisotropic Cancellation (4:19) Rather than integrating Navier–Stokes forward in time, the authors reverse the problem by choosing a diverging candidate velocity field $u$ and pressure $p$, then defining the force $f \coloneqq \partial_t u + (u \cdot \nabla)u - \nu \Delta u + \nabla p$. This satisfies the equation by construction, shifting the entire analytical burden onto proving that $f$ and all its derivatives remain smooth through $t = 1$. Standard self-similar profiles fail because differential operators scale identically, preventing divergence terms from canceling. The paper circumvents this lock by introducing the anisotropic parameter $h$, decoupling radial and vertical contraction rates. This separation yields independent scaling exponents that can be targeted by corrective, oscillating wave pulses. Outside the core, these zero-mean pulses generate non-zero average Reynolds stress fluxes. An iterative four-step correction cycle—resolving wave amplitudes, pulse amplitudes, mean flow/pressure, and five radial moment equations—improves residual decay order by order. Summing these corrections with cutoff functions produces a flat residual whose space-time derivatives vanish faster than any polynomial power near the singularity, allowing smooth continuation past $t = 1$. From Explicit Construction to Millennium Contradiction (12:26) To satisfy the Clay Millennium Prize problem formulation for the forced equations in full space (alternative C) and on the periodic torus (alternative D), the localized blowup must rule out any global smooth solution. Outside the active core and its outer annulus, the flow reduces to a pure swirl obeying the radial heat equation, requiring zero external forcing. This allows smooth spatial cutoffs to isolate the collapsing core without disrupting singular dynamics, while standard scaling extends the viscosity $\nu = 1$ construction to any positive viscosity $\nu > 0$. The proof bridges explicit construction and non-existence via Lemma 10.5. By weak-strong uniqueness principles, any hypothetical global smooth solution starting from rest under the same smooth, compactly supported force must coincide identically with the constructed profile before $t = 1$. Because the candidate solution inherits the unavoidable finite-time singularity, no smooth global solution can exist, completing the contradiction required for Theorem 1.1. Verification Limits and Physical Boundaries (13:24) The authors and hosts stress that this mathematical construction does not settle the behavior of unforced fluids. Alternatives (A) and (B) of the Millennium problem—governing whether an unforced fluid starting from smooth data can tear itself apart—remain entirely open. The forcing function in the paper is highly tuned to inject corrective pulses precisely when needed, meaning it does not represent an unforced natural flow. Furthermore, the result carries important institutional qualifications. While OpenAI reported that the 166-page proof was machine-checked in Lean using GPT-6 Astra and generated using roughly 10,000 internal agents over 88 hours (consuming 130 billion output tokens), the paper has not undergone external peer review. As of September 11, 2026, the Clay Mathematics Institute has not certified the proof or recognized the Millennium Prize problem as solved, and OpenAI stated it does not plan to claim the prize money. Priority for the related forced Euler blowup belongs to Levent Alpöge and Tristan Buckmaster. OpenAI's construction demonstrates that carefully engineered, smooth external forcing can drive a viscous fluid to infinite velocity with finite total energy, but it leaves the core physical question of fluid mechanics unresolved: whether an unforced fluid governed solely by inertia and viscosity can ever form a singularity.

Source quotes

[0:01] Theo: a vortex that eats itself. [1:34] Navier–Stokes is simply Newton's second law ($F = ma$) applied to a microscopic parcel of fluid. [1:34] Theo: water in a glass can blow up. [4:19] By defining $f$ as this exact residual, Navier–Stokes is satisfied by construction. [5:29] Because every term carries the exact same power of $ au$, they add together rather than canceling. [6:53] The exponent $h$ provides separate "handles" that allow corrections to target individual divergent components without disrupting the others. [8:29] Just as in the 1D product $u w_x$, zero-mean fluctuations yield a non-zero mean momentum flux. [13:24] Now the part where we say what this does not settle. Boundaries—give me the list. [13:24] Nobody claims this flow arises in nature. It settles the Millennium statement as written, not a physical question. [13:24] The one-dimensional toy model ($5/2 \, \tau^{-3/2}$) used earlier in the episode to illustrate the pulse cancellation strategy was created purely by the podcast hosts for pedagogy. [15:17] spending 160 pages making the leftover force smooth. [15:17] Strip away the carefully sculpted external force, leave a smooth parcel of fluid to evolve purely under its own inertia and viscosity, and mathematics still cannot say whether it stays smooth forever or tears itself apart.

Original transcript

[0:01] Welcome to Loop State. I'm Mara. On [0:04] September 8th, Open AAI posted a 166 [0:08] page paper titled Finite Time Blowup for [0:11] Navier Stokes. Theo, before any [0:15] equations, what does the thing they [0:18] built actually look like? [0:19] >> A vortex that eats itself. Fluid spirals [0:22] inward toward a vertical axis and shoots [0:25] out along it, up on one side, down on [0:28] the other. The core gets thinner and [0:30] thinner, faster than it gets shorter, [0:32] and the speed inside climbs without [0:34] bound. At time one, the speed is [0:36] infinite. [0:38] >> And that's the counter example. [0:41] A smooth fluid pushed by a smooth force [0:45] that reaches infinite speed in finite [0:48] time [0:48] >> with finite energy the whole way. That [0:51] last part is the point. The kinetic [0:53] energy in the core actually goes to zero [0:55] as the core collapses while the peak [0:57] speed goes to infinity. Small region, [1:00] enormous speed, tiny energy. [1:03] >> So the driving question for today is how [1:07] do you build a flow like that and why [1:10] did it take 166 pages? And one caveat up [1:15] front, the theorem is a theorem and [1:17] OpenAI reports it is formalized in lean, [1:20] but the Clay Institute has not certified [1:22] anything and as of today there is no [1:24] external peer review. We'll separate [1:26] what the paper proves from what Open AI [1:29] reports. [1:30] >> Let's start with what the equation even [1:33] says. [1:34] Theo, for someone who remembers high [1:37] school physics and nothing after, what [1:39] is Navier Stokes? Newton's second law [1:43] for a small parcel of fluid. Force [1:45] equals mass time acceleration divided by [1:48] mass. The acceleration of the parcel is [1:51] the time derivative of velocity plus the [1:53] transport term u do nabla u because the [1:57] parcel is moving while the flow changes [2:00] >> and the forces [2:03] >> three the pressure gradient minus nabla [2:06] pushing fluid from high to low pressure. [2:08] Viscosity new * the lelassian of U [2:11] friction between neighboring layers that [2:13] smooth sharp gradients and whatever [2:15] external force you apply called F plus [2:19] one constraint [2:20] >> divergence of U equals zero [2:22] incompressible. You cannot squeeze fluid [2:25] in one direction without stretching it [2:26] in another. That constraint makes the [2:28] pressure a global enforcer rather than a [2:31] local force. [2:32] >> And the hard part is the transport term. [2:37] The transport term is quadratic in U. [2:40] Everything else is linear. Viscosity [2:42] smooths. Transport sharpens. And the [2:44] whole 90-year question is whether [2:46] smoothing always wins in three [2:48] dimensions. [2:50] >> The clay millennium problem is not one [2:52] statement, it's four. What are they? [2:55] >> Fafferman's statement gives four [2:56] alternatives and you win by proving any [2:59] one. A and B say with zero external [3:02] force smooth initial data always gives a [3:06] smooth solution forever on all of space [3:08] or on a periodic box. That's the [3:11] regularity direction. [3:13] >> And C and D [3:16] >> C and D are the breakdown direction. [3:19] There exist a smooth initial velocity [3:22] and a smooth external force such that no [3:25] smooth solution with bounded energy [3:27] exists for all time. C is on all of [3:30] space. D is on the periodic Taurus. [3:33] >> And the paper claims which [3:37] >> theorem 1.1 establishes C. The fluid [3:40] starts at rest. The force is smooth and [3:42] compactly supported in space and time. [3:45] The energy stays bounded and the peak [3:47] speed goes to infinity as t approaches [3:49] 1. Because everything is compactly [3:51] supported, corollary 10.6 wraps it onto [3:55] the Taurus, which gives d. So it says [3:58] nothing about A and B. The unforced [4:01] problem is still open. [4:03] >> Correct. This is a designed force [4:05] producing a designed breakdown. It [4:08] settles the millennium statement as [4:10] written because the statement allows a [4:12] force. It does not say that water in a [4:14] glass can blow up. [4:16] >> Then how do you design a breakdown? [4:19] Here's the trick that made this click [4:21] for me. You don't solve the equation. [4:25] You run it backwards. [4:27] Exactly. Take any smooth velocity field [4:30] U and pressure P you like. Compute the [4:32] residual time derivative plus transport [4:36] minus viscosity plus pressure gradient. [4:39] Call that residual F. Then Navier Stokes [4:42] holds by construction with F as the [4:45] external force. [4:46] >> Um so the equation is never in danger. [4:50] The whole difficulty moves into F. The [4:53] whole difficulty is choose u so that the [4:56] speed blows up at time 1 while f and [4:59] every derivative of f stays smooth [5:02] through time 1. Individual terms in the [5:05] residual can diverge. Their sum must [5:07] not. [5:08] >> That's a cancellation problem not a [5:10] solving problem. [5:11] >> The paper says exactly that. Section two [5:14] first paragraph. The challenge is to [5:16] choose a flow that blows up while the [5:18] residual remains smooth. Everything [5:20] after that is engineering the [5:22] cancellation. [5:23] >> Let me try the dumbest possible flow and [5:27] watch it fail. Our cartoon, not the [5:30] papers. One space dimension, viscosity [5:34] one, no pressure. The force is u subt + [5:38] u * u sub x - u sub x. [5:42] And um I'll guess a Gaussian bump that [5:45] sharpens as tao goes to zero where tao [5:48] is 1 - t. The guess u0 equ= to the -2* [5:56] exponential of minus x^2 over tow. [5:58] Height grows like toao to the - 1/2. [6:01] Width shrinks like the roo of at [6:04] time 1. The peak is infinite. And the [6:08] force at the center [6:10] >> at xals 0 the transport term vanishes by [6:13] symmetry. The time derivative gives 1/2 [6:16] tow theus3 halves. The viscous term [6:19] gives 2 toa to theus3 halves. Total 5 [6:24] tow theus3s. It blows up. [6:28] >> So the velocity blew up and the force [6:32] blew up right along with it. Every term [6:35] carried the same power of toao, so they [6:37] add instead of cancelling. And you can [6:39] swap the Gaussian for any profile you [6:41] like. With one time scale, the terms [6:43] stay colinear and nothing can cancel. [6:46] You need a second scale, [6:48] >> which is where the paper's little [6:50] exponent h comes in. The paper has this [6:54] number h between 0 and 1 over 100. Why [6:58] does a number that small matter? [7:00] >> Because it breaks the tie. In the paper, [7:04] the core's radius shrinks like toao to [7:06] the 1/2, but its height shrinks like [7:08] toao to the 1/2 minus h. Two different [7:11] scales. The core becomes an increasingly [7:14] slender column. [7:16] >> And the velocities [7:19] >> swirl and axial speed grow like toao to [7:21] the minus one/2 minus h. Radial inflow [7:25] is slower. Order towa to the minus [7:27] one/2. So the swirl Reynolds number goes [7:30] to infinity while the radial Reynolds [7:32] number stays bounded. Fluid makes ever [7:35] more turns per diffusion time. But [7:37] viscosity still competes with the [7:39] inflow. In our cartoon that's the [7:42] anisotropic guess to the minus a * a [7:45] Gaussian in x where x is x over t to the [7:50] b and a and b are plus or minus h over [7:55] 3. And now the force at the center has [7:58] two terms with different powers of toao. [8:00] A tow the minus a minus1 plus 2 to the [8:04] minus a - 2 b. The exponents differ by [8:07] 2/3 h. You can no longer merge them, [8:10] >> which sounds worse. It still diverges. [8:15] >> It still diverges, but the terms are now [8:18] distinguishable. So a correction can [8:20] target one of them without touching the [8:22] other. That's what H buys you. separate [8:25] handles. [8:26] >> So what pulls the handles [8:29] cartoon? First I add a small odd pulse [8:33] to the velocity minus c of tow * toao to [8:38] the minus a * x * the gaussian. It's [8:41] zero at the center. How can it change [8:43] the force at the center? Through the [8:46] transport term. The poles has a slope at [8:48] the center and the background has a [8:50] value there. So their product u * w subx [8:54] is non zero at x= 0. Choose c of toao [8:58] equal to a + 2 to the 2/3 h and the [9:01] force at the center is identically zero [9:04] >> zero at one point everywhere else [9:08] >> still diverging. So you add another [9:10] pulse for the next spot and another. [9:13] That's the cartoon of the paper's real [9:15] mechanism [9:16] >> which is [9:17] >> in the paper the background vortex [9:20] leaves a residual only in an annulus [9:22] around the core and it is exactly minus [9:25] the divergence of a stress. The pulses [9:27] are oscillating waves whose velocity [9:29] averages to zero around the ring but [9:32] whose momentum flux does not. Outward [9:34] motion with a swirl surplus and inward [9:36] motion with a deficit both push angular [9:39] momentum outward. zero mean nonzero [9:42] product like our cartoons u * w subx [9:47] >> same idea one level up two-way families [9:50] give two flux directions and proposition [9:52] 7.5 puts the required stress strictly [9:55] inside the cone they span so both [9:57] weights are positive each pulse is [9:59] seated by an exponentially small force [10:02] amplified by the shear then killed by [10:04] viscosity as the shear shortens its [10:06] wavelength [10:08] >> um but that only cancels the leading [10:11] residual [10:13] after the pulses. What's left? [10:15] >> Everything the pulse has created. Adding [10:18] a correction W changes the residual by a [10:20] linear term, the pulse evolving in the [10:22] background and a quadratic term, the [10:25] pulse interacting with itself. The paper [10:27] writes that identity exactly and then [10:29] runs a fixed cycle. [10:31] >> What's in the cycle? [10:33] >> Four operations. Solve the wave [10:35] amplitude equations for the leftover [10:36] harmonics. Adjust the pulse amplitudes [10:39] so their cross covariance fixes the [10:41] average stress. Correct the angular mean [10:44] flow and pressure. Solve five radial [10:47] moment equations so everything stays [10:49] compactly supported and incompressible. [10:52] >> And each pass bys what? [10:57] >> 1/10th. The residual decays like Q to [10:59] the H sigma and sigma starts at 1/5 and [11:02] gains 1/10th per cycle. Sigma goes to [11:04] infinity. Then you sum all the [11:06] corrections with cutoffs that shrink [11:08] toward the singular point. [11:11] >> And the sum has a residual that is what [11:15] did the paper call it? Flat. [11:17] >> Flat. Every space-time derivative of the [11:20] residual is bounded by a constant time q [11:22] to the n for every n as you approach the [11:25] singularity. That residual is the force. [11:28] Flat means it extends smoothly through [11:31] time 1. Lema 10.3 realizes the extension [11:35] with a force supported up to time two. [11:38] >> So the force is smooth, the flow blows [11:42] up. And now I want to know why the [11:44] energy doesn't blow up too. [11:47] Infinite speed, finite energy. That [11:50] sounds like a contradiction. [11:52] >> It's a volume argument. The core has [11:54] radius to the 1/2 and height to the [11:56] 1/2us h. So its volume is to 3 - h. [12:01] Speed squar is to the -1 - 2 h multiply [12:04] energy of order toao to the 12 - 3h [12:08] >> which goes to zero because h is tiny [12:13] >> because h is less than 16. The same [12:16] bookkeeping makes the integrated [12:18] dissipation finite. Lema 10.4 four [12:21] proves both bounds for the localized [12:23] fields directly from the equation [12:26] >> and outside the core [12:29] >> outside the annulus the flow is a pure [12:32] swirl that solves the radial heat [12:34] equation exactly so it needs no force at [12:37] all that lets them cut everything off in [12:39] space with smooth cut offs keeping the [12:42] collapsing core untouched and a scaling [12:45] turns the viscosity one solution into [12:47] one for every positive viscosity One [12:50] more step. The theorem says no smooth [12:55] bounded energy solution exists. The [12:58] construction only gives one particular [13:01] solution. [13:02] >> Lema 10.5. Any smooth solution with the [13:05] same force and zero initial data and [13:08] bounded energy must agree with the [13:10] constructed one before time 1. So it [13:13] inherits the blowup and cannot be [13:15] global. That is the contradiction that [13:17] proves theorem 1.1. Now the part where [13:20] we say what this does not settle [13:23] boundaries give me the list [13:26] >> one it is the forced problem [13:29] alternatives C and D the unforced [13:32] problem A and B whether smooth data with [13:34] no force can blow up is untouched [13:38] >> um two the force is designed [13:42] >> designed smooth compactly supported and [13:46] tuned so the pulses land where they are [13:48] needed [13:49] Nobody claims this flow arises in [13:51] nature. It settles the millennium [13:53] statement as written, not a physical [13:55] question. [13:56] >> Three, verification. [13:58] >> Open AAI reports a lean formalization [14:01] checked in about 17 hours with GPT6 [14:04] Astra. That is Open AI's own report. As [14:08] of September 11th, there is no external [14:10] peer review and the Clay Institute has [14:12] said it is not treating the problem as [14:14] solved. Um for the process claims [14:20] >> also reported not proved an internal [14:23] model more capable than GPT6 Astra [14:26] roughly 10,000 concurrent agents 88 [14:29] hours to the result about 130 billion [14:32] output tokens. Open AAI says it does not [14:35] intend to claim the prize [14:37] >> and five the priority story. Two [14:41] mathematicians, Levent Alpaji at [14:44] Enthropic and Tristan Buckmaster at NYU, [14:47] independently resolved the forced oiler [14:50] problem with an anthropic model. Open [14:52] AAI recognizes their priority there. [14:54] Navier Stokes with viscosity is the [14:57] paper we discussed [14:59] >> and six our cartoon. [15:03] >> The one-dimensional toy is ours. It [15:05] shows the shape of the strategy, not the [15:08] paper's estimates. Anyone quoting five [15:10] halves tow to the minus3 halves as the [15:13] paper's result is quoting us, not them. [15:17] >> Takeaway: The trick isn't solving an [15:19] equation. It's choosing a flow and then [15:21] spending 160 pages making the leftover [15:24] force smooth. Infinite speed on a [15:27] shrinking column with energy that [15:29] vanishes as it does. [15:32] >> Open question. Take away the force. Does [15:34] a smooth fluid left alone ever do this [15:37] on its own? That's alternatives A and B. [15:40] And after this week, they are still [15:41] exactly as open as they were.