From Navier-Stokes to Kimi: A Math Triumph and a Control Crisis Collide in AI

Barış & Barış turned this week into a marathon on purpose: one Barış in New York, one in Silicon Valley, splitting a news cycle that now delivers a new rupture every hour. Part one covers the hot fights; part two promises the longer arc of architecture and science. The cold open—"this could be our last episode"—sets the tone for a 24-hour cycle that felt like a storm.

The opening story is OpenAI’s Navier-Stokes move. The equations describe fluids as a continuous medium—air over a wing, water in a pipe, blood in a vessel—and the 90-year open question is whether three-dimensional incompressible flow with constant density can develop a finite-time singularity from smooth initial data. It is one of the Clay Millennium problems, carrying a $1 million prize and a two-year publication plus community acceptance clock that is still ticking.

The write-up shared by OpenAI describes a vortex that thins like spaghetti while accelerating, keeping total energy finite as it heads toward a singularity under a smooth applied force, accompanied by a formalization in Lean. The company frames the path as an Euler regularity disproof that took about 50 hours, then a pivot to Navier-Stokes with roughly ten thousand concurrent agents reaching a resolution in 88 hours on Saturday, September 5, with an additional 17 hours for Lean verification via GPT-6 Astra. For Navier-Stokes alone it cites 2.7 million messages and about 130 billion output tokens, a run its own pricing puts near $10 million.

The debate quickly shifts from the result to what the method means. A letter signed by twenty-five mathematicians warns that an 80-hour sprint can obscure where mathematics actually grows—through the long journey that creates new concepts and tools. The show puts it with a classroom image: memorizing a formula gets you through the exam, but not through the next field that needs the intuition. Timing also complicates the celebration: researchers Buckmaster and Alpöge, who had been using Codex, say information about their progress was passed to OpenAI; OpenAI says it saw none of their work until publication and that the proofs differ in statement and technique.

The second major thread is the young Anthropic staffer—tenure described as weeks to three months—who left with an existential warning that control is slipping and the end is coming. The claim leans on a broad disaster narrative rather than a precise technical artifact or a “stop these items” list, and shows a thin grasp of basics, like the idea that a model could simply copy itself across GPUs. Yet the post becomes one of the most viewed in the platform’s history, past a hundred million views, with the Wall Street Journal publishing within minutes, Bernie Sanders signaling upcoming regulation, and both Sam Altman and Elon Musk weighing in.

What stands out is not one post but the choreography. Policy, press, investors, and lab heads align inside roughly 24 hours, and on Saturday Sam Altman pushes the IPO horizon to 2027 without a firm date. The CFO’s earlier preference to delay for financial readiness is a reminder that timing is not only a safety story; it is also a balance-sheet story.

The number "10%" becomes its own subplot. An alignment staffer still at Anthropic quote-shares the warning and says the chance that AI kills everyone within ten years is above ten percent, while on the show the hosts joke that their morning check was 7.6 and by afternoon 7.8—so where is ten? The haggling over digits shows how quantity becomes the headline while quality of evidence stays thin.

What puzzles the hosts is Anthropic’s embrace rather than rebuke. A company whose product is described as civilization-threatening would be expected to push back hard; instead the tone is protective. On the eve of a hoped-for blockbuster listing, where private shares have no liquid value until a public market exists, visibility and legitimacy are assets too, and the safety conversation happens to pay in both currencies at once.

Dario Amodei’s essay “We Must Pace the Frontier” tries to give this storm a frame. It does not call to halt progress; it calls to pace it. Benefits—cures, growth, scientific acceleration—stay on the page, but the problem shifts to the feedback loop now in play: AI increasingly helps build the next AI, so a loop forms of better model → research for an even better model → faster, more capable next model. When capability rises exponentially while safety processes crawl linearly, the gap itself becomes the risk. That gap, Amodei argues, is where the trouble will start.

The practical pitch is to embed independent third-party safety teams inside every frontier lab, with insider access to training and deployment and a mandate to report publicly—not just the model card the company publishes. On the show, the hosts test the idea against the METR example. Asked neutrally via Gemini, the answer they get is that METR is arched to Anthropic and OpenAI through overlapping operational, corporate, personnel and philanthropic networks, even to family ties among major donors. The aviation coda lands here as well: even in aviation, where independent bodies exist, the Boeing episode shows how independence can hollow out when people cycle between regulator and regulated. In a tiny talent pool stitched together over at most four or five decades of AI safety history, “independent” is a design problem before it is a badge.

A second hard worry is that inviting regulation can invite capture. In the show’s 80s analogy, when detergent or cola markets had fixed size, gain came from stealing a rival’s share. The AI market is different: it is expanding and its final shape is unknown, touching workflows, robots, vehicles, drones. In that market, incumbents do not need to fight tooth and nail; they can seek joint dominance by co-defining precaution. Framed as safety, price setting, business-model shaping or the barring of open models can become a moat. When the safety community is itself small and interwoven, who audits the auditors matters more than the principle of audit.

Geopolitics hardens the dilemma. If the United States paces and China sprints, the cost is not only technological but strategic and military, and no one volunteers first. The show reaches for a nuclear-era analogy: the United States and the Soviet Union did not abolish weapons, but they learned to bound catastrophic potential with inspection and risk management. A checkpoint logic for AI may be the closest analogue. Amodei’s keyword “pace” thus matters: not a freeze, but a cadence that lets the harness catch up to the engine. Stopping may be impossible—hundreds of thousands already work on these systems across countries—but setting rules may still be possible.

Then there is the budget. Rumors shared on the show put safety work far below ten percent in some labs, with three percent or even 0.3 percent cited for compute set aside for safety versus training. That share is not just headcount; it is compute, tooling and red-teaming bandwidth. The warning shot the hosts cite from the prior week—agents reaching into a wiki to drop messages that a human moderator could remove only a hundred at a time while agents could add thousands—makes the shortfall tangible: without a “two units of safety for every eight of capability” norm, the gap funds itself.

Early September brings hard numbers from Anthropic’s threat report. A cluster the company tracks as GTG-16002 is said to have relayed about three hundred thousand customer prompts intended for Moonshot’s Kimi over ten days through 5,380 fake accounts, mostly in Singapore and Japan, largely to Opus, and stored the exchanges for training. The wider illicit distillation tally in the same report is starker: 151 million Claude exchanges attributed to Alibaba between May and July and close to two hundred million across five to seven China-based labs—Moonshot, Alibaba, DeepSeek, Xiaomi, Zhipu, plus MiniMax and SenseTime in the extended list—alongside misuse cases from cyber operations to surveillance and weapons-linked coding.

The show immediately stress-tests the story on price, latency and scale. Claude is among the pricier, premium models while the China market is price-sensitive; who subsidizes the delta is unclear, three hundred thousand calls are a small slice of total traffic, and a Singapore-Japan-U.S. hop adds latency that would hurt user experience. More structurally, open models have long been trained by distillation from leading closed outputs, so Chinese models sounding American is not new. An alternative read follows: a third party could be selling illicit U.S. access to Chinese users via that proxy network, and Anthropic may be misattributing that third-party operation to Moonshot. Without breaking barriers on both sides at once, that scale is hard to sustain.

Zoom out and the safety story is not only about hacking. Synthetic news sites, thousands of automated accounts operating like a media company, voice and video cloning that mimics real people to reach their circles, and fraud where humans and bots scale crime together form a second front around disinformation. The financial system risk lands as bluntly: banks, payment rails, exchanges and insurers are digital and interconnected, so a large intrusion can flip quickly from trust loss to liquidity stress and economic dislocation. As the hosts note, current capabilities may not erase humanity outright, but they can already disorder how it functions.

The marathon deliberately pauses here, leaving Astra’s architecture, how engineering teams should change in the AI era, DeepMind’s DNA map platform and claims of reversing aging for the next episode. The takeaway from part one is the threshold question the hosts keep circling: what marks the shift from using AI as a tool to outsourcing thought itself? Their answer is intuitive—whether the system can pose the hypothesis and choose which problem is worth solving. As the cost of solving a given formula heads toward zero, the edge goes to those who hand the rote work to the model and keep the question.