The Fight Over Open Source AI, Anthropic's $1.5B Payout, NYC Socialists: Evictions = Violence?

The episode opens on one benchmark result and the political reaction it set off. Moonshot AI's Kimi K3 is described as roughly on par with Opus 4.8 and GPT 5.6 — not six months behind, not twelve, but level — and at about half the price. The panel frames the response as a second DeepSeek moment, and the evidence for that framing is institutional rather than technical: the story reaches the White House within days. The administration's posture arrives in fragments. Axios reported on a Monday that the White House was weighing a ban on Chinese open-source models; Wired reported on Wednesday that Commerce Secretary Howard Lutnick did not want one. Michael Kratsios, introduced as a friend of the show, is quoted saying there is information that Moonshot distilled for the development of K3. Polymarket's contract on a US ban of an open-source model in 2026 moved from 22% to 45% in a matter of days, which is treated as a real-time measure of elite opinion rather than of policy.

Sacks reads the sequence as a campaign rather than a debate. His evidence is rhetorical: Anthropic's blog posts, he argues, pair every threat description with an aside about guardrails being removable once a model is open, and the aside is doing work. "Any threat that they describe, they kind of go out of their way to take that shot at open source models." The mechanism he names is the placement of predicate facts in the public record so that a later action has its justification already on file; the ban then becomes a question of timing rather than of principle. Note what this argument does and does not claim. It does not allege a drafted rule. It alleges a preparatory narrative, which is a weaker claim and a harder one to falsify — a point the panel never quite addresses.

Sacks states, on what he calls good authority, that the White House has not decided to ban open-source models and wants that known. He separates the live question — what to do about Chinese distillation — from the ban, and describes the president as listening to a chorus of voices before deciding, with the implication that the chorus is still being assembled. His own position is unambiguous: action against the open-source ecosystem would be "a tragic mistake" that does nothing but hurt America's position, because "you cannot punish American developers for it." The framing matters, because it converts the dispute from a security question into a question about who bears the cost of a private company's enforcement failure.

The commercial argument follows from the numbers. Anthropic began the year at $10 billion of ARR and is now above $70 billion — on Sacks's reading, the fastest-growing company at scale in history, and therefore not one that needs protection from competitors Chinese or otherwise. He calls the lobbying regulatory capture. Then comes the structural tell, which is the most load-bearing sentence in the segment: if stopping distillation were the primary objective, Anthropic would push to block Chinese access to American models, because that is where distillation actually occurs and "you have to stop it at the source." Instead the proposal runs the other direction, restricting American access to Chinese models, which leaves the rest of the world using them and the distillation channel open.

Chamath supplies the mechanism before the argument: prompt a model, observe the output, train your own model on it, and multiply by tens of millions of queries until what has been moved is effectively trillions of question-answer pairs. If that were the concern, the remedy is identity verification at signup — a real account, some form of identification, a bounded credit card. Those steps would slow revenue traction, which is precisely why they are not taken. Sacks adds the operational picture: accounts created by students, aggregated and sold through dark-web channels, registered from Manila and India against American IP addresses. His inference is that scale makes the activity more visible, not less, so the absence of enforcement reads as preference rather than incapacity.

The reciprocity problem is the segment's sharpest turn. Everyone has distilled: Anthropic paid a $1.5 billion settlement to publishers, OpenAI is being sued by the New York Times, and Chinese labs distilled from Anthropic. Chamath's image is nine Spider-Men pointing at each other, and it is doing real analytical work — it shows that the accusation, applied consistently, incriminates the accuser. Friedberg declines the word stealing, on the ground that ownership of the underlying copyright is unsettled, and reframes the issue as one of speed: once a model publishes its performance criteria, rivals match it within weeks, which is a statement about commoditization rather than about theft.

One asymmetry goes unexamined. Sacks puts the burden of explanation on Anthropic — why is enforcement so poor, and why should the remedy fall on the public domain rather than on a company with 90% gross margins — while Friedberg puts it on the legal system, asking what actually counts as infringement. These are answers to different questions, and the panel treats them as converging because they generate the same policy conclusion. The convergence is real. The shared premise underneath it — that the present arrangement is stable if nobody intervenes — is the part that never gets tested, and it is the premise a policymaker would most want stress-tested.