AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse

The episode opens on an absence. A guest was booked and pulled out that morning, and the hosts spend their first minute diagnosing why. Sacks, who has a law degree he has not used in twenty-five years, says he sat up all night researching an interview that will not happen. The joke sets the register: this will be argued as a case, not reported as a story.

The facts as presented are narrow. Jacob Coxen worked as a researcher at OpenAI and then at Anthropic over roughly three years, and had been at Anthropic about six weeks when he resigned. That figure becomes the spine of the episode's skepticism, returned to at least four times with the same implication: too short a tenure to have seen anything, long enough to have planned something.

What makes the segment worth reading closely is not the news, which is thin, but the argumentative machinery the hosts build around it. Four separate claims get braided together — that the risk estimate is unfalsifiable, that the amplification was coordinated, that Anthropic cannot afford to disavow the claim, and that the whole affair is a lever for regulatory capture. Each is individually plausible. The interesting question is whether they add up to a refutation, or merely to a motive.

The resignation post itself is read out in two fragments. The first is a claim about belief rather than fact:

the people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.

The second is an accusation about conduct, aimed at both labs: that neither is acting responsibly, and that they are, in the words of the post,

racing straight to self improving superintelligence and gambling. Gambling with our lives.

Note what kind of statement this is. It does not allege a specific capability, a specific test result, or a specific decision. It alleges a disposition — that the people with the most information privately hold a view they will not defend in public. That makes it unusually hard to falsify, and it is the reason the rest of the conversation keeps sliding between two different questions: is the claim true, and is the claim evidence-backed.

The reach numbers are then used as an argument in themselves. The post is compared to a tweet that supposedly holds the all-time record, and the hosts press on the mechanism: how does an account with no audience reach a hundred million views in a day? The question is treated as rhetorical, but it is an empirical question about recommendation systems that nobody attempts to answer.

The most consequential line in the segment is not Coxen's. It is his former boss's. Evan Hubinger, who leads alignment science at Anthropic, replied that the resignation was substantively correct, and then attached a number:

I personally think it is over ten percent within the next decade. We do not yet have a plan to solve alignment.

That single reply converts a departing employee's opinion into something closer to a corporate position, and the hosts understand this immediately. It is also, as they point out, the moment the story escapes the tech press: within a day it is being quoted by senators and governors.

Sacks's opening move is to deny novelty. There is, he says, nothing here that the same crowd has not been saying for years, and the media's use of the word whistleblower is unearned. The challenge is put in the imperative:

This is all just vibes from him. Show us the data. Show us the reports. Show us the leaked information that the public didn't already have. Show us the facts.

It is a fair challenge in form, and it lands because the resignation post really does contain no documents. But the demand has a blind spot that the hosts never examine. A researcher inside a frontier lab is bound by confidentiality agreements, and the specific artifacts that would settle the question — evaluation transcripts, internal risk assessments, the reasoning behind a deployment decision — are exactly the artifacts he cannot lawfully publish. The only admissible evidence in Sacks's sense may be evidence that cannot exist.

The same asymmetry appears in the whistleblower framing. Sacks asks what has been blown open, and the answer is a belief, not a document. That is a real weakness in the resignation's evidentiary posture — and arguably the point of the post, since a claim that a belief is widespread and undisclosed is precisely the kind that produces no paper trail.

What the episode does not do is ask the symmetric question. If the resignation is unsupported assertion, so is the counter-claim that it was orchestrated. Both rest on inference from timing and affiliation. Only one of them is subjected to the show-me standard.

The evidentiary core of the psyop thesis is a timeline. The account, the hosts say, had almost no activity, almost no followers, and no posts before the resignation thread; if there were earlier posts, they were scrubbed. Then, within a day, it found a very large audience. That combination is offered as the anomaly that requires explanation.

The explanation offered is a set of three organizations and one funder. Nathan Calvin at Encode AI, which pushed California's SB 53 and a family of state AI laws. Peter Wildeford at the AI Policy Network, which pushes federal AI regulation and a federal AI department. Daniel Kokotajlo at the AI Futures Project, who released a podcast episode the same week using what the hosts describe as identical phrasing. All three, they say, are funded by Jaan Tallinn, described here as both an effective-altruism mega-donor and a co-lead of Anthropic's Series A.

Then the detail that carries the most weight: the Wall Street Journal story covering the resignation appeared minutes before the tweet thread. The hosts read this as proof of an embargo that was mishandled —

they were briefed under embargo

— and therefore as proof that the resignation was a staged media event rather than a spontaneous act. Later in the hour the claim is restated in weaker, more defensible form: the groups were alerted and probably participated, if not in creating the thread then certainly in amplifying it.

The inference is reasonable and incomplete. Embargoed briefings are how journalism works, and a source shopping a resignation letter to a major paper is ordinary. Fifteen-minute latency suggests coordination, but it is also what you get when a small, well-networked community reads the same feeds. The claim that such reach has never happened before is asserted on Elon Musk's authority and no data is shown. The psyop case is a correlation story told with more confidence than correlation supports.

The most revealing passage is almost incidental. Coxen had asked to come on the show, then canceled the morning of the recording. Sacks reads the cancellation as evidence of motive — he couldn't take the smoke — and lists the questions he would have asked. The first matters:

Who connected you with the Wall Street Journal? The answer would discriminate between a researcher who happened to be newsworthy and one who was placed. But it is a question a hostile interview would never get answered, and its unavailability is used as though it were an answer.

From there the tenure argument is deployed again. A relatively junior employee of three months, at most, would not ordinarily receive this kind of coverage; therefore the coverage was arranged. The reasoning is a syllogism with a missing premise — that the press only covers seniority — and it is exactly the premise the story falsifies, since the coverage exists.

One observation in this stretch survives the episode's own biases. In a normal whistleblower story the company pushes back and the press adjudicates; here the company's leadership amplified the claim. That inversion is genuinely unusual, and it is the strongest factual basis for suspecting the resignation was an act of alignment rather than opposition.

The pivot of the whole segment is the observation that Anthropic did not disavow the resignation. Chamath puts it as a question about corporate function — would a functioning company not immediately distance itself? — and Sacks answers with the word contradiction.

The mechanism he describes is that the safety lead did not merely tolerate the post but extended it, attaching a probability to a claim that had previously been qualitative. That move, the hosts argue, is what made the story travel: a vibe is arguable, a number is quotable.

This is the episode's best piece of analysis, and it is worth separating from the conspiracy framing around it. The legally salient fact is not that a six-week employee resigned. It is that a current senior executive endorsed the resignation and quantified it. Whatever one thinks of the underlying risk estimate, that endorsement is a corporate act, and it is the thing that later makes the disclosure question unavoidable.

The counter-reading is available and unaddressed: a safety leader might endorse a departing colleague's concern precisely because he believes it, at personal and institutional cost, which is what integrity looks like rather than what coordination looks like. The hosts never test that hypothesis, because their model of the company already excludes it.

Chamath asks for the steelman, and frames it as game theory: if the thing had not been exposed, how was it supposed to go? Sacks's answer is that the participants knew exactly what would happen, and that maximum unification was the goal. He then floats the stronger version — that a three-to-four-month employee may have joined already intending to run this campaign.

The stated endgame is institutional. These groups, Sacks argues, shape public perception, and the perception they want is negative enough to justify a new regulator. The ask is a federal department of AI — an FDA for AI — which he says Dario Amodei has himself called for. Once such a body exists, its advocates write the rules it enforces.