Jacob Coxon, Anthropic, and the AI Doomer Debate

A Twitter account with no followers and no posting history published a resignation letter on a Monday. By Tuesday it had 110 million views (01:33). The author, Jacob Coxon, had worked at Anthropic for six weeks before quitting. His message was simple: the people building artificial intelligence believe it could kill everyone by the end of the decade, and this is not a marketing stunt (01:04). Evan Hubinger, who leads alignment science at Anthropic, replied within the company's own comment section to say Coxon was correct, and that he personally put the odds of AI-driven human extinction above 10 percent within ten years (01:56). Bernie Sanders quote-tweeted it and announced legislation to ban superintelligence outright (02:24).

The hosts of this show do not buy the spontaneity of it. They point out that the account had been scrubbed of any prior activity, that three well-funded advocacy groups amplified the post within fifteen minutes, and that all three trace back to a single donor, Jan Tallinn, who also co-led Anthropic's Series A funding round (04:36). A Wall Street Journal story on the resignation appears to have gone live under embargo minutes before the tweet storm itself (05:02). None of this proves coordination. But it is, at minimum, a strange way for a spontaneous whistleblower moment to unfold.

The Mother of All Product Liability

David Sacks, known on the show as our former czar of AI, frames the deeper problem as a contradiction Anthropic cannot escape. The company is trying to go public at a valuation in the trillions of dollars while its own safety lead is telling the world its core product might end civilization.

"You're either on the Bernie Sanders side of this thing, in which case why are you IPO-ing? Or you renounce the argument." - David Sacks [13:30]

Investor Chamath Palihapitiya adds the mechanical detail: in the quiet period before an IPO, a company's S-1 filing is supposed to be an accurate snapshot of every risk it faces (47:36). If employees are publicly asserting a double-digit chance of extinction, that is not a boilerplate risk factor about chip supply or competition. It is, in David Sacks's words, the mother of all product liability lawsuits (11:14). Sacks argues Anthropic now has to choose: disavow the claim as hyperbole, or accept that its own logic points toward the Bernie Sanders position of freezing development altogether.

Friedberg supplies the historical pattern that makes the panel skeptical of the doomer framing generally. Al Gore's documentary An Inconvenient Truth leaned on IPCC forecasts that have since proven overstated (15:45). Public health officials predicted COVID would require total lockdown, with consequences that included the inflation and debt spiral that followed (16:12). After the Three Mile Island accident, the United States effectively stopped building nuclear plants, and the price shows it: producing one gigawatt of nuclear power costs about $15 billion in the US today, versus roughly $1 billion in China (16:36). Each case, Friedberg argues, followed the same shape: an unprovable existential claim, amplified until it became the default social assumption, used to justify centralizing control in the hands of whoever raised the alarm.

"The US will be the tribe in the Amazon that has never seen civilization. 'Cause civilization will evolve all around us." - David Friedberg [20:36]

The stakes of that argument, in this telling, are not abstract. If the United States pauses AI development, other countries will not, and the pause simply hands the frontier to someone else. Sacks connects this to what he calls the actual endgame of the doomer coalition: not safety for its own sake, but a Federal Department of AI, a new regulator that would police both closed labs and open-source models alike (07:50). Open-source weights, once published, cannot be recalled or centrally monitored, which is precisely why he thinks they are the real target (25:43). Open models also cut the cost of running AI by roughly 50 times, which is the opposite of what a company hoping to control a duopoly would want widely available (23:49).

To their credit, the panel does not simply dismiss the underlying fear. They spend fifteen minutes trying to steelman it, trading scenarios for how AI could actually kill everyone, from Skynet seizing nuclear command systems to airborne bioweapons brewed in hacked bioreactors. Every scenario runs into the same wall: a human still has to physically do something, and that human-in-the-loop step is exactly where today's systems fail. As Friedberg puts it, current AI assistants still cannot reliably book a hotel room without a person finishing the job.

"We solved one of the seven hardest math problems in the world, and you still can't get your intimacy kit from the front desk." - Jason Calacanis [62:35]

What AI Actually Did This Week

That joke is doing real work, because the same week produced a genuine technical result. OpenAI published a solution to a two-hundred-year-old fluid dynamics problem tied to the Navier-Stokes equations, the mathematics that describe how air and water move, underlying every aircraft and weather model in use today (58:40). The company reported using roughly 130 billion output tokens across 10,000 AI agents working in parallel to get there (59:38). Friedberg's read is important: this was not a flash of insight no human could have had. It was brute force, the equivalent of somewhere between 50,000 and 500,000 years of human labor compressed into a documented, readable exchange between machines (60:10). AI, in his framing, is an engine of leverage, not a new form of genius (61:26).

That same episode raised a quieter but more practical worry: whether the mathematicians who had been close to solving the problem had unknowingly fed their approach into the model that then beat them to it. OpenAI's own statement conceded it could not rule out that de-identified data from their usage helped improve its models (63:04). This is where Chamath's argument about Zero Data Retention, or ZDR, lands. ZDR is the promise that a company's prompts will not be retained. Chamath calls it a best-efforts, commercially framed promise, not a guarantee, since something as small as a user clicking a thumbs-up on a chat can pull that exchange back into the training pipeline (64:47).

"ZDR is not an effective solution, and there's a ton of leakage... increasingly, what we are all learning is that ZDR means literally nothing." - Chamath Palihapitiya [66:26]

His advice for any enterprise sitting on proprietary research, from drug design to legal strategy, is to stop treating a standard API contract as protection and instead stand up dedicated infrastructure, hosted through a vendor the company controls directly. Sacks pushes the argument toward policy: AI chat logs currently have less legal protection than email, since the government can obtain them with a subpoena rather than a warrant, even though people now use AI as a lawyer, doctor, and therapist all at once (71:01).

When the Story Replaces the Product

The Nike segment, on its surface, is a different subject. But it rhymes. Nike was removed from the S&P 100 index in September after eighteen straight years in it, following a stock decline of roughly 80 percent from its 2021 peak market value of $264 billion, a loss of more than $200 billion (80:09, 81:35). CEO John Donahoe, who took over in 2020, pushed an aggressive direct-to-consumer strategy that cut out the retail partners who had built the brand, leaving shelf space open for rivals like HOKA and On (80:34). China sales fell for eight straight quarters, losing ground to domestic brands Anta and Li-Ning (81:11).

Chamath's diagnosis is that Nike abandoned a clear identity built on athletic mastery in favor of campaigns built around messaging the company thought its audience wanted to hear, rather than the product itself.

"I aspire to be Michael Jordan. Always have, always will. I just think it's as basic as that. You're not going to buy the clothes of a brand that you don't aspire to be like." - Chamath Palihapitiya [85:14]

Friedberg's version of the same complaint is more literal: the shoes themselves got worse, falling apart within weeks, which is why he switched to Brooks, a running shoe brand owned by Berkshire Hathaway that has grown revenue by double digits for nine consecutive years to $1.6 billion, guided by one instruction from Warren Buffett to make the product better than it was the year before (88:35). The comparison the panel keeps landing on is the same one Bud Light faced: a company traded a specific, product-based identity for a general appeal to audiences it thought it should court, and lost the customers who actually bought the shoes.

Two very different stories, an AI resignation letter and a sneaker company's stock chart, end up making the same point. In both cases, the loudest claim in the room was not the most tested one. Anthropic's employees are asserting a number, greater than 10 percent, for an outcome nobody can verify and nobody can falsify either. Nike's marketing team assumed a narrative would carry a shoe that had stopped holding up on the trail. What actually moved the needle in each case was closer to the ground: the tokens OpenAI's agents actually produced, the stitching in a Brooks running shoe, the plain fact that no AI model yet can retrieve a toiletry kit from a hotel front desk without a human finishing the job. Confidence is cheap. The receipts, as this episode keeps insisting, are what is actually worth paying attention to.