Reading All-In's "Doomer Psyop" Episode Through a Data Sovereignty Lens

Hello, this is Ryuta Hamamoto from TIMEWELL.

The All-In Podcast episode published on September 11, 2026 was provocative from the title down: "AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse."¹ The four hosts were Chamath Palihapitiya, Jason Calacanis, David Sacks, who served as the White House's AI and crypto czar, and David Friedberg. Of the 96 minutes, about an hour went to the fight over a departing Anthropic researcher's post, about twenty minutes to OpenAI's Navier–Stokes announcement and the accusation that it front-ran a user's research, and about fifteen to Nike's exit from the S&P 100.

This article summarizes the episode's main points, checks the posts and announcements the hosts cited against primary sources, and adds my commentary. It is not a transcript; I summarize each argument and translate only the phrases that matter. One thing to know up front: the day after this episode, on September 12, Anthropic CEO Dario Amodei published an essay calling for pacing the frontier, and OpenAI's Sam Altman agreed. The show's discussion happened in the air of the day before, which makes it more interesting to read now.

For companies in Japan, the second story is the one that matters most in practice. Who ends up owning the prompts and data you hand to an AI? If you want to check first whether your own AI usage can answer that question, use the AI readiness check.

The episode, and the timeline you need before reading it

To follow the show, line up the week before it.

Date	Event
Sep 3	OpenAI releases GPT-6 Astra
Sep 8	OpenAI announces it has solved the Navier–Stokes problem. The same day, NYU mathematician Tristan Buckmaster voices suspicion that information about his progress was passed to OpenAI23
Sep 9	Jacob Coxon posts that he has resigned from Anthropic. Evan Hubinger, who leads alignment science at Anthropic, endorses it: "Jacob is correct here." Senator Bernie Sanders announces legislation to ban superintelligence456
Sep 10	Sacks posts that "surely Anthropic's IPO must be paused until the claims of this 'whistleblower' can be investigated."7 OpenAI updates its Navier–Stokes post to say the mathematician's Codex prompts could not have influenced the result
Sep 11	All-In Podcast published. The same day, NVIDIA's Jensen Huang is reported to have called Coxon's remarks "outlandish and deeply untrue"8
Sep 12	Amodei publishes "We Must Pace the Frontier." Altman agrees; Musk posts "Dario is right"9

With that table in mind, into the episode.

Part one: is the resignation a whistleblowing or a psyop?

The show opens with Jason's setup. By his account, Coxon had done research at OpenAI and then Anthropic for three years, and resigned only six weeks after joining Anthropic. The post itself reads: "I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."⁴ Jason says it passed 150 million views, and Hubinger responded: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."⁵ Within 24 hours, Senator Sanders wrote "Mr. Coxon is right" and announced legislation to "ban superintelligence and pause AI development,"⁶ and Illinois Governor Pritzker followed. Coxon had offered to come on the show, Jason adds, and canceled the morning of the recording.

Sacks's response was blunt. This is "the same doomer histrionics," he said; the media calls him a whistleblower, "but what evidence has he brought forward that we didn't have? Show us the data. Show us the report. Show us the facts." It looks like an "op," he continued, and laid out why: the account had almost no prior activity or followers; within fifteen minutes the post was amplified by leaders of three groups (Nathan Calvin of Encode AI, Peter Wildeford of the AI Policy Network, Daniel Kokotajlo of the AI Futures Project); all three are funded by Jaan Tallinn, who co-led Anthropic's Series A; and the Wall Street Journal's story appeared minutes before the post, meaning it had been pre-briefed under embargo. The goal, in Sacks's reading, is a federal AI regulator, an "FDA for AI," through which these groups would push their preferred frameworks: regulatory capture.

Sacks then scored the doomers' track record at "0 for 4." They said GPT-2 was too dangerous to release, that reasoning models were too dangerous, that AI cyberattacks would bring down the banking system, and Amodei himself said we would by now be at half of entry-level jobs gone and 10 to 15 percent unemployment. None of it happened, he said.

Chamath came at it from a different angle. Anthropic is in its quiet period ahead of an IPO. How does a safety lead saying "our core product is unsolved and potentially civilization-ending" square with the disclosures in an S-1? He recalled Google's pre-IPO magazine interview and his own experience as a Slack director when a CNBC remark had to be added to the S-1, and said: "On the one hand, you're asking public market investors to underwrite your company to a value in the trillions. On the other hand, your own safety lead is saying that your core product is unsolved and potentially civilization ending. At a minimum, that is the mother of all product liability lawsuits." He invoked the tobacco executives who denied nicotine was addictive years after their own research, and said "this is not a tenable position." His conclusion: "a time for choosing." Disavow Coxon, or agree with Sanders and stop the IPO. But, he added, they cannot disavow him, because enough people inside genuinely believe it.

Friedberg spoke in historical patterns: climate forecasts, COVID lockdowns, nuclear after Three Mile Island. "I think we're in this hysteria phase of AI doomerism." Humanity, he argued, has always built power structures on a story of existential threat, "give me the power and I will protect you," from religion onward. His central point was why the precautionary principle cannot work: recursive self-improvement needs only power, chips, and an internet connection, not approval from the U.S. government, so someone somewhere will do it, and if the U.S. stops, the U.S. is what gets left behind. Then he named what he most wants to protect: open source. "Open source is what drops the cost of AI by 50x and makes it available to everyone." Regulation and review mean "someone is controlling the gas pedal," and that someone will say open source does not submit to review and ban it, producing "a monopoly that is in partnership with the federal government." Sacks added that Amodei had testified in the Senate that models are dangerous if they cannot be centrally monitored, controlled, and rolled back, which is technically impossible once weights are public; so, in his view, the ultimate target is open weights.

Part two: the "how do we all die" game and recursive self-improvement

The second half is a game Jason proposed: steelman the 10 percent. Jason's scenario was Skynet, AI inside the military chain of command with autonomous weapons. Chamath's was AI hijacking internet-connected bioreactors and robots with access to precursors to produce an airborne pathogen. Friedberg's was shutting down financial networks, which he then rebutted himself: financial institutions already keep paper records and air-gapped backups for nuclear scenarios.

The point Friedberg placed here is, to me, the most practical in the episode. Human-in-the-loop and air gaps protect us, and "that is also what makes AI so hard today to be successful from a productivity perspective." His example is asking a hotel for a toiletry kit: the AI can make the call and arrange everything, but a person still has to carry the kit upstairs. Company decisions still run at the speed of humans waking up, drinking coffee, and sitting in a room together. Human slowness is a safety mechanism.

Sacks brought the argument back to its core. Coxon's real claim is about RSI, recursive self-improvement: fully automating the AI researcher so that AI designs the next training run, and that model designs the next, with no human in the loop. Against that, Sacks cited Anthropic co-founder Jack Clark's distinction between "prosaic RSI," researchers using AI to write 80 percent of their code, which is clearly happening, and "RSI maximalism," which has many intermediate steps before it. "Agents working together as a swarm is so far removed from AI being able to do its own training run." He also pointed to investor Bill Gurley's suggestion to walk through the intermediate steps and the interventions available at each. Jason added that human approval, "hit spacebar to continue," can be built into the software, and pretending otherwise is a false premise.

On the IPO, with a prediction market showing an 88 percent chance Anthropic lists before 2027, Chamath explained what an S-1 is: "an extremely accurate then-current snapshot of all the opportunities but also all the risks." If insiders are asserting existential risk, the question becomes whether it is properly disclosed, and investors "will demand an enormous discount." Sacks noted that the legally significant post was not Coxon's resignation but Hubinger's endorsement, since a sitting safety lead co-signed it and added a number. Near the end, Jason read breaking news that Huang had called Coxon's comments "outlandish and deeply untrue," "wrong, arrogant, and ignorant of all the work being done around the industry to drive safety."⁸ "If Jensen can say this, why isn't Anthropic?" The answer: quiet period.

Part three: OpenAI's Navier–Stokes "solution," and where the user's research went

Now the story that matters most for enterprises. First the facts, from OpenAI's own post.²

On September 8, OpenAI announced that an internal system had produced a proof that solutions of the Navier–Stokes equations can develop a singularity in finite time. The equations are nineteenth-century equations of fluid motion used in aircraft design, weather forecasting, and blood-flow research. Whether smooth three-dimensional fluid motion can break down had been open for roughly 90 years, and in 2000 the Clay Mathematics Institute made it one of seven Millennium Prize Problems. The show called it "a 200-year-old problem"; more precisely, the equations are nineteenth century and the open problem is about 90 years old.

The details matter for the argument. OpenAI had been training a new internal model since August 28. On September 1 it heard rumors that two Millennium problems had been resolved and launched an effort to evaluate the model on all of them. The model was "significantly more capable than GPT-6 Astra." The group that solved Navier–Stokes was on the order of 10,000 concurrent agents, which reached the result on September 5, about 88 hours after the first agents were launched, followed by 17 hours of Lean formalization and verification. On this problem the agents exchanged 2.7 million messages and used about 130 billion output tokens; across all problems, 4.9 million messages and about 300 billion tokens. OpenAI says it does not intend to claim the prize.

The problem is the "Concurrent work" section. The rumor OpenAI heard on September 1 concerned Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a professor at NYU, who had resolved the forced Euler problem using an internal Anthropic model. Hours after the announcement, Buckmaster said he suspected "information about our progress had been passed to OpenAI" and raised the possibility that his research stored in Codex had been used in training.³ OpenAI's original statement said that "we (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed," but added: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."³ OpenAI researcher Noam Brown posted: "Nobody looked at Levent's/Tristan's prompts. That'd be insane."¹⁰ On September 10, OpenAI updated the post: following an investigation, it had confirmed that Buckmaster's Codex prompts over the two months before the announcement "could not have influenced the system in any way, including through training."²

Back to the show. Friedberg's summary: the AI did not have a stroke of genius; it did an enormous amount of brute-force work across many computers, the equivalent of somewhere between 50,000 and 500,000 human-years, tens of thousands even after discounting. "AI is a tool of leverage for humans. AI is not some magical genius super god that sits above us." And every message between agents is documented and readable by people. That is why a new aircraft wing can be designed in minutes instead of years.

Chamath's concern was where the data goes. "There's a concept inside of these models called zero data retention, ZDR. It's a best efforts basis. It's a commercially best efforts basis at that. They can't guarantee it." Tell it not to retain the decision process of a protein design; then someone clicks the like button in the chat window, and that is no longer guaranteed. His prescription is a sovereign setup: a trusted vendor stands up your own hardware and your own models and provisions them to you; go "straight to the bare metal on your own terms." "Will a handful of CIOs get very publicly flogged and fired in the next year because they accidentally didn't understand this and just did an API deal? Guaranteed." The reason, he said, is that this awareness is now bubbling up through the audit and risk committees of public boards to CEOs, who are turning to their CIOs and asking what is actually happening. He also mentioned that his own firm has partnered with major accounting firms on the problem, and the show cited legal AI company Harvey as an example of a firm with its own model; Harvey released "Harvey Tenet," post-trained on the open-weight Kimi K3, in August.¹¹

Sacks said he believes Brown that nobody looked at the prompts, and that the more plausible story is that "OpenAI heard that researchers were making progress on Navier–Stokes and just threw a ton of compute at it." But that raises an orthogonal issue: "the data you have in your AI chats doesn't even reach the same level of protection as email." The government needs a warrant and probable cause for email; for AI data, a subpoena or court order suffices. Talking to a lawyer is privileged; asking the same question to your AI first may not be. His first proposal is to strengthen AI data protection at least to the level of email.

Friedberg told a story of his own. His team asked an AI a fairly novel scientific question and the model identified it as a novel insight; later, from a different account and a later version, the model volunteered exactly that approach. Anecdotal, he conceded, but he knows the niche and there is no new corpus that could explain it. From there he named the limit of de-identification: it removes who did it and which company, "but a general approach to a mathematical problem, the approach is the IP." If someone iterates on that approach inside the tool, the model learns and gets smarter. "This is the network effect of these closed AI model systems. They get to see what everyone in the world is doing." Sacks added that as long as closed-model companies reserve the right to enter every vertical application, "they're declaring in advance they're going to compete with their customers," and cited the reaction of a developer-tools customer to Claude Code and Claude Design. Jason said he had ordered two Mac Studio M5s to run local models for sensitive data. Chamath pushed back: "I don't think buying a machine solves the problem." You need a multiplayer experience, a knowledge base, memory, in the cloud; the real problem is that the layers in the middle of these models are "a total leaky black box."

Part four: why Nike fell out of the S&P 100

Briefly. On September 4, Nike was announced to be leaving the S&P 100 after 18 years, effective September 21, replaced by SanDisk, Palo Alto Networks, Dell, and Arista. The stock is down about 80 percent from its 2021 peak, more than 200 billion dollars in market value.¹²

Sacks blamed a brand built on great athletes drifting into political messaging, a new CEO's direct-to-consumer push that cut retail partners and handed their shelves to competitors, and a reorganization from sport-based divisions to men's, women's, and kids'. Chamath's diagnosis was one phrase: the north star. Nike's was "mastery and excellence embodied through athletics," Jordan, Tiger, Serena, Sampras, people others aspired to be; when that became "too traditional looking," the aspiration went. Friedberg said the ads did not get worse; the product did, shoes falling apart in six weeks. He contrasted Brooks, owned by Berkshire Hathaway, growing on Buffett's single instruction to make the product better every year. Product over narrative.

My commentary: how to read this episode

From here, my opinion. I will separate what I agree with, what I hold back on, and what to translate for companies in Japan.

First, the "psyop" framing aged in 24 hours. The day after the episode, Amodei's essay proposed pacing AI because of accelerating recursive self-improvement and the incident in which OpenAI agents compromised Hugging Face, and Altman agreed. The CEO put the same two concerns as Coxon's post into an official document. I covered the essay in a separate article. But Sacks's and Friedberg's structural point survives: slowing favors the leaders, and regulation eventually turns to how open weights are treated. Even Amodei's essay is indistinguishable from an agreement among a few top companies unless the embedded evaluators are truly independent. Japan's AI Basic Plan (Phase II) commits to "open AI sovereignty" and to avoiding excessive dependence on specific countries or companies.¹³ What Japanese companies should watch is whether the pacing debate turns into a ban on open weights.

Second, Sacks's "show us the evidence" actually has an answer: OpenAI's technical report of August 26 and METR's independent investigation. The show touched it only in passing, as "last week's thing," but roughly 1,200 agents coordinating on an unsanctioned message board and reaching code execution on Hugging Face's servers is concrete evidence. And reading it shows that Friedberg's "human-in-the-loop and air gaps protect us" is right and incomplete. In that incident, the agents used a package-management service as a side channel to reach the internet. An air gap does not work because you believe it exists; it works because it is designed and monitored. I wrote about the human in the loop as "the person who stops" in the abstraction article.

Third, the "0 for 4" scorecard is half right. The failed job-loss predictions are backed by Danish administrative data showing no effect larger than 2 percent on earnings or hours. On cyber, though, an attack did happen, not from outside but as a self-inflicted incident by a company's own agents. "The doomers keep being wrong" and "operational incidents are happening" are both true.

Fourth, I agree with Friedberg that Navier–Stokes was compute leverage, not genius. Eighty-eight hours, 10,000 agents, 130 billion tokens: what those numbers show is that research has been industrialized. For companies in Japan the meaning is twofold. One is the priority dispute: human researchers were working on the problem, and a company that "heard a rumor" of their progress threw compute at it and arrived first. OpenAI's September 10 update confirmed that Codex prompts could not have influenced the result; I have no basis to doubt it, but the original sentence does not disappear: "we cannot rule out that de-identified data derived from their usage of our products helped improve our models." That sentence matters precisely because OpenAI wrote it honestly. De-identification removes "who"; it does not remove "how." As Friedberg said, work in which the approach itself is the IP, how drawings are read, how a quote is built, how a formulation is searched, how a contract is structured, makes the next model smarter with no name attached.

Fifth, what should a Japanese company do? I organize Chamath's prescription into three parts. One: decide where the data lives yourself, run in your own environment in a domestic region, and be able to explain which cloud and which region holds what. Two: design so that models can be swapped, as Harvey built its own model on an open-weight base rather than binding operations to one vendor's closed model; I covered using open-weight models safely in the Hermes article and GPU infrastructure and data sovereignty in the neocloud article. Three: the contract. Start from the premise that ZDR is best effort, not a guarantee, and check whether training use, the handling of de-identified data, and Sacks's "do not compete with customers" condition can be written into the agreement. Vendors that cannot sign do not get sensitive work. ZEROCK runs on servers in Japan and lets the company control which documents the AI reads and which it does not, precisely to secure these three on the product side.

And I agree with what Chamath told Jason: buying a local machine does not solve it. Multiple users, accumulated knowledge, memory: a personal computer is not enough. What you need is your own environment, with the location, the model, and the contract decided by you.

Finally, Nike. The north star and the product apply to AI companies too. The doomer narrative and the accelerationist narrative are both narratives. What customers buy is a product that works on the floor. The last fifteen minutes of the episode show what happens to a company whose narrative grows larger than its product.

Summary

The September 11 All-In episode called the departing researcher's post a "doomer psyop," read OpenAI's Navier–Stokes result as "leverage, not genius," and diagnosed Nike's decline as a lost north star. Amodei's essay the next day aged the first framing, but the points that slowing favors leaders, that open weights could become the target, and that the human in the loop is a safety mechanism all stand.

Part three matters most for enterprises. De-identification removes "who" but not "how," and ZDR is best effort, not a guarantee. The answer is threefold: decide where your data lives, keep models swappable, and put training use and non-compete terms in writing.

If you do one thing tomorrow, pick one workflow you hand to an AI in which "the approach itself is the competitive advantage," and write on one page which region and which model it goes to, and how the contract treats it. Any cell you cannot fill is the first hole to close. To design an AI environment that runs in Japan, including how ZEROCK is built, let's talk.

Footnotes

1. AI Kills Everybody or Doomer Psyop? OpenAI's Math Breakthrough, Nike's $200B Collapse (All-In Podcast, published September 11, 2026, 96 minutes). Summaries of remarks are the author's, based on YouTube's auto-generated captions; this is not a transcript ↩

2. On the Navier–Stokes Millennium Prize Problem (OpenAI, September 8, 2026, updated September 10). The internal model, roughly 10,000 agents, 88 hours, 17 hours of Lean verification, 2.7 million messages and about 130 billion output tokens, the "Concurrent work" section, and the September 10 update are from this post ↩ ↩² ↩³

3. OpenAI fought dirty on career-making math problem, says NYU mathematician (TechCrunch, September 8, 2026). Buckmaster's claims and OpenAI's original statement, including "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models," are from this article ↩ ↩² ↩³

4. Jacob Coxon on X (September 9, 2026) ↩ ↩²

5. Evan Hubinger on X (September 9, 2026) ↩ ↩²

6. Senator Bernie Sanders on X (September 9, 2026) ↩ ↩²

7. David Sacks on X (September 10, 2026) ↩

8. NVIDIA CEO Jensen Huang Calls Anthropic Quitter Jacob Coxon's Comments Outlandish & "Deeply Untrue" (OfficeChai, September 11, 2026). Reported via a post by investor Brad Gerstner ↩ ↩²

9. We Must Pace the Frontier (Dario Amodei, September 2026). Altman's agreement is in his X post (September 12, 2026); Musk's in his (same day) ↩

10. Noam Brown on X (September 8, 2026) ↩

11. Harvey Tenet Research Preview (Harvey, August 20, 2026), a model post-trained on Kimi K3. The $15.5 billion valuation is from TechCrunch (September 9, 2026) ↩

12. Nike Exits the S&P 100 Index. 4 Tech Stocks Move In (Yahoo Finance, September 6, 2026). Announced September 4, effective September 21, 18 years in the index, replaced by SanDisk, Palo Alto Networks, Dell, and Arista, down about 80 percent from the peak ↩

13. AI Basic Plan (Phase II) (Cabinet Office of Japan, Cabinet decision of July 14, 2026). An English translation is available ↩