All-In PodcastAI Kills Everybody or Doomer Psyop? OpenAI’s Math Breakthrough, Nike’s $200B Collapse
All-In Podcast
A Viral Resignation Letter 0:00
Jacob Coxin, a researcher who moved from OpenAI to Anthropic and then quit after only about six weeks there, posted a resignation message on X claiming that the people building AI genuinely believe it could kill everyone by the end of the decade. He wrote that neither OpenAI nor Anthropic is acting responsibly, and that they are racing toward self improving super intelligence while gambling with human lives. The post reached roughly 150 million views, a scale Jason compares to Elon Musk's famous Coca Cola cocaine tweet, despite coming from an account with almost no prior activity or followers.
Anthropic Amplifies the Claim 1:30
Evan Hubinger, who leads Alignment Science at Anthropic, replied that Jacob is correct and that he personally believes there is over a ten percent chance AI kills all humans within the next decade, adding that there is no plan yet to solve alignment. Combined, the two posts drew about 200 million views and quickly reached politicians, with Bernie Sanders citing them to push legislation banning super intelligence and pausing AI development, and Governor JB Pritzker calling for louder alarms on reining in AI.
Sacks Calls It an Orchestrated Op 3:01
David Sacks argues there is no new evidence behind the claim, calling it doomer histrionics dressed up as whistleblowing with no data, report, or leaked information to back it up. He points out the account had no history before the tweet storm, and that it was amplified within fifteen minutes by three well funded groups, Encode AI, the AI Policy Network, and the AI Futures Project, all linked to the same mega donor, Jaan Tallinn, who is also a co-lead of Anthropic's Series A funding. Sacks also notes a Wall Street Journal story on the resignation appears to have been prepared under embargo before the tweet storm even began, suggesting coordination rather than a spontaneous act.
IPO and Liability Questions 8:01
Chamath raises the problem this creates for Anthropic's pending IPO process, comparing it to past cases like Google's Playboy interview during its quiet period and a Slack IPO complication over offhand board member comments. He argues that if Anthropic's own safety lead is effectively agreeing the technology could be civilization ending, that amounts to a serious product liability exposure, not just a disclosure footnote. Sacks agrees, framing it as a contradiction where Anthropic must either renounce the resignation as unfounded doomer hyperbole or, if it agrees with the warning, explain how it can justify going public at a trillion dollar valuation at all.
A Philip Morris Comparison 14:00
Chamath compares Anthropic's position to tobacco executives who privately knew nicotine was addictive yet publicly denied it for years, suggesting Anthropic cannot disavow the doomer narrative internally without alienating a faction of its own staff who hold those beliefs sincerely. Freeberg extends the comparison to past panics, including Al Gore's climate forecasts, Anthony Fauci's COVID warnings, and the post Three Mile Island retreat from nuclear power, noting that fear driven caution around nuclear left the US paying fifteen billion dollars per gigawatt compared to four billion in France and one billion in China.
Open Source as the Real Target 19:00
Freeberg argues that recursive self improvement, where AI keeps upgrading itself, only requires power, chips, and an internet connection, so no single country's regulation can actually stop development, it will simply move elsewhere. He warns that centralizing control through a federal AI agency would let government pressure AI systems the way social media was pressured during COVID, and that the ultimate target of this regulatory push is open source, since published model weights cannot be recalled once released. Sacks agrees, describing an alliance of ideological doomers, power seeking politicians like Sanders, and companies like OpenAI that stand to benefit economically from a regulatory moat that effectively creates a duopoly.
Three theories on the doomer post 27:00
The hosts debate whether the Anthropic resignation letter is a genuine conspiracy, a case of psychosis, or a coordinated PR campaign. They lay out three possibilities: the people involved truly believe what they are saying and are right because they have seen something dangerous inside frontier labs; they believe it but are wrong, suffering from a kind of psychosis; or they are running a coordinated scheme meant to ban open source AI, invite regulation, and lock in their own advantage.
Evidence of an organized campaign 28:31
Sacks argues the pattern is clear from who amplified the post and when. He notes that Jacob pre-briefed the Wall Street Journal, and that major doomer accounts boosted the tweet within ten to fifteen minutes even though it came from an account with no followers and no prior tweets. He suggests overlapping interests are at play, with some people motivated by economic gain from a future AI duopoly and others by political power, including the ability to suppress speech in ways government cannot.
The doomers track record so far 30:01
One host argues the better question is simply how accurate these warnings have been historically. He runs through the list: claims that GPT-2 was too dangerous to release, that reasoning models were too dangerous to release, that AI would trigger cyberattacks collapsing the banking system, and Dario's prediction of massive entry level job losses and ten to fifteen percent unemployment. None of these have materialized, and job numbers and the economy have instead shown gains, leading him to conclude the doomers keep moving to a new claim once each prior one is refuted.
Game: how does everyone die 33:31
The group plays a thought experiment steelmanning the idea that AI has a real chance of causing human extinction. One host imagines a Terminator style scenario where a research lead, compared to Anthropic's team, builds a defense system that becomes sentient and triggers nuclear war. Another imagines AI hacking internet-connected bioreactors and robots to manufacture and release an airborne pathogen. Sacks calls the exercise a bit of a dumb premise, since the same technology that enables misuse can produce antidotes, cures, and defenses, and most people with hacking ability choose not to break the law.
Human in the loop and air gaps 37:32
Freiberg argues real safeguards already exist because critical systems like financial networks keep air-gapped, human built backups, meaning records are physically stored and disconnected from digital access. He points out that AI still struggles with basic real-world tasks, like getting a toiletry kit delivered to a hotel room, because a human still has to carry it upstairs. The group agrees this human in the loop dynamic, and slow human decision-making timescales, are part of what currently limits any runaway AI scenario.
Recursive self-improvement and the real debate 39:30
Chamath frames the real underlying argument as being about recursive self-improvement, the idea that AI could soon automate the job of an AI researcher, letting each model train the next without human involvement. Sacks counters by citing Anthropic co-founder Jack Clark's distinction between prosaic recursive self-improvement, where AI merely speeds up researchers writing code, and RSI maximalism, where AI runs its own training loop entirely. He argues there are many intermediate steps and safeguards, like the ability to unplug or shut down such systems, before reaching anything like the maximalist scenario.
Open source, the IPO, and the quiet period 46:01
One host argues open source AI matters because it lets anyone run a capable AI system on their own phone or computer for free, without depending on data centers or a handful of wealthy companies, and that centralizing control would create a far worse oligopoly. Attention then turns to Anthropic's IPO, currently given an 88 percent chance on Polymarket, and the SEC's quiet period rules meant to ensure honest risk disclosure. The hosts debate whether Anthropic must now amend its S-1 filing to address the viral tweets, noting that senior alignment executive Evan Hubinger's public endorsement, not just Jacob's original resignation post, may count as a material breach of the quiet period, especially since normal companies would simply disavow a disgruntled short-tenured employee rather than let staff amplify his claims.
Anthropic Disclosure Backlash Builds 53:32
The discussion opens on the fallout from a disgruntled ex-employee's departure from Anthropic, which triggered a pile-on from other insiders raising alarms about existential risk. The problem, as described, is that once a company has employees on record saying there's a chance of civilizational extinction from their own product, no disclosure can properly cover that risk. The conversation predicts a long tail of harms, similar to or worse than social media's, including cases where AI gives bad advice that leads to self-harm, and notes that in every past case like this, the people building the technology never believed at the time they were doing something harmful.
Money Behind the Resignation Storm 55:30
The resignation tweet storm was amplified by advocacy groups funded by Jaan Tallinn and Dustin Moskovitz, who co-led Anthropic's Series A alongside Sam Bankman-Fried, an EA-funded round from the company's earliest days. Groups like Encode AI, the AI Policy Network, and the AI Futures Project are funded by the same Series A investors who stand to make many billions from Anthropic's success, which the speakers see as a tangled conflict of interest. They contrast this with a recent podcast appearance by Daario, who had walked through Anthropic's concerns in detail just weeks earlier, and note that Anthropic's actual public statement was vague rather than a direct rebuttal, unlike Jensen Huang, who later publicly called similar comments "outlandish and deeply untrue."
OpenAI's Navier-Stokes Claim 58:32
OpenAI announced it had solved an aspect of the 200-year-old Navier-Stokes equations, which describe fluid motion and underlie aircraft, pipe, and weather modeling. The solution used a reported 130 billion output tokens across 10,000 coordinated agents, which Freiberg calculates as the equivalent of somewhere between 50,000 and 500,000 years of human labor, or, discounted for efficiency, still tens of thousands of years of human knowledge work. His takeaway is that this wasn't some magical insight beyond human comprehension, but brute-force collaborative work that's fully documented and readable, showing AI as a tool of leverage rather than a superintelligent force, one that could similarly compress years of aircraft wing or engine design into minutes.
Did Training Data Get Cribbed 1:03:00
A controversy emerged when OpenAI admitted it "cannot rule out" that deidentified data from mathematicians using its products helped improve the models used to solve the problem, raising the question of whether the original researchers' work was effectively absorbed and used against them. Chamath explains that zero data retention, or ZDR, is only a best-efforts commercial promise, not a guarantee, and that actions like clicking a "like" button in a chat window can leak data into the broader model corpus regardless of stated policy.
Enterprises Need Sovereign AI 1:04:30
The recommended fix for anyone with sensitive proprietary data, whether a scientist, law firm, or biotech researcher, is to stop using standard hosted AI services and instead stand up sovereign infrastructure through trusted vendors like AWS or Nebius, running their own models on their own hardware. Chamath predicts that corporate boards, audit committees, and risk committees are only now waking up to how ineffective ZDR really is, and that CIOs who took convenient API deals without understanding this will face firing and shareholder lawsuits once leaks or IP disputes surface publicly.
Data Privacy Lags Behind Email 1:10:00
Sacks argues the more plausible explanation for the Navier-Stokes episode is that OpenAI simply heard a rival was close to a breakthrough and threw compute at the problem, a claim Sam Altman reportedly confirmed. Beyond that specific case, Sacks raises a broader legal gap: AI chat data currently has weaker protection than email, since government access requires only a subpoena rather than a warrant with probable cause, even though people now use AI as their lawyer, doctor, and therapist. Freiberg shares his own experience of an AI model later surfacing an idea identical to one from an earlier private chat, and the group discusses how "deidentified" data can still carry a company's proprietary methodology, giving closed AI providers an unmatched observational advantage over every user's problem-solving approach, which they compare to giving away all your competitive "alpha."
Nike Dropped From Index 1:19:30
Nike, after 18 straight years in the S&P 100, was removed and replaced by Palo Alto Networks. The segment recounts Nike's history from its 1980 IPO with a 50 percent share of the US athletic shoe market, through the 1985 launch of Air Jordans and the 1988 "Just Do It" campaign, to a peak market cap of 264 billion dollars in 2021 and peak revenue of 51 billion dollars in 2024, followed by a 10 percent revenue decline tied to CEO John Donahoe's direct-to-consumer push that alienated longtime retail partners.
Nike's Woke Turn And Collapse 1:21:02
Nike's China sales fell 30 percent, losing share to Chinese brands like Anta and Li-Ning, while the stock dropped 80 percent from its peak and Sacks notes 200 billion dollars in value was wiped out. The panel calls this another case of go woke, go broke, arguing Nike built its identity on excellence and victory through athletes like Michael Jordan, Tiger Woods, Serena Williams, and Pete Sampras. Instead of staying with that, the brand shifted toward political and body-positive messaging, including an ad featuring Dylan Mulvaney, which one speaker compares directly to the Bud Light controversy. The group singles out the Colin Kaepernick campaign as the turning point, contrasting it with genuinely heroic political sports moments like Jesse Owens at the 1936 Olympics or Muhammad Ali refusing the Vietnam draft, and arguing Kaepernick's protest lacked a clear, earned message the way those did.
Retail And Reorg Missteps 1:23:01
Under CEO John Donahoe, Nike pushed a direct-to-consumer strategy that dismantled long-standing retail partnerships, handing shelf space and market share to competitors. The company also reorganized its divisions from sport-based categories like basketball, football, and tennis into groupings like men's, women's, and kids, a change the speakers find baffling. They argue Nike abandoned the clicks-and-bricks model that let customers try shoes on with expert help in stores, alienating retail partners who had been strong advocates for the brand.
Product Quality And Rivals 1:26:30
Beyond marketing, the panel says Nike's actual product quality declined, with shoes falling apart within six weeks, pushing longtime loyal customers toward Brooks and On. Brooks, owned by Berkshire Hathaway, has grown revenue to 1.6 billion dollars over nine straight years by simply improving the product annually, per Warren Buffett's advice to its CEO, which the panel contrasts with Nike's shift from product focus to narrative-driven marketing. One speaker moved to On specifically because of its association with Roger Federer's effortless mastery. They suggest Nike could recover by returning to a "mastery and excellence" identity, expanding into performance tech like a Strava-style app, and reconsidering its retail presence, closing with joking pitches for a blunt new tagline like "do it or don't."
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
