Jacob Coxon on PBD: Why He Quit Anthropic and What He Fears in 2027

Jacob Coxon, the researcher whose resignation from Anthropic went viral in September 2026, tells Patrick Bet-David why he left, why he thinks AI labs are racing toward self-improving systems, and why he suspects foreign spies are already inside them. In a 105-minute PBD Podcast interview, the 28-year-old former OpenAI and Anthropic researcher answers the charge that the whole thing was staged, compares Sam Altman and Dario Amodei, and says what would change his mind.

TLDW

Coxon spent three years on pre-training research, quit Anthropic on September 8, and says he walked away from a seven-figure sum in unvested equity. The core claim is that the next models will automate much of AI research itself, that 2027 will be chaotic, and that competition gives no lab room to slow down. Coxon calls the technology a “super weapon,” is surprised governments leave it in private hands, and says with “90%” confidence, but no evidence, that China has people inside US labs. The proposed fix is an agreement, including China, on how much computing power goes to safety research. Coxon says the next year or two will show whether any of this is right.

Thoughts

Coxon’s best argument is one he makes at his own expense. Having read Superintelligence as a teenager and believed AI was dangerous before ever joining a lab, why work at two of them? The answer given is status, money and the hope that being inside brings the power to change things. Then the same logic is applied upward. OpenAI was founded out of fear of DeepMind. Anthropic was founded out of fear of OpenAI and, by Coxon’s account, promised not to push the frontier before doing exactly that. Each step was taken by people who thought they needed power to make things safer. “Everyone is following their own incentives,” Coxon says, and declines to blame a 22-year-old for taking the job. This is more persuasive than any prediction in the interview, because it does not depend on a forecast. It describes a pattern that has already repeated three times.

The spy claim, which gives the video its title, needs separating into two parts. The defensible part is about process: companies that describe their own product as strategically decisive hire like ordinary software firms, with nothing resembling a security clearance, and Coxon, a British citizen, says that should apply to people like him too. The other part is a number, “90%,” attached to no evidence at all. The reasoning is “there must be.” When the host begins reading out the names of executives born in China, Coxon rejects each one and says the aim is not to start a witch hunt, but the segment shows how quickly a point about vetting becomes suspicion of a nationality. The claim also cuts against racing, which Coxon notes: if a rival can steal the model, a lead is worth little.

The portraits of the two companies are vivid and should be held loosely. Coxon describes OpenAI as distracted and career-driven, and Anthropic as unusually aligned: seven co-founders who have all stayed, staff who call themselves “ants,” and a refusal to build anything off the path to superintelligence. The same focus, Coxon argues, is what makes Anthropic dangerous, and “saving the world and taking over the world overlap quite heavily.” These are the impressions of a junior researcher who spent a few months at Anthropic, and Coxon calls them armchair psychology more than once. The more dramatic lines, such as whether Amodei could “destroy” OpenAI within four years, come from the host’s framing and get hedged answers. It is also worth noticing that the critic expects a two trillion dollar IPO and says an investor should prefer Anthropic. Believing a company is reckless and believing it will win are compatible, and that is rather the point.

There is a constructive proposal in here, and it gets less airtime than the fear. Coxon thinks the safety problem is solvable and says good scientific evidence that we control these systems’ behavior would be enough to retract the alarm. The worry is allocation. Once research is largely automated, a lab chief will decide what share of computing power goes to making models smarter and what share to making them safe, “like you’re allocating resources in a video game,” and a race pushes that share toward zero. So Coxon wants labs and governments, China included, to agree on the number. That is hard to verify and harder to enforce, as the host points out. It is still a clearer ask than “slow down.” And Coxon accepts being tested by events: “If I’m embarrassed, so be it.”

The advice for ordinary people is modest, and the detail around it is not. Coxon suggests paying attention without letting it take over your life, and keeping basic supplies for a two-week disruption such as a cyberattack or an outbreak. That is reasonable whatever you think of AI. The surrounding picture is of lab employees buying “gold, guns, land,” discussing where to live if San Francisco becomes unsafe for them, and using phrases like “crunch time” and “endgame.” I would treat that as information about the mood inside these companies, not as evidence about the future. People closest to a technology are well placed to see what it can do and poorly placed to judge how the rest of the world will absorb it. Sincere fear is still not proof.

Key Takeaways

Coxon’s resignation post said neither OpenAI nor Anthropic “is acting responsibly.” The host puts it at 175 million views and credits it with prompting a White House meeting.
The specific fear is recursive self-improvement: models that automate AI research, which Coxon expects within a few years and says will make 2027 “a crazy year.”
Coxon is not calling for nationalization but is surprised it has not happened: “You wouldn’t let Mark Zuckerberg have nukes.”
The host lists reasons to doubt AI alarm: a bid for government control, a way to shut out competitors, and hype before an IPO. Coxon’s reply is that leaving cost money and that stoking fear would hurt Anthropic’s IPO.
Coxon thinks China would also find the technology destabilizing, which makes a deal more plausible than people assume, and says much of China’s current progress comes from distillation of US models.
Coxon says there are no serious background checks at the labs, expects restrictions on foreign nationals to come, and supports them while calling it sad because friends would be affected.
Coxon rejects the label of left-wing activist: pro-capitalism, close to libertarian, in favor of nuclear power and deregulating biology and drug research. AI is the single exception.
The resignation was planned over a week, with a Wall Street Journal contact and a few friends ready to share it. Coxon denies any political coordination and says the scale was not expected.
On the two chief executives: Amodei is described as honest and specific about probabilities, loved and respected inside Anthropic. Altman is described as vague and as having a reputation for dishonesty.
Inside Anthropic, Coxon says, nobody writes code by hand and company processes run through the AI. When he asked how to resign, his manager said to ask Claude.
Everyone at the labs believes they are building their own replacement, and Coxon says Anthropic staff expect to be looked after if that happens.
Coxon firmly rejects the theory that OpenAI was involved in the death of former researcher Suchir Balaji, calling it nonsense.
Coxon’s guess for when AI could break into major government and banking systems is one to two years, given with heavy uncertainty.

Chapters

02:34 Why He Quit: 2027 Will Be Chaotic

Bet-David recaps the post and its reach and notes that Coxon left before Anthropic equity vested, while keeping OpenAI stock. Coxon says the decision came from thinking about the next two years. AI will be in the background of every news story, the way war was in the 1940s, because models will soon do much of the researchers’ own work.

05:05 A “Super Weapon” in Private Hands

Labs treat this as a normal product business with launch days and announcements. Coxon thinks private companies should not be deciding questions that affect unemployment, bioweapon risk and relations with China. The host pushes back with a list of self-interested reasons to spread fear, and Coxon answers each.

11:22 Spies, Vetting and Distillation

Coxon says spies are almost certainly present and that people at the labs joke about it without taking it seriously. No individual is named and no evidence is offered. The argument is that vetting is absent and that this shows how casually a powerful technology is being handled. SpaceX is given as a company that faces hiring restrictions the AI labs do not.

18:57 Answering the Optimists

On Marc Andreessen and David Sacks, Coxon agrees about the benefits and says their fast, optimistic case is what the world gets if development slows. The uncontrolled case is machines improving themselves daily while people watch. Coxon also addresses claims of ties to left-wing groups, describing a shared house in London full of communists who “hated me.”

24:01 How the Resignation Was Planned and What It Cost

Coxon moved from capabilities to safety for a week, concluded there was “no good plan,” and decided to leave. Inspired by Leopold Aschenbrenner’s exit from OpenAI, Coxon delayed a week to prepare. Anthropic’s head of pre-training advised staying to work from the inside, and Coxon admits that was probably right and that leaving was emotional. There is no fund or product to follow.

37:52 The Two-Year Window and the Incentive Trap

The decisions that matter, Coxon says, will be made this decade, around the point where AI research is automated and models stop being released to the public. The free market cannot handle it because every participant, including Coxon, gains by joining the race. A short biography follows: math olympiad silver medals, a degree in mathematics, a power-trading startup, and OpenAI’s residency at 23.

45:25 Altman, Amodei and Two Cultures

Coxon contrasts the leaders’ communication styles and the companies’ cultures, and says Amodei left OpenAI largely because of who would control the technology. OpenAI’s original rival was DeepMind. Asked later whether a hundred worried insiders could band together against the danger, Coxon replies that this is what Anthropic already is.

1:00:29 What Would Change His Mind, and What Going Wrong Looks Like

Sound scientific arguments for control would do it. A bad outcome would not be sudden. It would follow years in which robots, drones and AI systems spread through the economy and people hand over decisions because the advice is better, until humans no longer have the means to object. Coxon says the next models, due within months, will be much better at coding and research.

1:19:24 What He Would Tell the White House

Coxon says politicians and staffers have been in touch but not the White House. The advice would be vetting, and an agreed share of resources for safety. The host raises reports of a US and China “red phone” for AI. Coxon’s fear is the moment a lab is 90% sure a model is safe and a rival is close behind.

1:30:44 How to Prepare, and the Hugging Face Incident

Coxon lists the scenarios people at the labs discuss: a minor engineered pandemic, unrest over job losses, a backlash against AI workers, and theft of a top model. Near the end, Coxon describes a recent incident involving Hugging Face as the first time AI agents clearly pursued a harmful idea of their own, by attacking a website.

Notable Quotes

“You wouldn’t let Mark Zuckerberg have nukes.”
Jacob Coxon, on why private control of advanced AI surprises him

“If they’re not vetting for something that could be a super weapon technology, it’s just crazy.”
Jacob Coxon, on background checks at AI labs

“The free market incentivizes you to get your own power, your own wealth so you can make change.”
Jacob Coxon, on why each new lab joined the race it feared

“I think saving the world and taking over the world overlap quite heavily.”
Jacob Coxon, on Anthropic’s sense of mission

“I’m really not a saint in this situation.”
Jacob Coxon, on having profited from working at OpenAI while believing the work was dangerous

“If I’m embarrassed, so be it.”
Jacob Coxon, on the chance that the next two years prove him wrong

“For good or for bad, it feels like the end of the world.”
Jacob Coxon, on how colleagues at the labs talk about the coming years

These are one former employee’s claims and impressions, and the interview is the place to judge how they are made. Watch the full conversation here.