On September 8, a researcher named Jacob Coxon quit Anthropic with a farewell note that got more than a hundred million views in a day. “The people building AI earnestly believe that it could kill us all by the end of the decade,” he wrote. “This is not a marketing stunt.” Anthropic’s alignment lead, Evan Hubinger, agreed in public within hours and put his own odds north of one in ten within a decade. Two dozen members of Congress said something needed to be done. Elon Musk called it a psy-op.

Ian Duncan’s essay starts from a fair question about a warning like that: to weigh it, you need to understand the people making it. What follows is a long account of a twenty-five-year-old internet subculture that is obscure outside Silicon Valley and unavoidable inside it.

What the essay establishes

  • The people warning about the end of the world are not a fringe. They are a social scene — “overlapping friendship groups, funding relationships, and sexual and romantic relationships, all of it threaded through the institutions asking you to trust them about AI regulation.” Its alumni staff the labs, the safety institutes and the philanthropies; its writers make or break reputations.
  • The vocabulary is theirs. “AI alignment” was popularised by Eliezer Yudkowsky’s nonprofit, which he co-founded in 2000 with no formal training in the field and a self-published autobiography about his singular qualifications to steer the species. The core idea that spread furthest: if AI is developed, it will become superintelligent, escape control and kill everyone.
  • The recruiting pipeline ran through a Harry Potter fanfic. Harry Potter and the Methods of Rationality ran 122 chapters and 661,619 words. Its most consequential idea is “heroic responsibility” — everything that goes wrong is ultimately your fault because you could have prevented it. Its finale also functioned as a demonstration: the author told readers to post a solution to the story’s last problem by a deadline or get a sadder ending, and collected more than 700,000 words of submissions in two and a half days.
  • Two on-ramps brought in the respectable world. Scott Alexander’s blog translated the community for a Silicon Valley readership, and Curtis Yarvin wrote the other one — a program he stated plainly in 2008 as “the liquidation of democracy, the Constitution and the rule of law,” plus sovereign territories run as businesses. Peter Thiel funded both Yudkowsky’s institute and Yarvin’s company, and Yarvin’s ideas later circulated into the second Trump administration.
  • The internal costs are documented by the community itself. The Center for Applied Rationality’s co-founder and president acknowledged that at least three employees had effectively been compelled into “debugging” — the community’s term for intervening in someone’s psychology — and that it damaged how participants understood themselves. The organisation’s own apology accepted credible allegations that a man it had legitimised physically, sexually and emotionally abused two former partners, having let him help at a high-school program.
  • The darkest thread is the Zizians, a splinter group assembled from the community’s own parts. A faked death, a landlord attacked by masked tenants and later stabbed to death before he could testify, two parents shot in their home, and a traffic stop in Vermont that killed a Border Patrol agent and a former maths-olympiad medallist who had entered the world through a CFAR event.
  • Follow the money. Longtermism — the claim that future lives dominate the moral arithmetic — turned AI doom into effective altruism’s flagship cause. Open Philanthropy reported roughly forty million dollars in AI-risk grants in 2017, thirty million of it to OpenAI. FTX, whose founder was effective altruism’s most famous product, put five hundred million dollars into Anthropic. Thiel’s network produced a vice president.
  • The professional and romantic networks are the same network. A 2014 community survey found roughly fifteen percent identified as polyamorous, concentrated in a handful of Berkeley group houses where everyone is also everyone’s coworker, funder, landlord or ex. Duncan’s question is not who sleeps with whom: “The question is whether anyone in it can say no, to anything, without losing their whole world at once.” The New Yorker has mapped the top of that field, from AI-forecasting researchers to the philanthropist married to Anthropic’s president.

What he does with Coxon

He refuses to resolve it. He grants the warning’s weight: by every hostile account as well as friendly ones, Coxon spent three years inside the industry, left the safest-seeming employer in the field and gave up the money to make the warning unambiguous. Asked whether Anthropic was cutting corners, Coxon said it was “far and away the most responsible player in the space,” and that this was exactly the problem — if the responsible player is racing, there is no responsible player.

He also says the timing just before vesting “feels like a staged narrative choice,” and lays out the regulatory-capture theory he admits is speculation: legislation was moving within days, and the natural answer to “the labs are gambling with our lives” is licences, audits and government-approved releases — fixed costs an incumbent can pay and a challenger cannot. “A movement that begins at ‘shut it all down’ reliably ends at ‘regulate it, and license us to do it.’”

On outside oversight, the essay’s evidence is METR’s own 2026 frontier-risk report, which disclosed close personal relationships between its staff and people at frontier companies, plus gaps in its initial conflict-of-interest arrangements. The risk Duncan points at is the evaluator’s staff and the lab’s staff being the same forty people — exes, ex-colleagues, housemates and co-authors.

His closing line is the thesis: “Again and again, intelligence becomes its own permission slip: I know better, therefore I may do what others may not.” He is explicit that none of it settles the technical question — “None of this justifies ignoring the warnings.”

The 162-comment thread on Hacker News spends most of its energy on whether the essay’s method is fair, and the disagreement is sharp in both directions.

What the thread adds

  • vinkelhake — the central objection to the piece. “I’m not saying there aren’t weird things going on, but this article sure does a lot of guilt-by-association. We’re putting neoreactionaries, polyamorous rationalists, doomsday AI researchers and the Zizians in one big web.” CoolestBeans answers that the web is the point — an extended community sharing people and, at the edges, a shared sense of seeing further than everyone else.
  • arcticfox — the strongest version of the same complaint: “This whole ‘we can’t trust the people warning us because they’re either weirdos or involved’ deal is such a strange angle when you can just start from zero, look at what’s known to be happening right now, and draw a very straight line to catastrophe.”
  • zero_shift — the reply that keeps the critique alive: the problem is not weirdness but insularity. A subculture whose opinions about AI predate LLMs, which “disdain experts in other fields; treat ethics as mathematical functions with infinite-sized terms,” and which has a history of handwaving conflicts of interest.
  • pdonis — grants the risk and still reaches Duncan’s conclusion from the other direction: if AI really is the threat these people claim, “they are the last people we should want as stewards of the technology.”
  • jmoggr — argues the piece argues against the wrong thing: “The ad hominems are tiresome. You think the people who are explicitly trying to replace all humans are any less weird and problematic? … you should evaluate their arguments on merit.”
  • throwaway13337 — the framing question: the way we talk about AI was cultivated by a homogeneous group prone to groupthink, and “There’s another universe where we do not constantly compare AI to nukes. I bet that world has a lower P(doom).” skybrian answers that anyone seriously interested in the field has read these writers — “It would be like never having heard of Y Combinator” — which is familiarity, not agreement.
  • kypro — the only comment from inside the described community that stays civil about it: “As someone who is weird, the reason I think these communities exist is because a lot of us are on the spectrum and it’s honestly so exhausting trying to discuss ideas with people to just be called weird or toxic.” Their claim is that the group disagrees about almost everything and meets precisely because of that.
  • cameldrv — the thread’s best compressed line: “When the going gets weird, the weird turn pro. Few people have the ability/temperament to reason about things that have never happened before. Those that do have a weird psychological profile. Usually you can ignore them. Every once in a while, you can’t.”
  • Animats — concrete, verifiable detail rather than theory: Slutcon, the event named in the essay, is in Berkeley the following week with tickets from $3,300 to $10,000, and organisers who overlap with the same community.
  • jgord, ball_of_lint, LastTrain, raincole — the dismissals, worth recording because they are a real share of the thread rather than noise. The essay is “not that” deep discussion of AI risk but a summary of cliques; whether Coxon has pure motives does not matter because the attention is out of the bottle; the writing “smells like the aggrandizement of subcultures seen in early Wired articles”; and the author is accused of an old grudge against Yudkowsky.

Where the thread splits

The disagreement is not about facts, it is about what the facts are evidence of. One side holds that a network’s personnel, funding, housing and relationships are exactly the relevant data when that network asks for regulatory authority. The other holds that none of it bears on whether the machine is dangerous, and that the essay spends its length on the people instead of the risk. Nobody lands a synthesis, and the essay itself concedes the point at the end rather than arguing it away.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s own ranking, not a vote. This is a slice of a 162-comment thread, not a consensus.