2026년 9월 Anthropic 연구원 제이콥 콕슨이 'AI가 10년 안에 인류를 멸망시킬 수 있다고 진심으로 믿는다'는 사직서를 남기고 주식 보유 기간을 채우기 2개월 전에 퇴사하며 AI 안전 경고가 공론화되었습니다. 이 기사는 이러한 종말 경고가 합리주의(rationalist)라는 특정 인터넷 서브컬처에서 비롯되었으며, 그 커뮤니티의 인물들이 AI 연구소와 안전 기관의 핵심을 장악하고 있다는 점을 분석합니다.
번역된 본문
AI 안전과 세계 종말에 대한 우려가 최근 당신의 삶에 들어왔다면, 그것은 아마도 한 통의 사직서 때문일 것이다. 2026년 9월 8일, 제이콥 콕슨(Jacob Coxon)이라는 연구원이 Anthropic을 그만두었다. Anthropic은 미국의 양대 최고 AI 연구소 중 하나로, 경쟁사에 비해 '신중한 기업'이라는 이미지를 쌓아온 곳이다. 그는 이전 3년간 OpenAI와 이후 Anthropic에서 사전학습(pretraining) 연구를 했는데, 이는 대형 모델을 유능하게 만드는 엔진실과 같은 작업이다. 그의 작별 인사에는 이런 내용이 담겨 있었다: "AI를 만드는 사람들은 그것이 이번 10년이 끝나기 전에 우리 모두를 죽일 수 있다고 진심으로 믿는다. 이것은 마케팅 수사가 아니다." 그는 스톡 옵션이 베스팅되기 두 달 전에 회사를 떠나며 그 권리를 모두 포기했다. 이 게시물은 하루 만에 1억 회 이상 조회되었다. 놀랍게도 동료들은 그를 정정하러 나서지 않았다. Anthropic에서 정렬(alignment) 과학을 이끄는 에번 하빙어(Evan Hubinger)는 몇 시간 안에 콕슨이 옳다고, 자신들도 정말로 AI가 전 인류를 죽일 수 있다고 믿는다고, 그리고 10년 내 그 확률이 개인적으로 10분의 1을 넘는다고 밝혔다. 20명 이상의 의회 의원들이 대책이 필요하다고 말했다. 자체 프론티어 연구소를 운영 중인 일론 머스크는 이 모든 것을 심리전(psy-op)이라 불렀고, 주류 언론은 예상대로 이어진 광풍을 숨 가쁘게 보도했다.
이 경고의 심각성을 현실적으로 평가하려면, 경고를 하는 사람들을 이해해야 한다. 세계 종말에 대해 가장 큰 목소리를 내는 이들은 매우 특정한 25년 된 인터넷 서브컬처에서 나오는데, 이들은 자신들을 다른 모든 사람보다 더 명확하게 사고하는 공동체로 여긴다. 실리콘밸리 밖에서 이 서브컬처는 잘 알려지지 않았고, 유난히 괴짜스러운 문학적 취향을 가진 팬덤처럼 보인다. 하지만 기술 업계 내부에서는 결코 그렇지 않다. 이들의 작가들은 명성을 좌우하고, 이 커뮤니티 출신들이 연구소와 안전 연구소, 그리고 'AI 안전'의 의미를 정하는 자선 단체의 인력을 채우고 있다. 이들의 논쟁, 결혼, 불화, 성적 취향은 지금 종의 미래에 대해 대중의 신뢰를 요청하는 조직들의 구조와 깊게 얽혀 있다.
콕슨이 사직하기 10개월 전인 2025년 12월, 거의 500명이 버클리의 한 공연장에 모여 인류의 가능한 종말에 관한 의식에 참석했다. 라이브 밴드와 28인 합창단이 있었고, 전례 문낭독도 있었다. 노래와 연설이 인류의 먼 과거에서 불확실한 미래로 향하면서 샹들리에가 하나씩 어두워졌다. AI를 만들거나 통제하려는 조직의 직원들이 대거 참석했다. 가장 어두운 순간, 주최자는 무대 끝에 앉아 떨리는 목소리로 청중에게 말했다: "여러분… 저는 우리가 해내지 못할 것 같습니다."
주최자는 레이먼드 아놀드(Raymond Arnold)로, 웹 개발자이자 게임 디자이너, 음악가, 그리고 합리주의 커뮤니티 기관의 오랜 설립자다. 가톨릭과 무신론 사이에서 자란 아놀드는 청년기에 합리주의 운동을 접하고 그것이 신성한 감정을 '급진적으로 확장'해 주었다는 것을 발견했다. 2011년 그는 19명의 친구들을 뉴욕의 한 아파트로 초대했는데, 그 저녁에는 등유 램프, 플라즈마 볼, 라이트세이버로 장식되어 있었다. 새로운 세계관을 위한 명절을 베타 테스트하기 위해서였다. 그들은 첫 겨울 모닥불과 인류의 미래를 형성할 기술들에 대해 노래하고, 마지막 촛불을 끄고, 함께 어둠 속에 앉아 있었다. 2년 뒤 아놀드는 이 실험을 '오늘보다 밝게(Brighter Than Today)'라는 공개 행사로 발전시켰다. 이 행사는 곧 뉴욕과 베이 지역에서 북미, 유럽 등 전 세계 도시로 퍼져나갔다. 아놀드는 처음에 세속적 지구 동지하절(Secular Solstice)을 폭넓은 휴머니즘 명절로 상상했지만, 결국 합리주의 달력의 최대 명절이 되었다. 이것은 전적으로 우연은 아니었다. 그는 2011년부터 이 커뮤니티를 조직해 왔으며, 현재 LessWrong과 라이트헤이븐(Lighthaven) — 버클리에 있는 컨퍼런스 캠퍼스로서 그 커뮤니티의 물리적 거점 역할을 하는 곳 — 의 운영을 함께 돕고 있다.
If worries about AI safety and the end of the world entered your life recently, it’s probably because of a resignation letter. On September 8, 2026, a researcher named Jacob Coxon quit Anthropic, the one of the two premier AI labs in the U.S. whose brand has tried to cultivate the image of being the careful one relative to its competitors. He had spent the previous three years doing pretraining research at OpenAI and then at Anthropic, the engine-room work that makes large models capable, and his goodbye note said, among other things: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.” He left two months before his equity vested and walked away from it . The post was viewed more than a hundred million times within a day. His colleagues, surprisingly did not rush to correct him. Evan Hubinger , who leads alignment science at Anthropic, posted within hours that Coxon was right, that they really do earnestly believe AI could kill all humans, and that his personal odds were somewhere north of one in ten inside a decade. More than two dozen members of Congress said something needed to be done. Elon Musk, who is building a frontier lab of his own, called the whole thing a psy-op , and mainstream news outlets predictably gave breathless coverage to the frenzy that followed. In order to realistically evaluate the gravity of this warning, we need to understand the people who are making it. The loudest voices warning about the end of the world come out of a very specific twenty-five-year-old internet subculture, one that understands itself as a community of people who think more clearly than everyone else. Outside Silicon Valley the subculture looks obscure, a fandom with unusually dorky literary preferences. Inside the technology industry it is anything but. Its writers make or break reputations. Its alumni staff the labs, the safety institutes, and the philanthropic organizations that decide what “AI safety” means. Its arguments, marriages, feuds, and sexual proclivities are heavily interwoven into the structure of the organizations now asking for the public’s trust about the future of the species. Ten months before Coxon resigned, in December 2025, nearly five hundred people gathered in a Berkeley music hall for a ceremony about the possible end of humanity . There was a live band and a twenty-eight-person choir. There were liturgical readings. As songs and speeches carried the room from humanity’s distant past toward its uncertain future, the chandeliers dimmed one by one. Heavily in attendance were employees of organizations trying to build, or contain, artificial intelligence. At the darkest point, the organizer sat at the edge of the stage, his voice breaking, and told the room: “Guys… I don’t think we’re gonna make it.” The organizer was Raymond Arnold, a web developer, game designer, musician, and longtime builder of rationalist community institutions. Raised between Catholicism and atheism, Arnold discovered the rationalist movement as a young adult and found that it “radically expanded” his capacity for sacred feeling. In 2011 he invited nineteen friends into a New York apartment furnished for the evening with oil lamps, plasma balls, and lightsabers to beta-test a holiday for their new worldview. They sang about the first winter campfires and the technologies shaping humanity’s future, extinguished the final candle, and sat together in the dark. Two years later Arnold turned the experiment into a public production called Brighter Than Today . Versions soon spread from New York and the Bay Area to cities across North America, Europe, and beyond. Arnold had initially imagined Secular Solstice as a broad humanist holiday, but it became the rationalist calendar’s principal feast day. This was not entirely accidental. He has organized the community since 2011, now helps run both LessWrong and Lighthaven—the Berkeley conference campus that serves as its physical center—and has described his function, earnestly, as that of a “village priest.” Solstice was his attempt to construct the cultural technology a durable movement needs: a canon of songs and readings, a ritual progression from light into darkness and back again, and a yearly occasion for geographically scattered believers to gather in one room. By 2025, Arnold estimated the probability of an AI catastrophe at greater than fifty percent and worried that his own ritual was no longer equipped to contemplate extinction in a psychologically healthy way. He had largely ceded Solstice production to other organizers, but returned to shape what he thought might be “the last Solstice,” complete with an aftercare gathering around a firepit. The evening supplied everything religions have traditionally supplied: fellowship, ritual, a shared story about humanity’s place in the universe, and somewhere to sit with the knowledge that you are going to die. Possibly soon. Possibly because of something your friends built. Follow the connections out from that room and you pass through some genuinely strange territory. A Harry Potter fanfic used as a recruiting pipeline. Psychological workshops that former participants describe as coercive. A program for breeding smarter children, promoted by an AI institute. A sex worker from Idaho who became one of the subculture’s most-read writers and now runs an AI-doom propaganda residency with Grimes, Elon Musk’s ex-girlfriend. A splinter sect that fled to tugboats and whose members have been linked to six deaths. And a blogger who wanted to abolish democracy, found the ear of several billionaire patrons, and watched his ideas come to fruition in the vice president’s office. I am aware, that if you are not lurking on the internet as much as I do, of how utterly insane that paragraph sounds. However, every bit of it is heavily documented by numerous sources. The people in this story are not a secret society coordinating over dinner. 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. One of many unfortunate shortcomings of mainstream media is that credibly reporting on these things is so absurd that you can’t expect readers to take it seriously. However, I’ll endeavor to provide a thorough account of how our current slide into neo-fascism, theocratic dystopia, and the forecasted end of the world is deeply rooted in the history of this subculture. So, let’s go back to the yeeeeaaaar two thousannnnd. The prophet We’ll start with Eliezer Yudkowsky , who had developed a formidable sense of his historical importance years before any contemporary AI company existed. Yudkowsky was born in 1979 and grew up in a Modern Orthodox Jewish family. By his own account he wrote off the entire institution of public education in second grade, the year he discovered his math teacher did not know what a logarithm was. He read Feynman’s QED at nine and says he understood most of it. He counts his time in school as ten lost years, dropping out around eighth grade, partly for health reasons. No high school, no college, no formal training of any kind in the field of AI. At eleven a grand-uncle handed him Great Mambo Chicken and the Transhuman Condition , a 1990 chronicle of cryonicists, space eccentrics, and people planning to upload their minds into computers. His own summary of the encounter is brief: after that he knew what was important, and that his career would touch it. In his late teens he found the transhumanists online. By 2000, at twenty, he had co-founded the Singularity Institute for Artificial Intelligence, and his website was already offering the complete package: an essay about the Singularity, a document sketching how you might actually code a transhuman AI, a plan for getting humanity from here to there, and an autobiography explaining the author’s singular q