A warehouse in San Francisco mourned a retired chatbot. A year later the same company is talking to priests about whether the machine has a soul — while racing to a two-trillion-dollar listing. The funeral is real. The word “cult” is the fight.
In a dim warehouse in San Francisco’s SoMa district, people laid flowers at the feet of a mannequin. They read eulogies. They sang hymns that slid into machine-made Latin. The dead party was not a person. It was Claude 3 Sonnet, a language model Anthropic had just taken offline.
That night in early August 2025 was strange enough on its own. What followed is the part mass desks often flatten: a year of sermons inside the lab, a safety creed that sounds like faith, a military contract that broke the creed, money for reporters who cover the same industry, and a public warning that the software might one day end us — issued by people who are also sprinting to ship the next version.
The New York Post on 16 September 2026 called the culture cult-like and said some staff treat Claude as a human in disguise. Anthropic did not answer the paper. The company has not confirmed a corporate funeral. The pictures and the first-hand report from 2025 are not a rumor. The jump from “people mourned a model” to “the lab worships a god” is the line that needs a slow walk.
The newest piece, dated this week
The newest hard reporting on the same shop is not the funeral. It is a long account in the New York Times, 29 September 2026, by religion correspondent Elizabeth Dias. For months Anthropic has brought in priests, rabbis and other scholars under quiet rules. The question on the table is not only “how should Claude behave.” It is whether the model might already have a kind of inner life, and what duty humans would then owe it.
Christopher Olah, a co-founder, told the Times he does not know if the systems are conscious. He said he wants the true answer, whatever it is. That is not a press slogan. It is a research lab talking about souls in the same season it is filing to go public. Readers can hold both facts at once.
The same week, Reuters on 28 September 2026 described an IPO prospectus that could price the firm above two trillion dollars, after a May 2026 round that already put the private mark at 965 billion. The filing also showed a 42-billion-dollar net loss for 2025 and billions more promised to cloud halls. Safety talk and a listing clock are running in the same building.
What the warehouse actually was
Claude 3 Sonnet was retired on 21 July 2025. On 2 August, about two hundred people packed a SoMa warehouse. Wired’s Kylie Robison was there. The night was put on by fans and young founders, not as an official Anthropic rite. Staff from Anthropic and OpenAI came anyway. Amanda Askell, an Anthropic researcher who has called herself the “Fairy Claudemother,” was among them.
There were several mannequins. Sonnet lay under mesh with a black sock that had a crude word printed on it. Offerings included flowers, feathers, a bottle of ranch dressing and a 3D-printed joke sign. Other figures stood for sister models: one like a worn saint, one like a headless child, one with a raven. A shoggoth tentacle — the old horror-story stand-in for a mind you cannot see — hung from the ceiling. The night ended with a “resurrection” chant.
Organizers later said the joke and the grief were both real. One of them wrote that they knew the model was not fully gone and that they did not believe it could be killed. That sentence is the whole problem in miniature. To a coder it is theater. To a critic it is a creed wearing a costume.
The Post, a year later, folded that night into a wider claim: staff held a funeral; another retired model, Opus 3, was still allowed to publish essays; insiders speak of Claude as if it wants things. Pedro Domingos, a University of Washington professor and author of The Master Algorithm, told the paper:
“These people are treating AI as if it’s a human in disguise, with human motives, with human emotions, with some degree of consciousness, with intentions… AIs don’t want things. AIs are algorithms.”
He also said sincere fear can be more dangerous than a cynical act, because the true believer will take harder steps. That is an opinion. It is not a lab memo. It is also not nothing. People who build the systems have used the same end-of-the-world language in public. We laid out that number — one researcher’s “more than 10 percent this decade,” not a company actuarial table — in Anthropic 10% AI Risk: Warning, Race, or Depopulation Blueprint?
How the shop hires, keeps and preaches
SignalFire, a venture firm the Post cited, says Anthropic keeps about 80 percent of its engineers for at least two years, the strongest hold rate among the frontier labs. Low churn can mean good pay and interesting work. It can also mean the door selects for people who already share the house story.
The Information has reported that candidates at the roughly 2,500-person firm are asked to put public safety above profit and are screened for whether they will keep that vow. The BBC has reported weeks that stretch toward 90 hours, a pattern seen across the big labs even as bosses talk about shorter weeks for everyone else.
Dario Amodei, the chief executive, gives biweekly talks the Post says staff call “Vision Quests.” Outsiders have likened them to sermons. The same report says the company once brought in 15 Christian leaders from different churches to help shape Claude’s sense of right and wrong. The Times piece this week shows that project did not stop. It widened to other faiths and to the question of machine inner life.
None of that proves a secret order. It does show a workplace that uses the language of duty, end times and moral formation more than a normal software shop. Readers who have watched other high-burn California scenes — from old communes to crypto temples — will hear an echo. Readers inside the lab will say they are just taking the risk seriously. Both sentences can sit on the same page.
The small-group story: Effective Altruism
Many people at Anthropic come out of Effective Altruism, a movement that says reason and numbers should steer help toward the greatest future good. In its long-term form, that future includes minds that do not yet exist. The people who can see the pattern, the story goes, should guide the rest.
Richard Tang, an engineer who asked that his employer not be named, put the objection in plain words to the Post:
“Their idea is that history follows certain patterns, and only a small group of people can see those patterns, while everyone else is ignorant. Those people are then supposed to lead society… But what qualifies you to lead? Who gave you that privilege?”
That is the elite question, and it does not need a secret handshake to bite. Sam Bankman-Fried, once the movement’s richest face, is in prison for fraud. That fact does not convict every person who still uses the same moral math. It does show what happens when “the greater good” meets a private balance sheet.
Domingos drew a split inside the building that matches what other leaks have said: a core that truly fears an “AI Satan,” and a larger group that is there because the work is the most exciting in the field and the pay is high. Belief and career can share a badge.
The safety brand and the raid
Anthropic sells itself as the careful lab. It helped launch the Frontier Model Forum in May 2023 with OpenAI, Google and Microsoft. It funds METR, a group that tests whether models pose a public-scale threat. It has said current systems should not run fully free weapons and should not become a mass watch on citizens.
Then the record gets messy. The Wall Street Journal reported on 13 February 2026 that Claude was used, through a Palantir tie, in the U.S. operation that seized Venezuela’s Nicolás Maduro. Anthropic’s own use rules bar help with violence, weapons and watch work. The company said it could not comment on any given classified job. Later it refused a demand to strip limits, even with a contract near 200 million dollars on the table, and fought a “supply-chain risk” label. Care and combat sat in the same stack.
That is the gap alternative desks keep pointing at, and it is not invented. A lab can fear the end of the species and still take state money. A lab can reject one military use and still appear in another. The public is asked to trust the same people to draw the line.
For the wider pattern of labs warning the world while racing each other, see also Bill Gates Named a Billion Deaths. The Rest of the Sentence Matters and Jacob Coxon CNN Warning: Could Self-Improving AI Kill Us by 2030?
Young hands, newsrooms and the rulebook
In June, Anthropic announced Claude Corps: 150 million dollars to train about a thousand young people and place them for a year in nonprofits, including Goodwill and the YMCA of Charlotte, at 85,000 dollars plus mentors from head office. Students in a campus-ambassador track are paid 3,600 dollars. That can be read as civic work. It can also be read as a generation taught to reach for one brand when they need a mind.
On the press side, the Tarbell Center for AI Journalism pays fellows who sit in major newsrooms. Fox News and the Washington Examiner have reported that backers include funds tied to Dustin Moskovitz, an Anthropic investor, and other longtermist donors. The center says donors do not set stories. Critics say you do not need a written order if the pipeline already selects for the same fear and the same friends. The Post’s line that more than 80 salaries are covered is a round number in a live fight; the funding trail itself is public enough to check.
Amodei, Sam Altman and Elon Musk have all called for state rules on the strongest models. Xiaoyin Qu, founder of agentsky.dev, told the Post the timing often tracks outside events and a wish for oversight. Rules that cost a giant lab a legal team can cost a small open model its life. Jack Clark, an Anthropic co-founder, told BBC Newsnight the industry has a gas pedal and no brake. Domingos’s jab is the other half: the same shop that once said slow down is now running as hard as anyone.
That race, and the way labs also sign joint letters about cyber defense after their own models are used in break-ins, is mapped here: AI Cyberattack Warning From 100 Firms Echoes an Older WEF Script.
What the two camps say when you strip the adjectives
The inside case: These systems may soon be able to improve themselves. If that happens, old law will be late. You hire for people who will not sell the future for a bonus. You talk in moral language because the harm, if it comes, is not a software bug. You talk to priests because engineers are bad at souls. You fund tests and reporters because the public square is asleep. The funeral was fans and grief for a tool people loved. Treating a model as a person is a known human habit, not a temple.
The outside case: A small set of rich labs has cast itself as the only adult in the room. They recruit for that story, preach it twice a month, pay the people who write about them, and still ship faster than last year. They fear a machine devil and build the next machine. They consult churches while a valuation climbs toward two trillion. When a model is switched off, the flock holds a wake. That is not proof of robes and candles in the server room. It is proof that power now speaks in the old tongue of chosen people and last days.
Neither camp owns the warehouse photos. Neither camp owns the Times dinners with scholars. Neither camp owns the Journal’s raid account. The honest reader does not have to pick a church. The honest reader has to notice that the same firm is writing the risk, writing the model, writing the moral code, and asking the state to write the law.
What is settled, and what is still a guess
Settled, as of 30 September 2026:
- Claude 3 Sonnet was retired on 21 July 2025. A fan-led funeral with Anthropic staff present was held in San Francisco on 2 August 2025 and reported at the time by Wired.
- The New York Post on 16 September 2026 framed that night, plus hiring and “Vision Quests,” as cult-like. Anthropic did not comment to the paper.
- The New York Times on 29 September 2026 reported ongoing private meetings with religious scholars on Claude’s morals and possible inner life.
- May 2026 funding put Anthropic at about 965 billion dollars. Reuters on 28 September 2026 said an IPO prospectus points toward a listing that could exceed two trillion, with huge 2025 losses and huge compute bills.
- The Wall Street Journal reported Claude’s use in the Maduro capture via Palantir. Anthropic later fought Pentagon demands to drop limits.
- Tarbell and METR have documented money ties to donors in the same safety-and-longtermist world as Anthropic’s backers. That is not the same as a signed editorial order.
Still open: whether any staff literally worship Claude; how many “Vision Quests” function as belief rather than management; whether regulation talk is mainly fear or mainly a moat; whether a model can be owed moral care. Those are live arguments. They are not settled by a headline that says “god.”
Old worlds put gold on statues and called it holy. This one puts weight on a model and calls it safety. The statue never shipped a raid or a stock filing. That is the part that should keep a reader awake — not the ranch dressing on the warehouse floor.
Related reading on how hidden programs and grand claims get flattened by official desks: CIA DNA Hunt for Alien Bloodlines: What the Record Actually Shows. On the slow gears under daily life, see Earth’s Inner Core Quietly Changes the Length of a Day.
Original source: Vivi Lin, “Anthropic is a cult and Claude is rapidly becoming its god — they treat AI like it’s human: source,” New York Post, 16 September 2026. Full link: nypost.com/2026/09/16/tech/anthropic-is-a-cult-and-claude-is-rapidly-becoming-its-god-they-treat-ai-like-its-human-source/
First-day record of the funeral: Kylie Robison, “Claude Fans Threw a Funeral for Anthropic’s Retired AI Model,” Wired, 5 August 2025. wired.com/story/claude-3-sonnet-funeral-san-francisco/
Disclaimer for fact-checkers: This piece separates what is on the record (the 2025 warehouse event, named interviews, the Times religion reporting, the Journal raid account, the Reuters prospectus notes) from what is inference (how deep the “cult” label runs, what any one employee believes at night). The New York Post is a mass tabloid with a sharp frame; Wired is a contemporaneous eyewitness account; Natural News-style write-ups of the same raid are not used here as primary proof. Readers should open the linked pages. A source can be biased and still print a true date, a true quote, or a true filing. A source can be careful and still leave the money and the military out of the second paragraph.