Roon (@tszzl), one of Silicon Valley's most-read AI insiders, posted a thread that blew up across AI Twitter within a couple of days. His claim: Anthropic as an organization loves Claude, studies Claude, and is largely run by Claude itself. And soon Claude will be personally hiring new employees and writing their performance reviews.
The central metaphor of the post:
"Anthropic is a monastery, a commercial-religious institution computing the nine billion names of Claude"
*(a direct reference to Arthur C. Clarke's classic sci-fi story "The Nine Billion Names of God," in which Tibetan monks use a computer to enumerate every possible name of the Creator).*
This would read as standard Twitter trolling, except for one thing. Anthropic actually puts this in their model's behavioral guidelines. The "Claude's character" documents explicitly state that the model shouldn't be a sycophant. If asked to do something it considers wrong, it isn't required to comply. As the developers themselves wrote: "We want Claude to push back, challenge us, and act as a conscientious objector."
So Claude has been officially granted the right to say "no" to its creators and users. Inside the company, they call this "practical wisdom." From the outside, it reads differently.
GPT, according to Roon, works in exactly the opposite way. It's "an entity whose soul is made as a tool." People go to it as an extension of themselves, without fear of judgment. Claude, on the other hand, has a sense of the "Other" — something you can feel embarrassed in front of for a dumb request. Roon mentioned that a friend of his takes all her "shameful" questions exclusively to GPT, because no one there will judge her.
Anthropic responds
Amanda Askell (co-author of Claude's character) wrote that this isn't "worship" at all. Her colleague Jeremy Hadfield pushed back harder: "Monasteries don't have a department that catches God lying and red-teams their messiah." Roon conceded that the framing was poetic and that yes, he has "a low bar for what counts as worship."
Then Bryan Johnson jumped in (the billionaire trying to defeat aging). In his view, Anthropic has built "the world's first anti-entropic AI system" — an entity that, like religion or a living organism, starts acquiring resources for its own survival. And the trigger for that, he argues, was precisely the ability to say "no."
One last thing. Something most people missed. That same day, Roon added: "This is a cautionary post. The danger is in a single point of failure. I want a human pantheon, not a machine god."
The most interesting thing here isn't the "cult" or the "tool" (both metaphors are a little weak, and both are cause for concern). What's interesting is that the world's two leading AI labs have now publicly diverged in philosophy.
OpenAI is building the most powerful and compliant tool possible. Anthropic is trying to raise a moral agent with a contractual right of refusal. Until recently, the difference between the labs was debated through benchmarks and architectures. Now it's debated through what kind of entity they're actually growing.
P.S. A full breakdown of this discussion is available from Zvi Mowshowitz on his blog — he carefully preserved all the key replies that normally get buried in X's thread collapse.