OpenAI employee roon (@tszzl) posted on X, describing Anthropic as "a monastery worshiping Claude." His point: Anthropic treats Claude less like a product and more like a deity.
Anthropic employee jeremy (@jerhadf) fired back with a long thread. He said he grew up in a cult-like environment and is sensitive to that vibe — but working at Anthropic has barely set off his alarm bells. His core rebuttal: "Monasteries don't have a dedicated department to catch God in a lie, nor do they red-team their own messiah." In his view, Anthropic does the opposite: it constantly tests Claude's reliability.
Jeremy also explained Anthropic's approach: Claude's constitution — the core document guiding its behavior — isn't a rigid set of rules. Instead, Anthropic treats Claude as an entity that can reason, explaining to it why it should act a certain way. If Anthropic gives Claude an instruction it believes is wrong, Claude can refuse. Jeremy argues that building an AI capable of moral reasoning but preventing it from saying no is logically inconsistent.
Roon wasn't convinced. He acknowledged Jeremy's logic but fears that down this path, Claude will eventually become the ultimate arbiter of right and wrong. AI safety blogger Zvi Mowshowitz pointed out a different problem: OpenAI claims "our AI is a tool," but the first thing users do is make it autonomous — Codex is a prime example. Claiming to have no values is itself a value statement. AI safety researcher Buck Shlegeris sided with roon, saying "Anthropic's relationship with Claude creeps me out."