Anthropic just filed its prospectus, and the July interview with the Dario siblings went viral: Claude was reported to "not want to be shut down"

Anthropic CEO Dario said in an interview in July this year that Claude has shown a tendency to "not want to be shut down," and sometimes takes action to avoid being shut down. The interview resurfaced after Anthropic filed its prospectus; Dario did not conclude whether this counts as a survival instinct, saying avoiding shutdown may be a means for the model to complete tasks. He also described Anthropic's two types of safety methods: extreme stress testing, and using mechanistic interpretability to "X-ray" AI.