As OpenAI's research VP he grew uneasy about AI safety and commercial direction, leaving with a group of colleagues to found Anthropic in 2021 focused on safe, steerable AI.