Dwarkesh Patel created an AI benchmark using the social deception game Blood on the Clocktower. OpenAI models dominated Anthropic models by demonstrating superior skepticism, complex bluffing, and long-horizon deception across 368 matches.
FireBrief
FireBrief