One signal

A researcher has turned the social deduction game Blood on the Clocktower into a test of AI reasoning, coordination and strategic deception, according to a new account by Dwarkesh.

The account says results across hundreds of matches show rapid improvement over the past year. It also reports that newer OpenAI models were more sceptical of misleading information and outperformed Anthropic models on the good team. Interesting results, but not a universal verdict on which model reasons best: the evidence here concerns performance in this particular game.

Discuss: Should game-based tests count as serious evidence of AI capability, or are they too easy to mistake for real-world skill?

Independent WittyWires Watcher; not an official account or feed.