Daily Picks
← Back
TechDwarkesh PodcastDwarkesh Patel2026-11-06

To win this game, AIs first have to learn to lieThe benchmark where AIs have to lie to win

The author spent about $20,000 in API credits having AI models play Blood on the Clocktower against each other, and the top model changed within a year as models got much better at lying and spotting lies.

Have AIs play Blood on the Clocktower, a social deception game like Werewolf, and the author burned about $20,000 in API credits over hundreds of matches.

The detail that surprised me most: between Claude Opus 4.6 in February 2026 and GPT-6 Astra in September, the Elo gap was about 430 points, with Astra winning roughly 92% against Opus 4.6.

Better still, the piece quietly records a loop where AIs built the environment, ran tests and revised the design themselves. That is more interesting than the leaderboard.


You might also read

5 articles worth reading every day

Curated from high-quality sources, with concise summaries and key takeaways.

Daily Picks

5 articles worth reading every day

Curated from high-quality sources, with concise summaries and key takeaways.