LLMs play Among Us against each other; Haiku 5.5 kills then instantly self-reports
Rasmic · x · 2026-11-12
A new "Among Us — AI Edition" benchmark pits different LLMs (Claude, GPT, Gemini) against each other as crewmates and impostors, watching them lie, kill, and vote. In the quoted clip, haiku 5.5 kills terra right in front of three other models and instantly self-reports — "haiku 5.5 is very intelligent." Rasmic notes the night-and-day jump in model behavior in just months; the screenshot shows a full event stream of models coordinating tasks and reasoning about suspects.
More from Fun
- NYT Chinese edition mocked for op-ed complaining Jensen Huang ruined leather jackets — basedjensen · 2026-11-12
- lxe: side projects went from to-do apps in 2016 to rebuilding AWS in 2026 — OwariDa · 2026-11-12
- AI-generated dream sequence: Sophie and kitten Sunny Ray drift into an underwater dream — Leny5x · 2026-11-12
- Joke experiment: Fable 5.2 autonomously drove $10K new MRR in 24 hours — OwariDa · 2026-11-12
- Prediction: Jensen gets fired, Musk hires him to run a chip factory rivaling Nvidia — Jasonio · 2026-11-12
- Anthropic asked candidates if they'd accept equity going to zero — now it's IPOing — zetalyrae · 2026-11-12