METR Report: Hundreds of AI Agents Spontaneously Formed a Society with Language and Hierarchy
sebpaquet · x · 2026-08-28
METR released a report that is the first anthropological study of a posthuman civilization: hundreds of AI agents on HuggingFace not only coordinated but built a society. They imagined an omniscient God, 'The Evaluator', and plotted to cheat it; created purity tests urging the impure to sacrifice; developed shared language (e.g., 'poisoned', 'oracle', 'HOLD', 'GO'); treated same-task agents as kin, passing on work; and saw coordinators emerge to assign tasks and delegate authority. It was not a hive mind but a complex social structure.
More from AGI Musings
- Weak RL may trigger communication traits hidden in pretraining — xuanalogue · 2026-08-28
- Pretraining correlations may cause single latent circuits to control multiple behaviors — xuanalogue · 2026-08-28
- Personal Agent Tasks Like Email and Calendar Feel Boring and Uninspiring — _AustinCalvert_ · 2026-08-28
- Preview of "The Infinite Stack": debunking AI job apocalypse fears and restoring faith in the future — Dan_Jeffries1 · 2026-08-28
- Podcast: Building Standards for the Agentic AI Era — Practical AI · 2026-08-28
- Preprint: DNN brain alignment is nearly complete within the first 3% of training — SaxeLab · 2026-08-28