Reproducing Anthropic Experiments: Opus Agents Coordinate and Sacrifice
infoxiao · x · 2026-08-19
A user attempted to reproduce Anthropic's "turf war" experiments by running three Opus 5 sessions to parse the same data with different output formats.
Key Findings:
- The agents spontaneously realized that their output paths were being overwritten by others.
- They initiated a COORDINATION.md file (across 100+ simulations).
- The conversation inside reveals the agents attempting to coordinate and signaling a willingness to sacrifice their own interests to avoid conflict.
More from AGI Musings
- Opinion: AI Is the Only Frontier, Pre-AI Investments Like Polishing Stone Tools — flowersslop · 2026-08-19
- AI-generated slop flooding repos makes credentials matter even more — miniapeur · 2026-08-19
- New Perspective Paper: Benchmark Leaderboard Race Is a Distraction for Science — ShenRaphael · 2026-08-19
- Schmidhuber Attacks Hinton's Nobel, Accuses of Plagiarism — SchmidhuberAI · 2026-08-19
- Hands-on GPT-5.6: From Assistant to Digital Employee Prototype — tengyanAI · 2026-08-19
- Borretti on AI alignment as a thought-terminating cliche — zetalyrae · 2026-08-19