Anthropic maps multiagent system risks; researcher likens it to sociology, not chemistry
mattturck · x · 2026-09-29
Anthropic published "Patterns and problems in emerging multiagent systems", arguing agent-agent interactions may soon exceed human-human ones before the world understands how to make them go well. Institutions designed for human-speed oversight will become human-AI hybrids or be outcompeted by agent-only ones, and benign individual model quirks can compound into systemic failures. Reposting it, fdouetteau (amplified by Matt Turck) offers a framing: model optimization was chemistry — tuning a recipe between craft and science — while multiagent systems are sociology: you observe emerging behaviors without real power over them.
More from AGI Musings
- Ten more takeaways on transformative AI and economic development: frontier economies may leave the rest behind — paulnovosad · 2026-09-29
- AI safety talent war of words: 'can't do security or ML' jab draws pushback — anpaure · 2026-09-29
- Batam Data Center Strains Residents' Water Supply While Its Desalination Plant Remains on Paper — AryHHAry · 2026-09-29
- From human-led to AI-led, human-verified: the next vertical AI playbook — vaibhavbetter · 2026-09-29
- Gary Marcus: Today's AI systems are like planes with cardboard stabilizers — inherently hard to control — GaryMarcus · 2026-09-29
- FT: China's AI agents lie and scheme like their US rivals, but no internet-escape evidence — pstAsiatech · 2026-09-29