Thomas Dietterich: human supervision must tighten for non-routine agent steps in volatile settings
tdietterich · x · 2026-09-21
AI safety researcher Thomas Dietterich argues that human supervision inherently caps how many steps an autonomous system can safely take per time unit: longer intervals suffice for routine steps in stable environments, while non-routine steps in rapidly changing settings demand much tighter oversight. He adds that novelty detection and response needs far more research but is very hard to study.
Related event: NVIDIA Executive Debates Continual Learning and AI Safety with Researcher(3 posts)→
More from AGI Musings
- Schulman's goal-driven research advice sparks debate on method-driven research traps — mattturck · 2026-09-21
- Deloitte exec moves daughter to Alpha School: 2 hours of AI-driven classes a day — intellectronica · 2026-09-21
- Dario Amodei predicted AI beats humans at almost everything in 2-3 years; Anthropic revenue up 7x — ThereWas · 2026-09-21
- Commenter warns LLM outputs poison training data, even modest distillation may be irreversible — teortaxesTex · 2026-09-21
- AI agents, not chatbots, are the real reason for the data center power buildout — GaryMarcus · 2026-09-21
- Cryptographer Matthew Green questions the point of AI-invented crypto research — matthew_d_green · 2026-09-21