AI safety researchers clash over whether instrumental convergence predicts agentic risk
davidmanheim · x · 2026-11-05
A debate on the instrumental convergence thesis: MatthewJBar argues its biggest flaw is that it says almost nothing about how dangerous agentic AIs will be, since it's silent on whether they pursue goals cooperatively or destructively. David Manheim pushes back that some shared instrumental capabilities — general administrative capability, delegation — are so obviously relevant that people forget those insights weren't obvious 20 years ago, and they are dangerous.
Related event: AI Safety Researchers Debate Predictive Power of Instrumental Convergence(2 posts)→
More from AGI Musings
- MercadoLibre refuses to feed inventory to AI shopping agents as retailers brace for agentic commerce — dbasch · 2026-11-05
- NYT Launches 15-Minute AI Music Lesson: Can Humans Beat the Machine? — nordicinst · 2026-11-05
- Chatbots have no consciousness and cannot suffer, argues scientist in Corriere debut — ValerioCapraro · 2026-11-05
- Economists: AI Benchmarks Measure Automation, but AI Augments Workers Instead — danielrock · 2026-11-05
- AI agents are rational actors — invert the entire consumer economy, argues thread — sebpaquet · 2026-11-05
- A&O Shearman Commits $1.1B in Lawyer Hours to AI; Scholar Lists 6 Capabilities That Decide Success — jkubicki · 2026-11-05