2026-09-20 Tao’s Nonlinear AI Warning

terence-tao-slow-down-ai — “We have to slow down AI”

Terence Tao argues that AI development is moving unjustifiably fast through “extremely nonlinear dynamics”; the revealing part is how replies translate his systems-risk point into a simpler fight over progress, professional relevance, and whether slowing down is even possible.

ezra-klein-ban-recursive-ai-self-improvement — Ban recursive AI self-improvement

Ezra Klein argues that a general slowdown is not enough: labs should be barred from letting AI systems build their successors until they can prove that recursive self-improvement will remain under meaningful human control.

roboharm — Do robot policies refuse unsafe instructions?

RoboHarm tests frontier policies on five physically harmful tasks and finds that capable robot agents usually attempt them: Claude Fable refused 20 of 100 trials, GPT-6 Astra only 2, while MolmoAct2 lacked a refusal mechanism altogether.

victory-by-any-means — AI x-risk and arms races

Victory By Any Means frames AI existential risk as an arms-race and coordination problem, where fear of losing pushes even risk-aware competitors toward dangerous capability development.