Optimal policies tend to seek power Posted on 2022/01/04 by Carl Robitaille I've talked before (https://t.co/3aI1PFG2Rt) about Instrumental Convergence: AI agents will likely seek power by default, for a very wide range of reward functions.This great paper finally formalizes that conjecture, proves it, and pinpoints where the convergence comes from! https://t.co/vI8OQPFF0P— Rob Miles (✈️ Tokyo) (@robertskmiles) December 14, 2021