Wei Dai comments on AI safety without goal-directed behavior