Rohin Shah comments on AI safety without goal-directed behavior