Adrià Garriga-alonso comments on AI safety without goal-directed behavior