AI safety without goal-directed behavior — AI Alignment Forum