Ultimate ends may be easily hidable behind convergent subgoals — AI Alignment Forum