Computational complexity of RL with traps — AI Alignment Forum