Why Simulator AIs want to be Active Inference AIs — LessWrong