Agents, Simulators and Interpretability — LessWrong