Mesa-optimization for goals defined only within a training environment is dangerous — LessWrong