Understanding and avoiding value drift — LessWrong