x
Future directions for narrow value learning — LessWrong