Understanding goals in complex systems — LessWrong