Towards an Algorithm for (Human) Self-Modification — LessWrong