Multi-Component Learning and S-Curves — LessWrong