Sparse autoencoders find composed features in small toy models — LessWrong