Improving Dictionary Learning with Gated Sparse Autoencoders — LessWrong