Improving Dictionary Learning with Gated Sparse Autoencoders — AI Alignment Forum