Examining Language Model Performance with Reconstructed Activations using Sparse Autoencoders — LessWrong