Scaling And Evaluating Sparse Autoencoder