Scaling and evaluating sparse autoencoders