Sparse autoencoders push past the toy-example era. What the features look like, what they tell us, and where the method breaks.