Sparse Autoencoders for Hypothesis Generation