“The success of using sparse autoencoders to extract features has significantly reduced the research risk associated with the field of mechanistic interpretability.”