“The success of mechanistic interpretability for AI safety is dependent on solving the problem of superposition.”