“The goal of mechanistic interpretability is to reverse engineer the weights of a neural network, analogous to a binary computer program, to understand the specific algorithms they execute.”