“AI interpretability research is more analogous to neuroscience than physics, but has an advantage because researchers can measure all activities in a neural network.”