“A key challenge for interpretability is to develop larger-scale abstractions for neural networks, analogous to organs or organ systems in biology.”