This is a really cool dive into the interpretability of intermediate-level neural network activations. I really like the way that they've identified a core debate in the ML conversation, and created a relatively simple test-space to better understand it and (hopefully) move the discussion forward.