“Interpretability research on simple CIFAR-10 image classifiers has revealed that their internal filters learn complex representations like Perlin noise, a concept only recently characterized by humans.”