Interpretability research tries to map what each internal layer of a model is doing. Some layers are well understood; others remain opaque even to the researchers who built the model.