“Superposition obscures the implicit modularity of neural networks, making it difficult to observe in the later layers of vision models and throughout language models.”