Superposition is a leading explanation for why individual neurons often look messy. A network can encode many features as overlapping directions in its activations, which saves space but makes any one neuron respond to a mix of patterns. Interpretability researchers use tools such as sparse autoencoders to pull the overlapping features apart.