The Geometry of Concepts: Sparse Autoencoder Feature Structure