Build from first principles

Build a Sparse Autoencoder Feature Lab

Reconstruct activations with sparse features, inspect what activates them, and test feature hypotheses through interventions.