Research

Interpretability

Opening the model to see what it computes: superposition, circuits, sparse features. Young, but already changing what we think a model knows.

In the canon