Toy Models of Superposition

Toy Models of Superposition asks how a neural layer can track more useful features than it has hidden dimensions. Read it to see how sparse features share non-orthogonal directions, why ReLU makes interference affordable, and why the clean geometry remains a controlled clue rather than a map of a real language model.

Reading focus: How a tied-weight ReLU autoencoder turns features into directions through a narrow hidden layer. Why sparse features can share non-orthogonal directions while dense features are pruned. How phase changes and geometric packing sharpen interpretability questions without proving what happens in real models.

Transformer Circuits Thread 2022. Nelson Elhage et al.. 34 min read, medium difficulty.