Course in preparation · Coming soon
Interpretability: Probes, Features and Circuits
Look inside a small open model and test what you think you see. Train probes, find features with a sparse autoencoder, trace a circuit, and run the control that could prove each claim wrong.
Who it is for. Researchers and engineers drawn to interpretability teams, who want hands-on work with model internals rather than reading about it.
The planned syllabus
Every module will meet the book's bar before it opens: the math derived on the page, a Colab notebook on Python 3.12, and measurements from real models.
- 01
Activations and the residual stream
Hook a small model, record its activations, and know what each one is.
- 02
Linear probes and their controls
Train a probe, then a control task, so a high accuracy means something.
- 03
Sparse autoencoders
Train one on a small model's activations and judge which features are real.
- 04
Circuits and patching
Activation patching and ablation: find which components carry a behavior.
- 05
Claims that hold up
Write an interpretability result with its evidence, its controls and its limits.
