Become an AI engineer

Interpretability

Interpretability is a unit of 18 lessons in Rabbithole's Become an AI engineer course, at level 5.

Lessons

1. Activation steering: clamping features to test what they doLevel 5
2. Attention Head CompositionLevel 5
3. Attribution graphs and circuit tracingLevel 5
4. Circuits: the QK and OV paths through a transformerLevel 5
5. Features and feature directionsLevel 5
6. Global Workspace Theory: finding the workspace layers in a language modelLevel 5
7. Introspective Awareness: Testing a Model's Self-Reports with Concept InjectionLevel 5
8. The Jacobian lensLevel 5
9. The Logit LensLevel 5
10. Reading a model's internals: yardsticks, Jacobians and the workspaceLevel 5
11. Model Diffing: Finding What Training ChangedLevel 5
12. Multilingual CircuitsLevel 5
13. Probing: reading concepts out of activationsLevel 5
14. SoLU: a design change for interpretable MLP neuronsLevel 5
15. Sparse autoencoders: training, dead neurons and dictionary tuningLevel 5
16. Mixture of Linear Transforms (MOLT)Level 5
17. Successor Heads: Finding and Testing ThemLevel 5
18. Superposition: packing many features into few dimensionsLevel 5

Sources

These lessons draw on 15 sources. The most used:

• Transformer Circuits Thread
• Verbalizable Representations Form a Global Workspace in Language Models
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations
• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
• On the Biology of a Large Language Model
• Circuit Tracing: Revealing Computational Graphs in Language Models
• A Mathematical Framework for Transformer Circuits
• System Card: Claude Sonnet 4.5

All units in Become an AI engineer