Mechanistic Interpretability

Chapter 1

Concepts

You will be able to define mechanistic interpretability, explain features, circuits, and universality, and describe the decomposability and linearity assumptions behind the lecture's reverse-engineering story.

Chapter 2

Features

You will be able to treat features as directions in activation space, calculate feature activation, compare vision and language features, and test whether a proposed feature label matches model behavior.

Chapter 3

Circuits

You will be able to define circuits as computational subgraphs, read curve-detector evidence, trace signed information flow, explain induction heads, and judge whether circuit evidence supports a behavioral claim.

Chapter 4

The Superposition Hypothesis

You will be able to explain privileged bases, polysemantic neurons, sparse features, feature interference, toy models of superposition, and why neuron-by-neuron explanations often fail.

Chapter 5

Representation Learning

You will be able to explain sparse representations, step through sparse autoencoder training, connect dictionary learning to feature discovery, and evaluate learned feature libraries.

Chapter 6

Mechanistic Interventions

You will be able to distinguish feature clamping from feature steering, use clamping to diagnostically test feature interpretations, and construct steering vectors that actively influence model behavior.

Chapter 7

Scaling MI to LLMs

You will be able to describe the scale problem in LLM interpretability, connect sparse autoencoder work to Claude and GPT-style feature dictionaries, explain automated circuit and feature workflows, estimate practical costs, and identify open research limits.