Kiru Lab / Mechanistic Interpretability
Features and Representation
What a model represents, why neurons rarely mean one thing, and what a circuit is.
In this module you will
- Define a feature as a direction rather than a neuron
- Explain superposition and polysemanticity
- Describe a circuit as components composing across layers
Features, Directions, and Superposition
A feature is a direction in activation space. Models pack more features than dimensions, and that is why neurons look confusing.
Concept · 40m
Circuits: Components That Compose
A circuit is a set of components whose composition implements an identifiable algorithm.
Concept · 35m