Back to Curriculum
Advanced•Modern AI
Mechanistic Interpretability
Sparse Autoencoders (SAEs), monosemantic feature dictionaries, circuit analysis, and activation steering.
Interactive Playground
Initializing Interactive Playground...
Research-Level Deep Dive & Equations
Neural network activations in Large Language Models exhibit **superposition**: a single neuron fires for multiple completely unrelated semantic concepts (polysemanticity).
•Why Superposition Occurs: LLMs need to represent millions of real-world concepts, but residual stream activations are constrained to lower dimensions (e.g. ). By storing non-orthogonal concept vectors in superposition, the model squeezes concepts into dimensions, relying on non-linear activation thresholds (ReLU, SwiGLU) to suppress interference noise.
•The Mechanistic Interpretability Goal: To reverse-engineer black-box neural networks into human-understandable circuits, we must decompose polysemantic residual activations into monosemantic concept features.
Key Equations
Test Your Knowledge
Check whether you have mastered this concept with a quick quiz.
Was this lesson helpful?
Your feedback helps us continuously improve the curriculum and interactive visualizations.