Mechanistic Interpretability

Reverse-engineer the transformer, one circuit at a time

Features, superposition, circuits, and sparse autoencoders — the emerging science of what neural networks actually compute, and why it matters for safety.

Advanced4 modules · 12 lessons~120 min1190 XPAligned with Anthropic/Transformer Circuits researchStart with: Deep Learning

Progress0/12

Module

The Atoms: Features & Superposition

Why the neuron is the wrong unit, where the real variables live, and how networks pack more concepts than dimensions

Features, Not Neurons9 min · 90 XPFeatures, Not Neurons

The Superposition Hypothesis10 min · 100 XPThe Superposition Hypothesis

Feature Geometry & Linearity9 min · 90 XPFeature Geometry & Linearity

Module

Seeing Inside the Transformer

Attention analysis, induction heads, the residual stream, and probes — the first generation of working microscopes

Attention Pattern Analysis10 min · 100 XPAttention Pattern Analysis

Induction Heads & In-Context Learning11 min · 110 XPInduction Heads & In-Context Learning

Logit Lens & the Residual Stream9 min · 90 XPLogit Lens & the Residual Stream

Probing Classifiers9 min · 90 XPProbing Classifiers

Module

Circuits & Causal Tools

From 'what's represented' to 'what's computed' — circuit analysis, patching, and the dictionary microscope

Circuits: Reverse-Engineering Algorithms11 min · 110 XPCircuits: Reverse-Engineering Algorithms

Attribution & Activation Patching10 min · 100 XPAttribution & Activation Patching

Sparse Autoencoders & Dictionary Learning12 min · 110 XPSparse Autoencoders & Dictionary Learning

Module

From Microscope to Stakes

Why interpretability is a safety bet, what oversight actually looks like, and where the frontier of failure is

Interpretability for Oversight10 min · 100 XPInterpretability for Oversight

Limits, Failure Modes & the Frontier10 min · 100 XPLimits, Failure Modes & the Frontier

Capstone project

Open the box

Take a trained model and explain three of its decisions using feature importance, SHAP, or attention inspection. Find one place the model uses a signal you didn't expect.

Deliverable: A notebook with explanations and a write-up of the surprising signal you found.

Milestones

Finish every lesson to earn the Mechanistic Interpretability Certificate.