Mechanistic Interpretability

Sparse Autoencoders (SAEs), monosemantic feature dictionaries, circuit analysis, and activation steering.

Not started3 min explanation

Visualize, practice, and deep-dive material are optional—use only what helps you learn.

Explanation

A focused 3-minute explanation using the topic's authored material.

Learning goals and prerequisites

After this lesson

Run causal activation experiments
Train and evaluate sparse feature dictionaries
Use interpretability evidence responsibly

Helpful before starting

Transformer internals and linear algebra
Causal intervention and experimental design
Representation learning and sparse models

Start here

Mechanistic Interpretability, in plain language

Sparse Autoencoders (SAEs), monosemantic feature dictionaries, circuit analysis, and activation steering. Mechanistic interpretability studies internal computations and can generate causal hypotheses about model behavior, though current methods are incomplete and easy to overstate.

For a small example, a tiny model has a neuron correlated with bracket depth. Measure activation, intervene by patching or ablating it, and check whether the predicted behavior changes causally. This is the mechanism to keep in view as the lesson becomes more technical. Before moving on, identify the input, transformation, output, and one observation that would falsify your conclusion.

Key points

Activations, residual stream, heads, MLPs, features, circuits, and probes.
Correlation vs causation and intervention-based evidence.
Activation patching, ablation, attribution, and steering.

Use it well

When Mechanistic Interpretability helps—and where it breaks

A model switches from recall to copying under a prompt pattern. Form a circuit hypothesis, localize components, run counterfactual activations, and reproduce across examples and seeds. A useful result still depends on checking the assumptions and evidence below rather than treating one successful output as proof.

Key points

Treating a probe as proof of model use. Better approach: Intervene on the representation and measure behavior.
Naming features from a few examples. Better approach: Use broad samples, negatives, and causal validation.
Using interpretability as a safety certificate. Better approach: Treat it as one evidence source in a layered assurance case.

Was this lesson helpful?

Submit to the team when server feedback is available; otherwise this browser keeps a local copy and tells you so.

Back to Topic Library