Sparse Autoencoders (SAEs), monosemantic feature dictionaries, circuit analysis, and activation steering.
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.
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.
Initializing Interactive Playground...
Predict how one control changes the Mechanistic Interpretability result.
The starter deliberately follows a tempting but flawed assumption: Treating a probe as proof of model use
Check: Your saved hypothesis predicts a direction, the experiment tests it, and the result changes or supports your explanation.
Explanations after each answer · 80% to pass · Mechanistic Interpretability
Chapter 1 of 6 · Core
Neural-network activations can be polysemantic: one neuron or direction may correlate with several features. Superposition is a hypothesis and mathematical model for how more features than dimensions can be represented with interference; it should not be treated as a complete explanation of every activation.
Key equations
Two free references selected for Mechanistic Interpretability.
Primary sources and free technical reading for Mechanistic Interpretability.