Learn mechanistic interpretability

what is mechanistic interpretability?

A neural network is a few million (or billion) numbers - its weights. When you push data through them, those numbers produce useful behaviour: a classification, a sentence, a prediction. Nobody designed them by hand; gradient descent did. So the model works, but on first inspection the weights look like noise.

Mechanistic interpretabilityis the project of looking inside trained models and reverse-engineering what those weights are actually computing. Not vague stories like “this neuron likes cats”, but algorithm-level descriptions clean enough that a human could implement them by hand. The bet is that if we can read what a model is doing, we can catch it doing something we don't want before it ships.

This curriculum walks you from the simplest possible case (literally reading digit templates off a weight matrix) to the frontier of the field today (sparse autoencoders pulling monosemantic features out of real language models). Six steps. No prerequisites beyond Python.