Mechanistic interpretability is the branch of AI research that tries to reverse-engineer trained neural networks into human-understandable algorithms - to figure out not just what a model does but how, by identifying the specific internal computations that produce its behavior. The goal is to treat a network’s learned weights as something that can be read and understood, the way one might decompile a program, rather than as an opaque black box.
The agenda was crystallized in “Zoom In: An Introduction to Circuits,” published on Distill in March 2020 by Chris Olah and colleagues, which laid out three claims. First, features: networks are built from meaningful units such as a curve detector or a dog-head detector that can be studied directly. Second, circuits: these features are wired together by weights into computational subgraphs that implement understandable algorithms. Third, universality: similar features and circuits recur across different models and tasks. The claims were offered as speculative but testable hypotheses.
The field has since produced concrete results. Induction heads were identified as an attention circuit driving in-context learning; the theory of superposition explained why single neurons respond to many concepts; and sparse autoencoders were used to pull millions of interpretable features out of production models. A recurring theme is that understanding is hard-won - features are tangled together in superposition, and reading them out takes specialized tools.
Why business readers should care: mechanistic interpretability is the most direct attempt to make AI auditable from the inside. If it succeeds, it could let organizations verify what a model has actually learned and catch dangerous or deceptive behavior that ordinary testing would miss.