Mechanistic Interpretability

Mechanistic interpretability is the attempt to reverse-engineer what a trained neural network actually computes, described as features and circuits, rather than treating it as a black box characterised only by its inputs and outputs. A feature is a direction in activation space corresponding to something describable; a circuit is a subgraph of features and weights implementing a behaviour end to end. A published example is the induction head, which detects a repeated prefix and copies whatever followed it last time.

The central obstacle is superposition. A model represents far more distinguishable features than it has dimensions, packing them into near-orthogonal directions and tolerating the interference. Individual neurons are therefore polysemantic — one unit fires for several unrelated things — so reading neurons directly yields no explanation. Sparse autoencoders attack this by training a very wide, sparsely activating dictionary on a layer's activations, recovering features closer to monosemantic. Clamping one and watching the output change is the causal test separating a feature from a correlation.

The limits are real. Recovered dictionaries are incomplete, reconstruction is lossy, feature sets differ between training runs, and most fully traced circuits come from small models. Nondeterminism in training and sampling makes exact replication awkward, and a Token-level story about one prompt rarely generalises.

What none of this licenses is folk prompting advice. Discovering that a Large Language Model carries a feature for a concept says nothing about which phrasing recruits it, and no published result yet justifies an incantation. Claims about Prompt Engineering still need an Evaluation Harness and measured outcomes; interpretability supplies an explanation to be earned, not a Plausible Mechanism to be assumed. Its firmest practical contribution so far is negative — showing that intuitive stories about why a model produced an output, including many about Hallucination, are simply wrong.