Mechanistic interpretability
An effort to understand how a neural network works by taking its internal machinery apart and identifying the actual pieces — the patterns it represents and the wiring that connects them — rather than guessing at its logic from the outside. The claim is that a model can be opened from within and read like a circuit, down to isolating a single internal feature and turning it up to steer the model's behaviour. This sets it against approaches that leave the model sealed and only fit an approximation to what it does.
The programme of reverse-engineering a neural network's internal computation into human-understandable parts — features (directions in activation space) and the circuits of weights connecting them — rather than approximating its behaviour from outside. Stated in Olah and colleagues' "Zoom In" (Distill, 2020) and scaled to a production language model in Anthropic's Scaling Monosemanticity (2024).
Mechanistic interpretability pursues the claim that the box can be opened from the inside. Where post-hoc explainability leaves the model intact and fits an approximation to it, the circuits work traces individual features inside a vision network and the weighted connections that build simple detectors into complex ones, advancing the claim that these are the network's real, studyable units.
Scaling Monosemanticity carries the approach to Claude 3 Sonnet using sparse autoencoders to extract millions of features, and demonstrates control as well as reading: clamping the "Golden Gate Bridge" feature high makes the model steer every exchange toward, and identify as, the bridge. The standing challenge to the programme is Jonas and Kording's finding that analysis methods can reveal structure without recovering function — the circuits programme is in part a claim to possess the tools that their neuroscientists lacked.