2023
Do we understand what we built?
The cuts
Treat the network like an organism: map its circuits, name its features, do biology on the weights.
Interpretability as a field — the admission, made rigorous, that we grew something we cannot yet read.
Zoom In: An Introduction to Circuits2020The tensions
Interpretability inverted the classic question. For every previous machine, we understood the mechanism and asked about the behavior. Here the behavior works and we ask about the mechanism. We built it, we trained it, we can watch every weight — and we still cannot say, in the way an engineer means it, why it does what it does.
