CONCEPT
Inference from Traces
The method—perfected by J.J. Thomson—of reconstructing a hidden interior from controlled perturbations of a system’s visible outputs: the ancestral logic of both experimental physics and modern AI interpretability.
Thomson never saw an electron; he saw a glowing patch shift on a screen under known fields and reasoned his way to a particle a thousand times lighter than the lightest atom. The method was the same across his whole career: the cathode-ray deflection that found the electron, the constrained sphere of positive charge that generated the plum pudding, the two parabolas on a photographic plate that revealed isotopes in neon gas. In each case a hidden structure—invisible to any direct instrument—was made legible by its effects on things that could be measured. This is not merely a historical technique; it is the operative logic of
AI interpretability research today, which works by sending controlled probes through an opaque model and reading its internal organisation off the pattern of what bends. The limit Thomson himself identified is part of the concept: inference from traces works because matter is lawful, but the latent structure of a
language model—far less stable than an atom under perturbation—requires correspondingly more