E8Interpretability & alignment science

Interpretability and alignment science asks what these models are actually computing inside, and how to keep systems steerable as they grow more capable. Where the rest of the pillar trains and serves LLMs, this discipline opens them up: it treats a trained network as an object of study, looking for the structures and circuits that produce its behavior.

You will train probing classifiers on frozen activations, then move to mechanistic interpretability, reproducing the indirect-object-identification circuit and the induction heads that implement in-context learning. You will model superposition and polysemanticity, decompose the residual stream into monosemantic features with sparse autoencoders, and intervene on behavior through activation patching and representation steering. The discipline closes with scalable oversight (debate, amplification) and weak-to-strong generalization, the empirical proxy for aligning a model smarter than its supervisor.

This matters because capability without understanding is fragile: the techniques here feed directly into safety evaluations in pillar I and AI governance in pillar J.