Feature enumeration as a safety audit
A list of everything a model could do wrong is a strange kind of safety guarantee, but it's the one this paper closes on: if every feature inside a model were catalogued and understood, dangerous behavior would have nowhere left to hide unnoticed.
This theme names that ambition and the two pieces offered toward it, a scalable way of enumerating features instead of guessing at them, and a method for tracing how those features causally depend on each other, read against a corpus that has mostly audited models by judging their outputs rather than inspecting their insides.
Cunningham et al. (2023) close by naming the ambition their method serves: enumerative safety, Elhage et al.'s (2022) proposal that a complete, human-understandable list of a model's features could guarantee it will not perform dangerous behaviors such as deception. Sparse autoencoders are offered as the tool for building that list -- a scalable, unsupervised way of enumerating features rather than guessing at them -- and feature-circuit-detection as a way of tracing how listed features causally depend on each other across layers, a step toward the paper's stated goal of an end-to-end account of how a model computes its outputs. Read against the rest of the corpus, this reframes the oversight problem: earlier papers audit a model by judging what it outputs -- reward modeling, red-teaming, revised responses -- while enumerative safety proposes auditing it by listing and inspecting its internal parts instead, with mechanistic interpretability named as the research agenda this project frames itself as advancing.