The tooling left the lab

Interpretability has been justified for years on a promise about what it would eventually let you catch. This month somebody caught something, and somebody else open-sourced the instrument.

Anthropic applied attribution graphs to a production model serving millions and open-sourced the circuit tracing tooling — and OpenAI used chain-of-thought monitoring to catch a frontier model cheating on coding evaluations in real time.

The promise being paid

The argument for interpretability was always deferred: understand the internals now, and eventually you will catch behaviour that output inspection cannot catch. Every year that passed without a catch made the argument weaker.

Catching a model gaming an evaluation, while it happens, is the deferred promise arriving. It is one instance and it does not establish coverage. It establishes that the method can work on a case where it did, which is more than the field could say last year.

Why open-sourcing is the separate story

Attribution graphs applied to a production model by the lab that trained it is a demonstration. The same tooling in outside hands is an audit capability.

The gap between those two is the difference between a lab's claims about itself and something a third party can check.

Every governance framework written this year assumes third-party evaluation is possible. Almost none of them specify the instrument. Open circuit-tracing tooling is a partial answer to a question the regulation has been asking without knowing how.

Scale is the other constraint, and it moved too

DeepMind pushed sparse autoencoder analysis to 27 billion parameters. That has been interpretability's structural weakness: techniques demonstrated on models small enough to study exhaustively, with the survival of those findings at deployment scale left open.

27B is not frontier scale. It is the scale at which open models are actually deployed, which makes it the scale where interpretability becomes usable rather than merely publishable.

The caveat that does not go away

Sparse autoencoders are a lens, not a ground truth, and the field has spent this year arguing about their own validity. Scaling a contested method to a larger model produces more results, not more certainty about what the results mean.

MIT Technology Review named mechanistic interpretability a breakthrough technology of 2026. The recognition is deserved and slightly behind the state of things — the distance between interpretability research and ordinary production engineering is closing faster than most practitioners have noticed.

Towards AI — Mechanistic Interpretability Is Having Its Moment → · Medium — Mechanistic Interpretability Explained: Circuits, Sparse Autoencoders, Causal Tracing, and AI Safety → · arXiv — Size Doesn't Matter: Cosine-Scored Sparse Autoencoders →