Circuits, groups of features, allow researchers to trace an AI model's thinking steps.

Judge as of 2025-02-18. True: published interpretability research by then (e.g., Anthropic's circuit and attribution-graph work, other circuit studies) shows that circuits (connected groups of features) let researchers trace the intermediate computational steps a model uses for specific behaviours. Mostly true: circuit tracing was demonstrated but only for limited behaviours or models, a limit the wording slightly glosses over. Mixed: evidence shows circuit methods trace some model computations but with major disputes about whether this reflects the model's 'thinking'. Mostly false: circuit tracing was only shown in toy models or narrow cases, and stating it as a general ability for AI models overstates it. False: no circuit-based method had shown any ability to trace model computations.

Set before the research began.

1. E1 anthropic-docs-mirror/anthropic-research/tracing-thoughts-language-model.md at main · RobGruhl/anthropic-docs-mirror GitHub (third-party mirror of an Anthropic research post) Own record

   “Even on short, simple prompts, our method only captures a fraction of the total computation performed by Claude, and the mechanisms we do see may have some artifacts based on our tools which don't reflect what is going on in the underlying model.”

2. E2 anthropic-docs-mirror/anthropic-research/tracing-thoughts-language-model.md at main · RobGruhl/anthropic-docs-mirror GitHub (third-party mirror of an Anthropic research post) Own record

   “Our method allows us to artificially change the intermediate steps and see how it affects Claude’s answers. For instance, in the above example we can intervene and swap the "Texas" concepts for "California" concepts;”

3. E3 Circuits Updates - March 2024 Transformer Circuits (Anthropic interpretability team) · March 2024 Own record

   “We investigate an 18-layer model, with a sparse autoencoder with expansion factor of ~256 trained on the layer 9 residual stream (as in Cunningham et al.).”

4. E4 Circuits Updates - March 2024 Transformer Circuits (Anthropic interpretability team) · March 2024 Own record

   “They then ask how, on specific dataset examples, the features in later layers are computed as functions of features in earlier layers by multiplying through attention OV and MLP weights (with a fixed gating mask).”

The March 2024 Anthropic Circuits Update, which predates the claim, applies feature-based circuit analysis to an 18-layer model (E3). It also describes work examining how later-layer features are computed from earlier-layer features on specific dataset examples (E4). So circuit methods had traced intermediate model computations before 2025-02-18, but the dated evidence shows this only for specific models and examples, a limit the general wording glosses over. E1 and E2 describe broader tracing in Claude, but they are undated, so they can't show the work was published by the judge date. E1 also says the method captures only a fraction of the computation.