Anthropic published all of its interpretability work so others could benefit.

Judge as of 2025-03-10. True: Anthropic's interpretability research up to that date was publicly released (e.g. the Transformer Circuits publications, papers, blog posts, and released code or tools where relevant), and there is no evidence of significant interpretability work being kept unpublished. Mostly true: the large majority of the work was published, with only minor or incidental internal work left unreleased. Mixed: substantial work was published, but a significant portion is documented as withheld. Mostly false: only selected results were published, while Anthropic or credible sources acknowledge that significant interpretability work was kept internal; this is a part stated as the whole. False: little of the work was published.

Set before the research began.

1. E1 Circuits Updates - March 2024 Anthropic interpretability team (transformer-circuits.pub) · March 2024 Own record

   “While these were improvements to the training process we used at the time, other changes have superseded them. We intend to publish these changes in a future update once we’ve vetted them further.”

2. E2 Circuits Updates - March 2024 Anthropic interpretability team (transformer-circuits.pub) · March 2024 Own record

   “Others are minor points we wish to share, since we're unlikely to ever write a paper about them.”

3. E3 A mathematical framework for Transformer Circuits Anthropic · 22 December 2021 Own record

   “Dec 22, 2021”

4. E4 Open-sourcing circuit-tracing tools Anthropic · 29 May 2025 Own record

   “Today, we’re open-sourcing the method so that anyone can build on our research.”

5. E5 Circuits Updates - March 2024 Anthropic (transformer-circuits.pub, Anthropic interpretability team) · March 2024 Own record

   “Some of these are emerging strands of research where we expect to publish more on in the coming months. Others are minor points we wish to share, since we're unlikely to ever write a paper about them.”

6. E6 Circuits Updates - March 2024 Anthropic (transformer-circuits.pub, Anthropic interpretability team) · March 2024 Own record

   “While these were improvements to the training process we used at the time, other changes have superseded them. We intend to publish these changes in a future update once we've vetted them further.”

7. E7 Toy models of superposition Anthropic · 14 September 2022 Own record

   “In this paper, we use toy models — small ReLU networks trained on synthetic data with sparse input features — to investigate how and when models represent more features than they have dimensions.”

Anthropic's own pages show that several interpretability works were published before the judging date, including the Transformer Circuits framework (E3) and Toy Models of Superposition (E7), and that minor findings were shared in a March 2024 update (E2, E5). The same update says some training-process changes were held back pending further vetting (E1, E6), and Anthropic open-sourced its circuit-tracing method only in May 2025, after the judging date (E4). No independent source describes the total scope of Anthropic's interpretability work, so the evidence cannot establish whether 'all' or most of it was published.