Anthropic released a writeup on Garcon, the infrastructure tool its team used for interpretability research.

Judged as of 2022. True: Anthropic published, on or before the date the statement was made, a writeup describing Garcon as the infrastructure its team used for interpretability research (for example, on transformer-circuits.pub). Mostly true: the writeup exists, but a minor detail differs, for example it was published as part of another article rather than on its own, or it was released by members of the team rather than by Anthropic as such. Mostly false: Garcon is only mentioned in passing in other work, without a writeup, or the writeup appeared only after the statement. False: Anthropic has no published material on a tool called Garcon.

Set before the research began.

1. E1 Garcon Transformer Circuits Thread (transformer-circuits.pub); the page does not name a publisher separately · 22 December 2021 Own record

   “This document describes one of the core pieces of infrastructure we've built to enable interpretability research at Anthropic: a tool we call Garçon.”

2. E2 Garcon Transformer Circuits Thread (transformer-circuits.pub); the page does not name a publisher separately · 22 December 2021 Own record

   “We've found Garçon very helpful for our research. We now use it for essentially all of our interpretability work, and other teams have also adopted it for ad-hoc exploration of models.”

3. E3 A Mathematical Framework for Transformer Circuits Transformer Circuits Thread (transformer-circuits.pub); the page does not name a publisher separately · 22 December 2021 Own record

   “The crucial piece of software infrastructure which made our analysis of models possible is a library called Garcon (described in another article) created by Nelson Elhage with help from Tom Brown, Sam McCandlish, and Chris Olah.”

E1 is a standalone Garçon writeup on transformer-circuits.pub, dated 2021-12-22, which is before the 2022 statement. It describes Garçon as "one of the core pieces of infrastructure we've built to enable interpretability research at Anthropic." E2, from the same page, says the team uses it for essentially all of its interpretability work. E3 is a companion paper that names Garcon as the key infrastructure and points to the separate article. The evidence is the published writeup itself, which matches the criteria for true.