When Anthropic was founded, its founders made mechanistic interpretability a central part of the company's direction, focused on LLMs.

Judge as of 2025-02-18 against evidence from Anthropic's founding in 2021. True: records from 2021 or shortly after show that interpretability research on language models was a central, stated research direction at founding. Such records include early hiring of interpretability researchers such as Chris Olah, founding announcements or mission statements naming interpretability, and the early Transformer Circuits publications (e.g. 'A Mathematical Framework for Transformer Circuits', December 2021). Mostly true: interpretability was a prominent early focus but one of several co-equal priorities, or it started within the first year rather than strictly at founding. Mixed: interpretability work existed early but founding-era materials give it only a minor role. Mostly false: interpretability became central only well after founding (e.g. 2023 or later). False: founding-era evidence shows interpretability was not part of the company's direction.

Set before the research began.

1. E1 Anthropic is the new AI research outfit from OpenAI's Dario Amodei, and it has $124M to burn | TechCrunch TechCrunch · 28 May 2021

   “Anthropic, as it’s called, was founded with his sister Daniela and its goal is to create “large-scale AI systems that are steerable, interpretable, and robust.””

2. E2 Anthropic is the new AI research outfit from OpenAI's Dario Amodei, and it has $124M to burn | TechCrunch TechCrunch · 28 May 2021

   “The company is a public benefit corporation, and the plan for now, as the limited information on the site suggests, is to remain heads-down on researching these fundamental questions of how to make large models more tractable and interpretable.”

3. E3 A Mathematical Framework for Transformer Circuits Transformer Circuits Thread (Anthropic) · 22 December 2021 Own record

   “One avenue for addressing these issues is mechanistic interpretability, attempting to reverse engineer the detailed computations performed by transformers, similar to how a programmer might try to reverse engineer complicated binaries into human-readable source code.”

4. E4 Anthropic raises Series B to build robust AI Anthropic · 29 April 2022 Own record

   “On interpretability, it has made progress in mathematically reverse engineering the behavior of small language models and begun to understand the source of pattern-matching behavior in large language models.”

Independent reporting at Anthropic's founding in May 2021 says the company's stated goal was to build "large-scale AI systems that are steerable, interpretable, and robust" (E1). It also says the company planned to stay "heads-down" on making large models "more tractable and interpretable" (E2). In December 2021, Anthropic published a mechanistic interpretability paper on transformers (E3), and in 2022 it reported progress on reverse engineering language models (E4). This shows interpretability was a stated, central direction from founding, but E1 lists it as one of three properties with no ranking among them. Under the criteria, one of several co-equal priorities counts as mostly true.