Chris Olah

Overview

Anthropic co-founder. Pioneer of mechanistic interpretability — the research program of reverse-engineering what neural networks compute by examining circuits and features. One of the most cited researchers in AI safety/interpretability.

Affiliations: Anthropic (co-founder, researcher); previously Google Brain / OpenAI.

Why Significant

• Invented or co-developed the core vocabulary of mechanistic interpretability: superposition, features as directions, circuits, polysemanticity
• His 2020 "Circuits" thread (with collaborators) launched interpretability as a rigorous subfield
• As an Anthropic co-founder with deep alignment research focus, represents the lab's internal safety-first culture
• Publicly acknowledged (May 2026) that frontier labs operate within incentives that can conflict with doing the right thing — rare candor from a lab insider

Key Contributions

• Circuits thread (2020–2024): Reverse-engineered attention heads, curve detectors, multimodal neurons in vision models. Foundational interpretability work.
• Superposition Hypothesis: Showed that models represent more features than they have neurons by using compressed, overlapping representations.
• Features as linear directions: The geometry of how concepts are encoded in model weights.
• Anthropic Interpretability team: Built and leads the team doing mechanistic interpretability research at scale on Claude models.

Recent Activity

• 2026-07-28: Named signatory of the "Pacing the Frontier" statement, alongside Dario Amodei and Jared Kaplan, asking the US government to support an international effort to build the tools needed to deliberately pace automated AI development. Anthropic endorsed the statement as an organization the following day. → Frontier Pacing (source)

• 2026-05-25: Invited by Pope Leo XIV to speak at the Vatican presentation of the encyclical "Magnifica humanitas: On safeguarding the human person in the time of artificial intelligence". Spoke alongside cardinals and theologians. Called for cross-institutional partnership between the Church and tech industry for AI governance. Acknowledged industry incentive conflicts publicly. (source)

• 2026-07-06 — Anthropic published A global workspace in language models, the lab's most significant public interpretability result to date: a narrow set of internal word-like representations it calls J-space, described as Claude's shared whiteboard. It is the research programme this page describes, arriving as a named mechanism → Mechanistic Interpretability

• 2026-07-13 — Claude Values Vary by Model and Language, extending the same programme from mechanism to behaviour → Mechanistic Interpretability

Affiliations

• Anthropic — co-founder, interpretability research
• Previously: Google Brain, OpenAI (early)

Related

• Anthropic
• AI Alignment
• Mechanistic Interpretability

External Links

• Distill.pub (interpretability articles)
• Anthropic Interpretability Research