Chris Olah

Chris Olah is a Canadian machine-learning researcher and co-founder of Anthropic, where he leads the interpretability research program; previously head of the OpenAI Clarity team and co-founder of the Distill journal at Google Brain.
Nextomoro
9 min read
Chris Olah

Chris Olah

Chris Olah is a Canadian machine-learning researcher who co-founded Anthropic in 2021 and leads its interpretability research program. He worked previously at Google Brain, where he co-founded the Distill journal in 2017, and at OpenAI, where he founded and led the Clarity team focused on neural-network interpretability. As of May 2026, Olah is one of the named co-authors on Anthropic's Transformer Circuits Thread and is widely cited in industry coverage as one of the pioneers of mechanistic interpretability.

At a glance

Origins

Olah is Canadian and was raised in Toronto, where he attended The Abelard School and graduated as a National AP Scholar in 2010. He enrolled at university but left at age 18 without completing a degree, an episode he has discussed publicly on the 80,000 Hours podcast in August 2021 in an interview titled "Chris Olah on working at top AI labs without an undergrad degree." The interview frames the period as a personal commitment that pulled him away from formal study rather than a strategic career choice.

In 2012 he was awarded a Thiel Fellowship, the program founded by Peter Thiel that grants $100,000 to people under 22 to pursue work outside university. Olah has cited the fellowship as the funding that allowed him to switch focus from earlier interests including 3D-printing software toward machine learning. The combination of self-directed study, an active personal blog at colah.github.io, and the Thiel funding produced an unusual path into industrial research: by his early twenties Olah had built a public reputation as a clear and visual writer on neural-network internals without the credential profile of a typical research-lab hire.

Career

Olah's first long-running research engagement was at Google Brain. In the 80,000 Hours interviews he has described the entry point as an invitation from Jeff Dean to intern after a talk Olah gave at Google, with the internship extending across roughly two years before he transitioned to full-time research roles. He spent approximately five years at Google Brain in total, with primary press coverage placing him there from around 2014 through 2018. The work at Brain centered on neural-network visualization and interpretability, including a co-authoring contribution to the June 2015 DeepDream blog post on Inceptionism, and the Feature Visualization and Building Blocks of Interpretability papers published on Distill.

In March 2017, while still at Google Brain, Olah co-founded the Distill journal with Shan Carter and Arvind Satyanarayan as editors-in-chief, with backing from Google, OpenAI, DeepMind, and Y Combinator Research. Distill was an open-access journal designed for clear, interactive machine-learning articles published natively on the web rather than as conventional PDFs. The journal went on indefinite hiatus in 2021.

Olah moved to OpenAI around 2018 to found and lead the Clarity team focused on neural-network interpretability. The Clarity period produced the Circuits thread published on Distill, including the lead article "Zoom In: An Introduction to Circuits" in March 2020, and the May 2021 reverse-engineering work on the CLIP image classifier. The thread established a research methodology built around individual neurons, attention heads, and small computational subgraphs as the unit of interpretability analysis.

In December 2020 Olah left OpenAI alongside Dario Amodei, Daniela Amodei, and several other senior OpenAI staff. The group formally incorporated Anthropic as a Delaware Public Benefit Corporation in early 2021 with seven public co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Sam McCandlish, Jared Kaplan, Jack Clark, and Olah. Olah took responsibility for the company's interpretability research program, which has since produced the Transformer Circuits Thread papers as the methodological successor to the Distill Circuits thread.

In 2024 Olah was named to the TIME 100 AI list, which described him as one of the pioneers of mechanistic interpretability. He also serves as an advisor to Goodfire, the San Francisco mechanistic-interpretability startup founded in 2024 by Eric Ho, Daniel Balsam, and Tom McGrath, which raised a $150 million Series B in February 2026 at a $1.25 billion valuation.

Affiliations

  • Google Brain: Research Intern, Research Associate, and Research Scientist, approximately 2014 to 2018.
  • OpenAI: Researcher and lead of the Clarity team, approximately 2018 to 2021.
  • Anthropic: Co-founder and interpretability research lead, 2021 to present.
  • Goodfire: Advisor.

Notable contributions

Olah's body of work is concentrated on neural-network interpretability research and on the publication formats and venues that communicate it. Many of the most-cited papers below are co-authored, and Olah is credited as one of several authors rather than the sole author.

  • DeepDream and Inceptionism (June 2015). Co-authored Google Brain blog post on visualizing neural-network features by gradient ascent on input images.
  • Distill journal (March 2017). Co-founder and editor-in-chief alongside Shan Carter and Arvind Satyanarayan. The journal was a venue for clear, interactive machine-learning articles published natively on the web; it went on indefinite hiatus in 2021.
  • "Feature Visualization" (November 2017). Distill article co-authored with Alexander Mordvintsev and Ludwig Schubert on visualizing what neurons and channels in vision models respond to.
  • "The Building Blocks of Interpretability" (March 2018). Distill article co-authored with Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, and Alexander Mordvintsev on visualization tools for vision models.
  • "Zoom In: An Introduction to Circuits" (March 2020). Lead article in the Circuits thread on Distill, co-authored with Nick Cammarata, Shan Carter, Gabriel Goh, Michael Petrov, and Ludwig Schubert. The canonical methodological statement of mechanistic interpretability as a research program.
  • "Multimodal Neurons in Artificial Neural Networks" (March 2021). OpenAI Distill article reverse-engineering CLIP, co-authored with Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, and Alec Radford.
  • "A Mathematical Framework for Transformer Circuits" (December 2021). Foundational article in the Transformer Circuits Thread, co-authored with Nelson Elhage, Neel Nanda, Catherine Olsson, and others, on the mathematical decomposition of small attention-only transformers.
  • "In-context Learning and Induction Heads" (March 2022). Transformer Circuits article identifying induction heads as a circuit pattern underlying in-context learning.
  • "Toy Models of Superposition" (September 2022). Transformer Circuits article on the phenomenon by which neural networks represent more features than they have neurons.
  • "Towards Monosemanticity" (October 2023). Transformer Circuits article applying sparse dictionary learning to extract interpretable features from a small transformer; lead-authored by Trenton Bricken with Olah among the senior authors.
  • "Scaling Monosemanticity" (May 2024). Transformer Circuits article scaling the dictionary-learning approach to a production frontier model; lead-authored by Adly Templeton with Olah among the senior authors.
  • "Circuit Tracing" (March 2025). Transformer Circuits article presenting attribution-graph methodology for tracing internal computations in production language models; lead-authored by Emmanuel Ameisen with Olah among the senior authors.
  • TIME 100 AI (2024). Olah was named to the TIME 100 AI 2024 list, with the citation describing him as one of the pioneers of mechanistic interpretability.
  • Public-talk record. "Looking Inside Neural Networks with Mechanistic Interpretability" at the FAR.AI San Francisco Alignment Workshop in February 2023; the November 2024 Lex Fridman Podcast #452 appearance alongside Dario Amodei and Amanda Askell; the August 2021 80,000 Hours interview on interpretability research.

Investments and boards

  • Goodfire (AI): Advisor. Mechanistic-interpretability company founded in 2024 by Eric Ho, Daniel Balsam, and Tom McGrath. As of February 2026, Goodfire has raised approximately $209 million across three rounds, including a $150 million Series B at a $1.25 billion valuation, and counts Anthropic as a strategic investor through the April 2025 Series A.

No public personal angel-investor activity on record outside the Goodfire advisory role in AI, semiconductors, datacenters, software, or energy as of May 2026. Olah's footprint in this section is concentrated in his research and operating role at Anthropic rather than a parallel investing program.

Network

Olah's longest-running professional relationships are with his six fellow Anthropic co-founders, all of whom worked alongside him at OpenAI in the period before the 2021 founding: Dario Amodei, the company's chief executive; Daniela Amodei, the company's president; Tom Brown, the lead author of the GPT-3 paper at OpenAI; Sam McCandlish; Jared Kaplan, the chief science officer; and Jack Clark, the policy lead. The Anthropic interpretability team also includes a long-running research network of repeat co-authors: Catherine Olsson, Nelson Elhage, Trenton Bricken, Adly Templeton, Nick Cammarata, and Neel Nanda among others, with Nanda having gone on to become a prominent independent voice on mechanistic interpretability after a period at Anthropic.

His Distill collaborator network from the Google Brain and OpenAI period includes Shan Carter, his co-founding editor at Distill and a frequent paper co-author; Arvind Satyanarayan of MIT, the third Distill editor-in-chief; Alexander Mordvintsev, the DeepDream and Feature Visualization co-author; Ludwig Schubert and Michael Petrov, recurring Circuits-thread co-authors; and Gabriel Goh, the OpenAI multimodal-neurons co-author. The broader interpretability research community in which Olah is a frequent reference point includes the Goodfire founding team led by Eric Ho, Daniel Balsam, and Tom McGrath, and the DeepMind interpretability research group from which McGrath joined Goodfire.

Position in the field

As of May 2026, Olah is the named author most frequently associated in industry coverage with the mechanistic interpretability research program, a body of work that treats the internal mechanisms of trained neural networks as the unit of analysis, in contrast to behavioral evaluation or training-data analysis. Industry coverage including the TIME 100 AI 2024 entry credits him as one of the pioneers of the field, and Anthropic's chief executive Dario Amodei has referred to Olah on the November 2024 Lex Fridman Podcast as a co-founder of the discipline.

The credential profile is unusual among senior industrial-research staff. Olah does not hold an undergraduate degree, and his transition into research went through the Thiel Fellowship and a self-published technical-blog body of work rather than a doctoral program. The 80,000 Hours podcast in August 2021 ran a dedicated interview on the path, and Olah is a frequent reference point in coverage of AI research careers without traditional academic credentials.

The publication-format choices have been similarly distinctive. The Distill journal, co-founded in 2017, was a deliberate move away from PDF-format conference papers toward interactive web articles. After the journal went on hiatus in 2021, the Transformer Circuits Thread at Anthropic continued the same publication style for in-house interpretability research. Anthropic's commitment to interpretability as a structural research investment, rather than a side program, is one of the company's distinguishing features in the frontier-lab landscape, and Olah's role at the head of that program is the institutional anchor for the commitment.

Outlook

Open questions over the next 6 to 18 months:

  • Next Transformer Circuits Thread releases. Successor papers to the Circuit Tracing methodology and the Scaling Monosemanticity dictionary-learning approach, including any results on production frontier models in the Claude 4.x and successor lines.
  • The Anthropic interpretability roadmap. Whether and how the published mechanistic-interpretability research program is wired into capability-release decisions under Anthropic's Responsible Scaling Policy, as foreshadowed in Dario Amodei's April 2025 essay on the urgency of interpretability research.
  • Goodfire trajectory. The trajectory of the Goodfire mechanistic-interpretability platform on which Olah serves as advisor, including any further commercial validation of interpretability tooling beyond the February 2026 Series B at a $1.25 billion valuation.
  • Public-talk cadence. Frequency and substance of Olah's conference-talk and podcast appearances, given the comparatively rare long-form interview record before the 2021 80,000 Hours episode.
  • Independent mechanistic-interpretability community. The growth of the broader mechanistic-interpretability research community, including the work of former Anthropic-adjacent researchers such as Neel Nanda, and any contribution that emerges from external academic groups.
  • Anthropic governance and capability releases. Whether the interpretability research program continues as a senior research priority through any future Anthropic governance changes including a potential public listing.

Sources

Person Anthropic
About the author
Nextomoro

Nextomoro

Read next

Jeff Clune

Vik Bajaj

John Schulman

Karan Malhotra

Arthur Mensch

AI Research Lab Intelligence

nextomoro tracks progress for AI research labs, models, and what's next.

AI Research Lab Intelligence

Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to AI Research Lab Intelligence.

Your link has expired.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.