Chris Olah

Overview

Before Chris Olah co-founded Anthropic — and before the billions — he was a researcher with no university degree who taught the world neural networks from his personal blog. In 2012 he took a $100,000 Thiel Fellowship, a grant designed to push gifted young people straight into research or entrepreneurship instead of university, and managed everything that followed without a degree: his blog posts and threads attracted millions of readers trying to understand cutting-edge machine learning (80,000 Hours). Four years at Google Brain, where he was second author on the 2015 launch of DeepDream, led to an interpretability team of his own at OpenAI — and in 2021, to co-founding Anthropic, where he leads the effort to reverse-engineer what the most powerful AI models are actually doing inside (80,000 Hours).

Olah’s field now has a name — mechanistic interpretability — and he is credited as one of its founders. The question driving him is disarmingly simple. “Imagine if some alien organism landed on Earth and could go and do these things,” he said in a 2021 interview. “Everybody would be rushing and falling over themselves to figure out how the alien organism was doing things… it seems to me that really the thing that is calling out in all this work for us to go and answer is, what in the wide world is going on inside these systems?” (80,000 Hours)

The dropout who taught the field

Olah’s path into research ran through explanation, not credentials. He built his early reputation on colah.github.io, a personal blog of unusually clear, visual technical essays — his 2015 post “Visualizing Representations: Deep Learning and Human Beings” was still being cited in Distill’s references years later (Distill). By the time he joined Google Brain in 2015, he had already won a readership most professors never get (80,000 Hours).

At Google Brain he spent four years building tools to visualize what was happening inside neural networks. He was second author on the 2015 “Inceptionism: Going deeper into neural networks” post — the launch of DeepDream, the image-hallucination technique that let the public see what a neural network was “seeing,” which, as one interviewer put it, “almost everyone has seen” (80,000 Hours; Distill). He also co-authored the well-known paper “Concrete Problems in AI Safety” (80,000 Hours).

He then helped found Distill, an academic journal built around a simple idea: machine learning deserved to be published as interactive web articles with visualizations baked in, not as static PDFs (80,000 Hours). The journal became the home of his signature research line — “Feature Visualization” (2017), “The Building Blocks of Interpretability” (2018), the activation atlases (2019), and “Zoom In: An Introduction to Circuits” (2020), the methodological statement of the circuits program (Distill). A paper published in Distill is what led Olah to his next job: in 2018 he left Google Brain to found and lead OpenAI’s interpretability team (80,000 Hours).

Reading the mind of a model

At OpenAI, Olah’s team pushed interpretability past visualization and into something stranger: finding individual neurons inside models that seemed to mean things. Their most famous result came from CLIP, OpenAI’s model for pairing images with captions. Digging through it, they found multimodal neurons — units that responded to the same concept whether it arrived as a picture, a drawing, or written text. There was, for example, a “yellow” neuron that fired for the color yellow, for the written word “yellow,” and for words like “lemon” and “banana”; and a Spider-Man neuron that fired for the word Spider-Man rendered as an image, for photographs of Spider-Man, and for drawings of him. Olah noted that this mirrored a real neuroscience phenomenon — human brains contain neurons that respond to a person whether seen or named (80,000 Hours).

The popular version credits Olah alone with the discovery, but the paper’s own author notes tell a more precise story: Gabriel Goh was the research lead who first discovered multimodal neurons, and it was Goh together with Michael Petrov who discovered the original Spider-Man neuron in the dataset examples. Olah’s role was the overall framing of the article, mentoring each team member through their work, and writing the region-neuron sections (Distill).

His motive for all of this is empirical, not philosophical. He has said he holds no very confident theory of AI safety — theorizing about it a priori, he argued, “can really easily lead you astray” — and prefers to study the systems directly, because models that do the right thing for the wrong reasons are exactly the failure mode you cannot catch by testing alone (80,000 Hours).

Anthropic and the move from OpenAI

In December 2020, Olah left OpenAI along with several colleagues. The group had decided to start a new lab: Anthropic, an AI research company focused, in Olah’s words, on making large models “reliable, interpretable, and steerable” — training frontier-scale models and studying interpretability and human feedback in that context, on the argument that safety research disconnected from the most capable systems will always lag behind them (80,000 Hours). Anthropic was set up as a for-profit public benefit corporation — meaning, Olah explained, that it is not legally obligated to maximize shareholder value — and its Series A was led by Jaan Tallinn, a funder focused on AI safety, joined by individual backers rather than venture firms (80,000 Hours).

The company was created in 2021 by a team of former OpenAI employees led by siblings Dario and Daniela Amodei. Its seven co-founders, besides the Amodeis, are Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah — and the company, known for its chatbot Claude, reached roughly a $1 billion annual revenue run rate by late 2024 (The Business Times). Olah has anchored Anthropic’s interpretability program ever since, continuing the Transformer Circuits research thread the team began at OpenAI.

That program’s flagship result arrived in 2024 with “Scaling Monosemanticity,” which applied sparse autoencoders — a dictionary-learning technique — to Claude 3 Sonnet, Anthropic’s production model. The team trained autoencoders with roughly 1 million, 4 million, and 34 million features and found highly abstract, often multilingual and multimodal concepts: famous people, countries and cities, code with security vulnerabilities. Some features were safety-relevant — deception and power-seeking, sycophancy, bias, dangerous content — and forcing a feature to high values (“steering”) measurably changed the model’s demeanor, preferences, stated goals, and in some cases its ability to circumvent safeguards (Anthropic). It was the first demonstration that the black box of a frontier language model could be opened at scale, not just studied on toy systems.

The market noticed. On March 3, 2025, Anthropic closed a $3.5 billion round led by Lightspeed Venture Partners at a $61.5 billion valuation — a deal that, per the Bloomberg Billionaires Index, gave each of the seven co-founders a fortune of more than $1.2 billion (The Business Times).

What this profile does not claim

This profile does not assert Olah’s birth year or exact birthplace — no source read for this profile confirmed them. It lists his country as Canada on the basis of multiple tertiary biographies, but no tier-1 or tier-2 source read here stated his citizenship directly, so that field carries less confidence than the rest of the profile. It does not claim a TIME 100 AI listing: the TIME page could not be fetched and the listing was not independently verified. It gives no founding year for Distill, because sources read and consulted disagree on it. It does not repeat the widely-quoted claim about why he left university as a teenager, which this profile could not verify from a readable source. And it credits the multimodal-neuron discovery to Gabriel Goh and Michael Petrov, per the article’s own author notes, rather than to Olah alone.