Chris Olah is a machine learning researcher regarded as one of the founders of the field of mechanistic Mechanistic interpretability, and a co-founder of Anthropic, where he leads interpretability research.
Early work and Distill
Olah did not complete a traditional university degree, instead building a research career through independent and industry work, including time at Google Brain, where he contributed to early visualization and feature-interpretation techniques for convolutional neural networks, and at OpenAI. In 2017, he co-founded Distill, an online journal focused on clear, interactive visual explanations of machine learning research, which became influential for its editorial emphasis on genuine understanding over publication volume before ceasing regular publication in 2021.
Interpretability research
Olah's research centers on mechanistic interpretability: the attempt to reverse-engineer the internal computations of trained neural networks into human-understandable algorithms, rather than treating them as opaque black boxes. His early circuits-based work, conducted partly at OpenAI and continued at Anthropic, aimed to identify specific "features" and "circuits" inside vision and language models corresponding to interpretable concepts. At Anthropic, his team has published widely cited work on extracting monosemantic features from large language models using sparse autoencoders, part of a broader research agenda aimed at eventually being able to audit a model's internal reasoning rather than relying only on its outputs.
Anthropic
Olah co-founded Anthropic in 2021 alongside Dario Amodei, Daniela Amodei, Tom Brown, and Jared Kaplan, several of whom he had worked with at OpenAI. Interpretability has been positioned within Anthropic as a core pillar of its safety strategy, alongside Constitutional AI and its Responsible Scaling Policy (see Responsible scaling policies), on the premise that understanding a model's internals could eventually catch dangerous behavior, such as deceptive or reward-hacking tendencies, that black-box evaluation would miss.
Influence
Olah is widely credited with turning interpretability from a niche academic pursuit into a well-funded, central research program at a frontier lab, influencing similar interpretability efforts at Google DeepMind and elsewhere, and shaping how the AI safety field thinks about the difference between behavioral alignment and genuine understanding of a model's internal reasoning. His self-taught, non-traditional path into the field is also frequently cited as a counterexample to the assumption that frontier AI research requires a conventional academic pedigree.