Research Scientist, Interpretability

About this role

<div class="content-intro"><h2><strong>About Anthropic</strong></h2> <p>Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.</p></div><h2 class="heading">About the role:</h2> <p>When you see what modern language models are capable of, do you wonder, "How do these things work? How can we trust them?"</p> <p>The Interpretability team at Anthropic is working to reverse-engineer how trained models work because we believe that a mechanistic understanding is the most robust way to make advanced systems safe. We’re looking for researchers and engineers to join our efforts.&nbsp;</p> <p>People mean many different things by "interpretability". We're focused on mechanistic interpretability, which aims to discover how neural network parameters map to meaningful algorithms. Some useful analogies might be to think of us as trying to do "biology" or "neuroscience" of neural networks using “microscopes” we build, or as treating neural networks as binary computer programs we're trying to "reverse engineer".</p> <p>A few places to learn more about our work and team at a high level are <a class="text-accent-secondary-100 underline" href="https://www.youtube.com/watch?v=TxhhMTOTMDg" target="_blank">this introduction to Interpretability</a> from our research lead, <a class="text-accent-secondary-100 underline" href="https://colah.github.io/about.html" target="_blank">Chris Olah</a>; a <a class="text-accent-secondary-100 underline" href="https://open.spotify.com/episode/5UF79Uu94ia0fwC32a89LU" target="_blank">discussion of our work</a> on the <a class="text-accent-secondary-100 underline" href="https://www.nytimes.com/column/hard-fork" target="_blank">Hard Fork podcast</a> produced by the New York Times, and this <a class="text-accent-secondary-100 underline" href="https://www.anthropic.com/research/engineering-challenges-interpretability" target="_blank">blog post</a> (and accompanying video) sharing more about some of the engineering challenges we’d had to solve to get these results. Some of our team's notable publications include <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2021/framework/index.html" target="_blank">A Mathematical Framework for Transformer Circuits</a>, <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html" target="_blank">In-context Learning and Induction Heads</a>, <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2022/toy_model/index.html" target="_blank">Toy Models of Superposition</a>, <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2024/scaling-monosemanticity/" target="_blank">Scaling Monosemanticity</a>, and our Circuits’ <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2025/attribution-graphs/methods.html" target="_blank">Methods</a> and <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2025/attribution-graphs/biology.html" target="_blank">Biology</a> papers. This work builds on ideas from members' work prior to Anthropic such as the <a class="text-accent-secondary-100 underline" href="https://distill.pub/2020/circuits/" target="_blank">original circuits thread</a>, <a class="text-accent-secondary-100 underline" href="https://distill.pub/2021/multimodal-neurons/" target="_blank">Multimodal Neurons</a>, <a class="text-accent-secondary-100 underline" href="https://distill.pub/2019/activation-atlas/" target="_blank">Activation Atlases</a>, and <a class="text-accent-secondary-100 underline" href="https://distill.pub/2018/building-blocks/" target="_blank">Building Blocks</a>.</p> <p>We aim to create a solid foundation for mechanistically understanding neural networks and making them safe (see our <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2023/interpretability-dreams/index.html" target="_blank">vision post</a>). In the short term, we have focused on resolving the issue of "superposition" (see <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2022/toy_model/index.html" target="_blank">Toy Models of Superposition</a>, <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2023/toy-double-descent/index.html" target="_blank">Superposition, Memorization, and Double Descent</a>, and our <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2023/may-update/index.html" target="_blank">May 2023 update</a>), which causes the computational units of the models, like neurons and attention heads, to be individually uninterpretable, and on finding ways to decompose models into more interpretable components. Our subsequent <a class="text-accent-secondary-100 underline" href="https://www.anthropic.com/news/mapping-mind-language-model" target="_blank">work</a> found millions of features in Sonnet, one of our production language models, represents progress in this direction. In our most recent work, we develop methods that allow us to build circuits using features and use this circuits to understand the mechanisms associated with a model's computation and study specific examples of multi-hop reasoning, planning, and chain-of-thought faithfulness on Haiku 3.5, one of our production models.” This is a stepping stone towards our overall goal of mechanistically understanding neural networks.</p> <p>We often collaborate with teams across Anthropic, such as Alignment Science and Societal Impacts to use our work to make Anthropic’s models safer. We also have an <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2024/april-update/index.html#interpretability-architecture" target="_blank">Interpretability Architectures project</a> that involves collaborating with Pretraining.</p> <h2 class="heading">Responsibilities:</h2> <ul> <li> <p>Develop methods for understanding LLMs by reverse engineering algorithms learned in their weights</p> </li> <li> <p>Design and run robust experiments, both quickly in toy scenarios and at scale in large models</p> </li> <li> <p>Create and analyze new interpretability features and circuits to better understand how models work.</p> </li> <li> <p>Build infrastructure for running experiments and visualizing results</p> </li> <li> <p>Work with colleagues to communicate results internally and publicly</p> </li> </ul> <h2 class="heading"><strong>You may be a good fit if you:</strong></h2> <ul> <li> <p>Have a strong track record of scientific research (in any field), and have done <em>some</em> work on Interpretability</p> </li> <li> <p>Enjoy team science – working collaboratively to make big discoveries</p> </li> <li> <p>Are comfortable with messy experimental science. We're inventing the field as we work, and the first textbook is years away</p> </li> <li> <p>You view research and engineering as two sides of the same coin. Every team member writes code, designs and runs experiments, and interprets results</p> </li> <li> <p>You can clearly articulate and discuss the motivations behind your work, and teach us about what you've learned. You like writing up and communicating your results, even when they're null</p> </li> </ul> <p>To learn more about the skills we look for and how to prepare for this role, see our blog post – <a class="text-accent-secondary-100 underline" href="https://transformer-circuits.pub/2025/april-update/index.html#work" target="_blank">So You Want to Work in Mechanistic Interpretability?</a></p> <p>Familiarity with Python is required for this role.</p> <h2 class="heading"><strong>Role Specific Location Policy:</strong></h