Research Engineer, Interpretability

When you see what modern language models are capable of, do you wonder, "How do these things work? How can we trust them?"

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. 

 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".

 A few places to learn more about our work and team at a high level are this introduction to Interpretability from our research lead, Chris Olah; a discussion of our work on the Hard Fork podcast produced by the New York Times, and this blog post (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 Mathematical Framework for Transformer Circuits, In-context Learning and Induction Heads, Toy Models of Superposition, Scaling Monosemanticity, and our Circuits’ Methods and Biology papers. This work builds on ideas from members' work prior to Anthropic such as the original circuits thread, Multimodal Neurons, Activation Atlases, and Building Blocks.

We aim to create a solid foundation for mechanistically understanding neural networks and making them safe (see our vision post). In the short term, we have focused on resolving the issue of "superposition" (see Toy Models of Superposition, Superposition, Memorization, and Double Descent, and our May 2023 update), 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 work 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.

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 Interpretability Architectures project that involves collaborating with Pretraining.