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’s mission is to reverse engineer how trained models work, and Interpretability research is one of Anthropic’s core research bets on AI safety. We believe that a mechanistic understanding is the most robust way to make advanced systems safe.
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, or as treating neural networks as binary computer programs we're trying to "reverse engineer".
We aim to create a solid scientific foundation for mechanistically understanding neural networks and making them safe (see our vision post). 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 which found millions of features in Claude 3.0 Sonnet, one of our production language models, represents progress in this direction. In our most recent work, we developed methods that allow us to build circuits using features and use these 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 Claude Haiku 3.5, one of our production models.” This is a stepping stone towards our overall goal of mechanistically understanding neural networks.
A few places to learn more about our work and team are this introduction to Interpretability from our research lead, Chris Olah, Stanford CS25 lecture given by Josh Batson, and TWIML AI podcast with Emmanuel Ameisen.
Some of our team's notable publications include and our Circuits’ Methods and Biology papers, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, Towards Monosemanticity: Decomposing Language Models With Dictionary Learning, A Mathematical Framework for Transformer Circuits, In-context Learning and Induction Heads, and Toy Models of Superposition. This work builds on ideas from members' work prior to Anthropic such as the original circuits thread, Multimodal Neurons, Activation Atlases, and Building Blocks.
