My work focuses on understanding the representations and algorithms underlying human and machine cognition. For the first time, we have access to models that can flexibly accomplish complex cognitive tasks across a wide range of domains. My research uses insights from cognitive science in order to uncover the mechanisms supporting this seemingly-intelligent behavior, and, reciprocally, uses techniques from mechanistic interpretability to better characterize the similarities and differences between minds and machines. Ultimately, this research program aims to transform black-box neural networks into explicit and useful cognitive models of linguistic and visual processing, while also driving the development of more human-like artificial intelligence systems.

Here are some specific questions that I like to think about:

• How do modern neural networks seem to produce compositional/symbolic behavior?
• To what extent do language models learn to represent a coherent model of the world?
• How can we best employ pretrained models as (components of) cognitive models?
• How much of human cognition is truly symbolic, and how much is statistical?
• What is the role of inductive biases in modern AI? Is scale all you need?