Research

I’m interested in the place where formal and informal systems meet: in particular, where the rigor and clarity of computer science collide with the fundamental, human-centered questions of philosophy and psychology.

My research pursues this junction through different perspectives (and fields). I’m interested in computational cognitive science—formal models of human cognition—and in AI alignment—formal models of human norms, desires, goals, and values. Recently, that has meant examining the implicit models of human decision making that are inverted by reinforcement learning from human feedback (RLHF), and exploring reward models as rich but flawed artifacts for expressing human preferences, subject to both computational and representational constraints. I am interested in how these models both succeed and fail.

My work has explored what chatbots do and don’t capture about conversation at its best (and worst); charted the computational structure of everyday decision-making dilemmas and shown that humans embody a resource-rational approximation of optimal solutions; and interrogated how reward models operationalize human preferences, norms, and values by turning normative and non-normative dialogues into scalar reward numbers. Through it all I am most engaged with pursuing rigorous notions of optimality, structure, and (mis)specification as they pertain to the core questions of the human experience. This is not my only interest, but after 20 years it is clear that it is inexhaustible.

There has never been a richer or more fascinating—or more urgent—time to pursue this particular set of questions. It is more than a life’s work, and it demands of us—I have always believed—that we be willing to trespass over traditional disciplinary lines to pursue these questions wherever they lead.