Biography

I am a CS PhD candidate at the Center for AI Safety at Stanford University, under the guidance of Prof. Clark Barrett. I want to understand how values form. In reinforcement learning, decision theory, and game theory, the utility function is fixed at the outset and sits outside the agent, whereas the way values form in organisms, within a lifetime and across lifetimes, is highly structured. My view of AI alignment follows from this: values should be raised, not specified. Most recently I gave a definition of exploration that is not relative to any terminal goal: one policy explores at least as much as another if running it eventually teaches you what would have happened under the other. I used it to show where standard intrinsic rewards fall short and when they succeed. I am now asking whether placing care into an AI’s environment leads to robustly good values.

Before this, I trained language models to use their chain-of-thought as a natural-language hidden state, so that the reasoning is load-bearing for prediction (ICLR 2026). I came to alignment from formal verification and programming languages at Stanford, after studying interactive theorem proving at CMU with Simon DeDeo, where we published in Cognition on why abduction is needed to believe the corpus of known mathematics. I majored in computer science and electrical engineering at MIT.

My primary character trait is curiosity, and I really love math.