I’m fascinated by intelligence in all its forms. Broadly speaking, I want to understand how it works. Intelligence already exists in nature, and I think much of our job is to explore and discover its principles. In my day-to-day research, I do this by studying neural nets mechanistically, in much the same spirit that we study how brains and other biological systems process information.

Of course, understanding matters: it’s basic to doing science, and it matters beyond AI safety and governance. AI models are becoming tools for scientific discovery; work built around them has already won Nobel Prizes. But if we come to rely on their predictions without understanding how they arrive at them, what happens to science as a search for explanations? I want us to learn from these systems, not just use them. Opening them up and mapping their internal workings may be one of our clearest paths towards that understanding—and towards making advanced AI safe and governable.

Some questions I’m working on include:

• When an AI system does something unexpected, can we trace it to the internal circuits or representations responsible? And is there a useful level of abstraction at which to explain its behaviour?
• How can we detect subtle failures that standard benchmarks might miss—including behaviours described as deception, reward hacking, or specification gaming? I know some of these terms can sound anthropomorphic, but the underlying behaviours still need to be studied.
• What would give us confidence that a safety intervention has actually worked, rather than just making an unwanted behaviour less visible during evaluation?
• What concrete evidence would independent auditors and regulators need to verify safety claims before a system is deployed?
• As we build and use more capable systems, how do we avoid gradually handing over important decisions in ways that weaken human control, democratic norms, and freedom?

Beyond interpretability, AI safety, and governance, I’m drawn to computational cognitive science, philosophy of science, and poetry, among many other things. My work has been published at NeurIPS, ICML, ICLR, ACL, EMNLP, FAccT, and Science Robotics. You can find my publications on Google Scholar, or read more about what I hope to work on in my Research Agenda.

My work is supported by OpenAI, Anthropic, Schmidt Sciences, NVIDIA, and others.