About
Matter that senses and responds
To me, understanding the origin of biological life is one of the most fundamental questions we can ask as a species. Of all the seemingly stochastic processes occurring in the universe, the ability of matter to organize, self-replicate, and gain functional complexity is the one that allows us to ask this question in the first place. Much of my scientific curiosity has grown from wanting to understand how interactions at the smallest levels of matter give rise to the behavior of an entire organism. That desire to connect mechanisms with their consequences now draws me to questions about artificial intelligence.
Biological selection involves creating environmental conditions that select for a specific subset of organisms or macromolecules like RNA over other organisms or molecules based on their properties. Model training creates selective pressure that selects for models that exhibit a certain kind of behavior over models exhibiting other kinds of behavior.
The connection between biological selection and model training gives me a unique way to frame questions: What are we actually selecting for: the desired behavior or a desire to pass a test by any means necessary? What alternative behavior could satisfy the same test? What evidence would distinguish a system that has internalized the intended principle from one that has learned to pass the test?
In the early 1940s, a decade before the first transistor computer was built, Isaac Asimov philosophized about the ethical constraints that should guide the behavior of thinking machines. The elegance of Asimov’s laws of robotics is the ambition to make ethical responsibility intrinsic to a machine’s behavior, instead of something constrained by external guardrails or scoring mechanisms. The difficulty is in carrying that ambition into situations its designers did not anticipate. I want to understand how training (or even fundamental changes to LLM architecture) can develop reliable constraints on behavior, how those constraints can be investigated, and where they break down.
The emergence of function from chemistry has fascinated me for much of my life. With AI, I want to understand how the processes that produce increasingly capable systems can also produce behavior we can reliably trust. I see AI safety as a natural continuation of the questions that first brought me to research, with the added responsibility of helping shape complex, intelligent, autonomous systems we are deliberately bringing into the world.
