Direction

I am fascinated by how little we understand the internal mechanisms of advanced machine-learning models. Opaque optimization has proven effective at achieving impressive capabilities- at the expense of understanding what internal algorithms are learned by models. This opacity could become an increasing concern as capabilities improve and models are deployed in high-stakes settings. 

The north star of my research career is to better understand how concepts such as goals, and truth are represented within frontier models. While interpretability insights are valuable on their own, my focus is to translate them into practical model editing techniques. The ability to intervene on a model's internals is essential for rapidly addressing undesirable behavior without undergoing costly retraining. Model editing is analogous to the critical safety and utility benefits that patching provides in traditional software. I am especially interested in studying concept editing across model scale and distribution shifts so as to predict how such insights will generalize to future systems and environments.