Researchers used to define objectives for artificial intelligence (AI) agents by hand, but with progress in optimization and reinforcement learning, it became obvious that it's too difficult to think of everything ahead of time and write it down. Instead, these days the objective is viewed as a hidden part of the state on which researchers can receive feedback or observations from humans — how they act and react, how they compare options, what they say. In this talk, Anca Dragan, Associate Professor of Electrical Engineering and Computer Sciences at UC Berkeley, discusses what this transition has achieved, what open challenges researchers still face and ideas for mitigating them. Dragan discusses applications in robotics and how the lessons there apply to virtual agents like large language models. Series: "Data Science Channel" [Science] [Show ID: 39350]

Ver todos los episodios