CS Distinguished Lecture: Tuomas Sandholm (Carnegie Mellon University, Strategy Robot, Inc., Optimized Markets, Inc., Strategic Machine, Inc) - What Can and Should Humans Contribute to Superhuman AIs?
I will discuss what humans can and should contribute to superhuman AIs—not general ones intended to be like humans, but ones for specific applications that make the world a better place. I will discuss how the application should drive research. I will present extensive experiences from having fielded superhuman AIs for combinatorial markets, organ exchanges, and imperfect-information game settings. I will discuss inventing and scoping novel AI applications. I will discuss how humans should supply the value framework while leaving policy optimization and combinatorics for AI. I will cover a framework that separates those ends and means, and conducts future-aware optimization in very-large-scale dynamic problems in a scalable way. I will wonder about the future of science when theorems (not just proofs) and empirical theories need to be so long that they are beyond human comprehension. I will discuss human overconfidence in humans over AI. I will discuss what explainability could be and why in many AI applications it should not be required. Finally, I will suggest flipping ethics around from an ex post discussion activity to a system-design discipline that I coin pre-design ethics.
This lecture satisfies requirements for CSCI 591: Research Colloquium.