Departmental Colloquium

 This is a featured talk

The questions of AI safety can seem very empirical, with the focus often being on (important!) topics like cybersecurity and politics. However, at its core the question of "how do we coexist safely with a superintelligence" is crying out for conceptual foundations, and there are fundamental theoretical questions to be worked on by mathematicians and other theorists.

We give an introduction to mathematical approaches to AI safety by focusing on two specific approaches: Safety-via-Debate (part of an approach known as "scalable oversight") and Open-Source-Game-Theory (a part of the theory of multi-agent dynamics), in which we present some new results.. The former asks for a protocol by which a less powerful computational agent may trust a more powerful one. The latter asks to understand interaction dynamics between agents that may directly model each other by accessing their source code. We motivate the relevance of both approaches to AI safety, and describe the underlying mathematical problems.

Note the room change: SS1069

Note different room: SS1069