University Seminar- Computability, Complexity, and Algebraic Structure-Universal Artificial Intelligence: Yudkowsky's Contributions to Solomonoff Induction and AI Safety

Time: Friday, February 14, 2:30 – 3:30pm

Place: Phillips Hall, Room 209

Speaker:  Keshav Srinivasan, GWU

Title: Universal Artificial Intelligence: Yudkowsky's Contributions to Solomonoff Induction and AI Safety

Abstract:  In this age of ChatGPT and other Large Language Models (LLM's), technology companies are racing to develop Artificial General Intelligence (AGI).  But this heightened fascination with AI has also brought renewed interest in the theoretical work of Marcus Hutter and others in developing AGI relative to an oracle for the halting set.  We will discuss Solomonoff Induction, a formalization of inductive reasoning which Hutter used to develop AIXI, a theoretical AI which surpasses all Turing-computable AI's in intelligence.  We will also review Eliezer Yudkowsky's proposed modifications to Solomonoff Induction and AIXI, based on logical non-omniscience, impossible world semantics, and Newcomb's paradox.  In the final part of the talk, we will assess Yudkowsky's claims of the relevance of these contributions to the AI safety problem and its potential resolution.