Open-Ended and AI Generating Algorithms in the Era of Foundation Models

Jeff Clune, a Professor of Computer Science at the University of British Columbia and Senior Research Advisor at DeepMind, discusses the exciting new opportunities presented by foundation models for developing open-ended and AI-generating algorithms. He explores how these models can drive continuous innovation and learning in AI systems. The talk covers concepts such as Quality-Diversity (QD) algorithms, like MAP-Elites and Go-Explore, which aim to produce a diverse set of high-quality solutions. It also delves into open-ended algorithms, exemplified by POET, designed for endless innovation, and the AI-generating algorithms approach where AI creates better AI. The presentation highlights recent work including Omni (Open-endedness via Models of human Notions of Interestingness), VPT (Video Pre-Training), and OmniEpic for generating diverse environments and tasks. Clune also touches upon the potential for AI to automate the scientific process and design agentic systems, emphasizing the shift from hand-crafted pipelines to learned ones.