Date: Thursday, June 27, 2024
Speaker: Prof J. Mark Bishop
Title: Deep Stupidity: A Provocation on the Things LLMs Can and Cannot Do
Abstract: Artificial Neural Networks (ANNs) have soared to heights of "grandmaster" and "super-human" performance in games ranging from Go to Starcraft, heralding a new era of applications in business and becoming synonymous with the AI brand. However, when these systems falter-resulting in anything from autonomous vehicle accidents to biased automated processes-the repercussions quickly dominate headlines, raising questions about the reliability and understanding of AI.
Judea Pearl argued that the fundamental flaw is that these systems merely "fit curves" without understanding causality. Echoing this sentiment, Gary Marcus and Ernest Davis advocated for AI systems that inherently grasp concepts of time, space, and causality. Revisiting core themes from his provocative papers, Professor Bishop will discuss not just the limits of AI in terms of causality but will propose that the real issue is more profound: AI, in essence, does not "understand" anything at all.