Multidisciplinary Institute in Artificial intelligence

THE "HARD PROBLEM" OF ARTIFICIAL INTELLIGENCE

ABSTRACT

Everyone is talking about Artificial Intelligence these days: automatic image recognition, autonomous vehicles, ChatGPT, DALLE-E-2 and their potentially transformative effects on society. We even hear regularly about The Singularity, the fateful day when computers become smarter than us, widening the intelligence gap between us and them so far that the survival of the human race might be endangered. A number of well-known and respected computer scientists, starting with Ray Kurzweil, chief scientist at Google, believe that "The Singularity is near” and that we need to start preparing for it. In contrast, other leading researchers (Rodney Brooks, Melanie Mitchell, Gary Marcus, etc.) think that computers still have a long way to go. My opinion on the matter lies between these two poles. But no one could reasonably deny that recent advances in AI, and, specifically, in ChatGPT, Midjourney, etc., are both promising and worrisome. With the help of examples from everyday life, I hope to show that there are still significant difficulties to overcome, whether it be in the field of image recognition or text/image generation and that these difficulties are related to the lack of understanding of these programs. But what exactly does "understanding" consist of? The key question for AI in the 21st century will be to know to what extent these programs will be able to acquire an understanding of the world similar to ours relying only on billions of texts, images and sounds drawn from the Internet, without ever having perceived the world as we perceive it, through our visual, tactile, olfactory, auditory, and gustatory senses. I believe that there is still a long way to go before these programs will achieve what we would call true AI, even if the AI that eventually emerges is not necessarily identical to human intelligence.

ROBERT FRENCH

Robert M. French is an emeritus Research Director at the CNRS.  He has worked for the last four decades in the area of computer modeling of human cognition.  His research is multi-disciplinary, having published work in analogy-making, in catastrophic forgetting in artificial neural networks, in child development, and has written numerous articles on the Turing Test. He has also published technical articles on various topics in machine learning. He holds a B.S. and Master's degree in mathematics, worked as a translator in Paris for many years, culminating with the co-translation into French of Douglas Hofstadter's Pulitzer Prize-winning magnum opus, Gödel, Escher, Bach.  Hofstadter invited him to join his AI research group in the U.S. and he accepted, doing a doctorate under his supervision on the computational modeling of analogy-making.  After a post-doc in psychology at the University of Wisconsin, he accepted a position as professor of statistics in the psychology department at the University of Liège in Belgium.  In 2004, he joined the CNRS as a Research Director and since that time has worked at the Laboratory for Research on Learning and Development (L.E.A.D., UMR 5022) at the University of Burgundy.

VIEW THE REPLAY