Dimitri Coelho Mollo: "Philosophical Issues around Large Language Models"

Dimitri Coelho Mollo, Umeå University: "Philosophical Issues around Large Language Models"

Abstract
Large Language Models (LLMs)---such as ChatGPT, Claude, Copilot---have brought back into relevance the fierce philosophical debates between symbolicists (Fodor, Pylyshyn) and connectionists (Smolensky, Rumelhart, McClelland) that took place in the 80s. The key issue then, as today, centres on the question of cognitive architecture: what sort of cognitive organisation underlies human (and biological) intelligence? can it be replicated in artificial systems? In addition, the dissemination of LLMs across society raises a series of ethical questions importantly tied to the nature of these systems.

In this talk, I will present some of my work on core theoretical and ethical issues around Large Language Models. I will argue that current AI research has a special, explorative role to play in advancing our understanding of cognition and intelligence that partly differs from what can be achieved through cognitive and comparative psychology---a role that is nicely illustrated by LLMs. I will also argue, based on recent co-authored work, that a popular method to try and achieve safe and ethical LLMs, namely Reinforcement Learning through Human Feedback (RLHF), is plagued by fundamental problems.