#frAIday: Is LLM Meaning an illusion?

#frAIday hybrid
This lecture will be held both on Zoom and in Galaxen, Universum at Umeå University. We will offer free lunch wraps - limited number available, so come in time to secure yours. Welcome!

Abstract
When you ask one of today’s Large Language Models a question, it gives you what seems like a meaningful answer. But is this just an illusion? It has been suggested that because LLMs believe nothing, understand nothing, and intend to communicate nothing, they cannot produce meaningful texts. (Bender et al 2021, Bender and Koller 2020) Rather, it is we, human users of these models, who project meaning into the texts they produce. In this talk, I investigate this suggestion using a well-known distinction in the philosophy of language, between “subjectivist semantics” and “consumerist semantics” (Kaplan 1989). Subjectivist semantics holds that linguistic events and objects (e.g. utterances, inscriptions, signings) are endowed with meaning by the beliefs and intentions of those who produce them. Consumerist semantics holds that linguistic events and objects get their meaning from the language that is used, independently of the users’ beliefs and intentions. I argue that the view that LLM meaning is an illusion stems from an exclusively subjectivist approach to meaning. However, linguistic meaning is both subjectivist and consumerist. In the case of LLMs, it is exclusively consumerist. Appreciating this highlights an important difference between LLM meaning and human meaning, without reducing LLM meaning to an illusion.