Kellogg and AI Learning Lectures: How LLMs Understand Text: Context and Meaning in the AI Age

This event is part of a series, “Kellogg and AI Learning Lectures”, which is organised by Kellogg College and the AI Competency Centre at Oxford. Open to all Oxford University staff and students, these sessions are designed to boost understanding of developments in AI and build confidence in using emerging technologies.

Large Language Models represent an entirely new paradigm of semantic information processing. They are the first time a computer can process the meaning of text without mark-up. Their ability to “understand” text is at the same time familiarly human and completely alien.

They can write poems and extract subtle clues from opaque texts but at the same time they fail at seemingly trivial tasks like counting words in a sentence. They seem to possess deep knowledge of the world yet fall for the most trivial of spatial puzzles. They can write code that will emulate a calculator but not multiply numbers. Based on casual use in a seemingly similar context, two people can easily come to the conclusion, that the models are practically infallible or entirely useless.

This talk, presented by Dominik Lukeš, an AI Consultant at the AI Competency Centre, will outline how Large Language Models “read” and “understand” text, and how they generate their response. This contrasts with how humans perform the same tasks (which is not always the same as how they perceive them). The talk will explore how humans deploy attention, working memory and external tools to deal with complex text in contrast to the models which only use a form of attention.