HumAI: Can Language Models Represent the Past? How Would We Know?

About the seminar

Most commercial chatbots are trained overwhelmingly on text drawn from the Western (especially, English-language) internet. When models try to represent the perspective of another place or time, they often produce a caricature, still really seen from the outside. In this talk, I will suggest that the challenge of producing “historical language models” that can understand the past from within may give us insight into the broader task of helping them understand cultural difference. To make things more interesting, we will explore the generative Talkie-1930 model, which simulates conversation with someone from the past.

Ted Underwood will give his talk on zoom and if you wish to participate in the lecture hall the seminar will be on a big screen and with a moderator in the room.

The Talkie-1930 model: https://talkie-lm.com/chat

Blog post introducing the model: https://resobscura.substack.com/p/are-vintage-llms-the-start-of-a-new

Ted Underwood is Professor of Information Sciences and English at the University of Illinois, Urbana-Champaign. He has written three books on literary history and is currently working on ChronoLogic, a benchmark that evaluates a language model’s understanding of historical contexts from 1800 to 2000 by asking it to draw inferences and solve problems like a writer with a specified social background at a particular place and point in time.