which will be given by Prof. David Smith of Northeastern University. David Smith is an Associate Professor in the Khoury College of Computer Sciences at Northeastern University in Boston. His interdisciplinary work focuses on natural language processing and computational linguistics, particularly through applications in the humanities and social sciences. He is currently a Visiting Professor at the Faculty of Historical and Cultural Studies.

As suggested by Alison Gopnik and others (e.g., Farrell et al., 2025), large language models and similar AI artifacts are "cultural technologies". Like language and writing—and also bureaucracies, democracies, and markets—AI transforms our relationship to memory and our interactions with each other. More particularly, state-of-the-art models are trained on archives collected from the digitized human record. Model trainers are recreating, intentionally or not, processes for selection, categorization, and source criticism that resemble some archival practices. After surveying some of the consequences of this archival view of AI, this talk will present work from our research group that traces the effects of training data composition on training dynamics and of mixtures of genres on high-level LLM capabilities. I will also discuss the ways in which analyzing large-scale patterns in the human record can help us build better models.

Attendance may be in person or online. A link will be circulated a few days before the event to online attendees, and refreshments will be available after the lecture for in-person attendees. Please indicate your mode of attendance during registration.