Description:

Monday / CS Seminar
January 5 / 12:00 PM
Hybrid / Mudd 3514

Speaker
Ari Holtzman, University of Chicago

Talk Title
Seeing Like a Language Model

Abstract
How does a language model perceive its input? What aspects of reality does it find legible and which elude it? How can we know? Current approaches to studying LLMs—focused on engineering progress—are insufficiently exploratory. I will discuss some projects touching on various aspects of LLMs: Absence Blindness, the localization of memories, and more sensitivity to subtle social signaling. I will discuss what it means for interpretability approaches to be predictive rather than mechanistic, defend prompting as a form of scientific inquiry, and caution against formalizing concepts too early, without doing the required amount of stamp collecting. Along the way I’ll make the case that AI has been given a new mandate—to articulate the ineffable, by describing aspects of communication and computation that we previously had no words for because they were stuck too deep inside human cognition to be easily referenced.

Biography
Ari Holtzman is an Assistant Professor of Computer Science and Data Science at the University of Chicago where he runs Conceptualization Lab. His lab's motto is 'Untangling the ineffable.' because he believes one of the primary purposes of all human endeavor is to create new vocabulary to help articulate the world more clearly. His research focuses on generative models of language, including contributions like Nucleus Sampling and work on understanding what language models are actually doing and how they perceive the world. Ari completed his PhD at the University of Washington, where he won the William Chan Memorial Dissertation Award and was part of the team that won the inaugural Amazon Alexa Prize in 2017.

Research Area(s)/Interests: LLMs, Machine Communication, AI Driven Narrative Art

---
Zoom Link
Panopto Link