One of the fascinating discoveries about AI's inner workings is that neural networks represent meaning with geometry. Words that are close in meaning sit near each other in a model's representation space, and relationships between words, such as antonym pairs or the mapping from "man" to "woman," appear as directions in that space.
Two ideas push this observation further. The linear representation hypothesis argues that many concepts are encoded as such directions, and the platonic representation hypothesis suggests that different models, trained on different data, converge toward the same underlying representation of the world. If so, the geometry that these models learn may reflect the structure of the culture they were trained on. This talk offers a gentle introduction to the geometry of AI's semantic space and then shows, through several examples, how it can be used to read culture at scale.

Yong-Yeol (YY) Ahn is the Quantitative Foundation Distinguished Professor at the School of Data Science, University of Virginia. He studies complex social, cultural, and biological systems by combining network science, machine learning, and large-scale data. His work has examined the flavor networks behind world cuisines, how rap lyrics reflect social conditions, how beliefs fit together and spread, the science of science, and the geometry of AI representations, and has appeared in journals including Nature, Science, PNAS, and Nature Physics. Before joining UVA in 2025, he was a professor of informatics at Indiana University Bloomington and a visiting professor at MIT. He received his Ph.D. in physics from KAIST and was a postdoctoral researcher at Northeastern University and the Dana-Farber Cancer Institute at Harvard. He is a Kavli Fellow and a Microsoft Research Faculty Fellow, and has received the ACM IMC Test of Time Award and the Indiana University Trustees' Teaching Award.