Embedded Realities: Visualizing AI Embeddings Across Time and Language
This project set out to bridge the gap between AI research and public understanding by translating complex, evolving discourse into experiences people can see and feel.
Developed through the Duke Bass Connections Program, the initiative brought together graduate and undergraduate researchers across Computer Science, Statistics, and Visual & Media Studies under the guidance of Brinnae Bent and William Seaman. The team was organized into three interdisciplinary tracks: Historical Drifts, Geolingual Shifts, and AI Representation.
The Historical Drifts and Geolingual Shifts teams used large language models to analyze how AI discourse evolves over time and across languages. In parallel, the AI Representation team explored new media approaches for translating these findings into immersive, visual forms.
The project culminates in a public exhibition at the Rubenstein Arts Center, featuring two digital installations that bring these insights to life. As part of the AI Representation team, I designed and built the projected visualizations in TouchDesigner, transforming abstract data into interactive, spatial experiences. The installations are currently in testing and refinement, with the exhibition opening on April 28.
Watch our highlight reel here: https://www.youtube.com/watch?v=lnxlCOY_Jvc
Advisors
Dr. Brinnae Bent
Dr. Bill Seaman
AI Representation
Alexis Golart
Qifeng Cheng
Madeline Tsang
Historical Shifts
Bochu Ding
James Sohigian
Geolingual Drifts
Junyu Zhang
Vivienne Foley
Neha Shukla
Location
Duke Rubenstein
Arts Center
Timeline
2 Semesters
Installation 1: Temporal Tunnel
The Temporal Tunnel transports viewers through the evolving history of AI discourse, immersing them in how AI has continuously shaped, and been shaped by, the world around us. As viewers move forward through the space, they travel along a living timeline, physically progressing through decades of conversation around AI.
The visualization is built by transforming embedding data in TouchDesigner and linking it to ultrasonic sensors that track the viewer’s position in real time. Each point represents a word drawn from scholarly and public discourse for a given moment in time. The viewer occupies the role of “AI” itself, positioned at the center, with surrounding words dynamically arranged in relation to it—revealing how meaning, proximity, and context shift over time.
Visual inspiration taken from this tutorial: https://www.youtube.com/watch?v=xU4Xb2A-bAo
Installation 2: Geolingual Growth
The Geolingual Growth exhibit invites viewers into a living landscape of language, where meaning takes root and grows differently across cultures. Each interaction begins as a single concept—such as “family”—which unfolds into a branching tree that reveals how that idea is expressed and associated across multiple languages. As viewers engage with the space, individual trees emerge and gradually expand into a dense forest of interconnected meanings.
The visualization is built by mapping word embeddings across languages and translating these relationships into organic, root-like structures. Each branch represents a different language, with nearby words positioned according to their semantic proximity within that language’s embedding space. Using handheld objects and sensor-based triggers, viewers activate the growth of each tree in real time—bringing abstract linguistic relationships into a tangible, evolving form that highlights both shared meaning and cultural nuance.
Final Exhibit
The final exhibit, Lost in Embedding Space, was shown at the Duke Rubenstein Arts Center on April 25th. Both installations were open to the public, encouraging visitors to learn about the ways data and AI shape our lives.
Design Progression
