Untangling the ineffable.
- Generative Models as a Complex Systems Science: how should we study computational systems that emerged rather than being designed?
- Predictive Interpretability: using models that are weaker than LLMs but predict aspects of behavior as explanations of LLMs.
- Machine Communication: studying the changing communication landscape as human-machine and machine-machine communication become more prevalent.
- Fundamental Laws of LLMs: can we take the Bell’s Theorem mindset and consider certain verified facts about LLMs to describe invariants about what they must be doing and learning?
- Making LLMs Reliable Tools: tolerance, error handling, delegation, automation.
Recruiting PhD students and Postdocs! See bottom for details.
What?
A research lab at the UChicago CS Department and Data Science Institute, founded on the belief that new conceptual models of how generative models (LLMs, image diffusion, etc.) work are fundamental to progress in building, interpreting, and using such models. The current AI ecosystem tends to favor engineering progress, theoretical guarantees, or complete mechanistic explanations. But as UChicago Philosopher of Science William C. Wimsatt noted, humans are limited beings and must work with piecewise approximations of reality. This lab is dedicated to exploring and exploiting new conceptual models of how generative AI works, and even how human communication does (or will) work.
We are one of the founding labs of the Communication & Intelligence group at UChicago, which recently had
Where?
The beautiful UChicago campus in the bustling city of Chicago, where you don’t need to own a car because public transport actually exists!
With UChicago’s push into AI, Data Science, and HCI starting in 2016, the university now hosts a number of professors thinking deeply about generative models in Computer Science (Chenhao Tan, Mina Lee, Rana Hanocka, & co.), Statistics (Victor Veitch, Aaron Schein, & co.), and other disciplines (James Evans in Sociology, YC Leong in Psychology, & co.). UChicago is known for unconventional, interdisciplinary collaboration, which is just what the study of generative models demands. Conceptualization Lab also has close ties with TTIC, an independent institute on the UChicago campus, where there are a bunch more folks thinking about generative models.
UChicago has an awesome intellectual discourse, where people are always talking about new ideas, the history of knowledge, and what the future of science, art, communication and everything else will look like. It’s a place to have discussions into the night at a teashop, a cocktail bar, or anything in between, as the city has a lot to offer and campus is right there in the middle of it all.
When?
All the time! I am actively recruiting PhD students to the CS Department PhD program and the Data Science Institute PhD program, whose deadlines are December 16th and 17th respectively (they are separate applications). I am also recruiting Postdocs via the DSI Postdoctoral Scholars program, which has will accept applicants on a rolling basis starting on 12/6.
Who?
I’m looking for clear thinkers who also love doing, enjoy viewing problems from different angles, and want to bring their (possibly non-CS!) perspective to bear on the phenomena we study. Previous experience with research, especially regarding generative models, is a plus but is not a requirement. During the PhD you’ll be expected to write code and formalize ideas mathematically (in papers and on a whiteboard), but the PhD will also help you get better at these things. If you feel like you need more work on one of these, are willing to do the work, and have a unique perspective to give, please apply! I’m interested in recruiting a diverse set of perspectives.
Unfortunately, I won’t be able to respond to most emails regarding PhD applications, as there are simply too many, but if you have a particular issue you can email me. If you’re interested in what I’ve written at the top of the page, I highly encourage you to apply!
Ari Holtzman, PI.
Why?
Why not?