Meet The Postdoc is a monthly series highlighting the AI Lab’s postdoctoral researchers.

Kawin Ethayarajh is a postdoctoral researcher for Princeton Language and Intelligence, where he works on understanding how AI will operate as a part of human systems. This summer, he will start a position as an assistant professor of applied AI at UChicago Booth. He holds a doctorate in computer science from Stanford University.

What brought you to the AI Lab?

I was looking to do a brief postdoc before starting a faculty position, and I was already familiar with some of the faculty at PLI.

What was your Ph.D. thesis about?

My thesis was on "Behavior-Bound Machine Learning." Traditionally, we think of AI being bound by what hardware and software can do, but I argued that what it can accomplish in the real world is often bound by the behavior of real-world actors, such as workers and firms. By trying to understand this behavior and incorporating it into how we develop AI, we can create systems that work well in practice and not just under idealized conditions.

What has the experience at the AI Lab been like so far?

Amazing! There are a ton of interesting people to collaborate with, as well as a lot of autonomy to shape your postdoc experience.

How has your postdoc experience differed from your Ph.D.?

I was fortunate to have a lot of autonomy during my Ph.D., but I feel that I have even more of it during my postdoc. PLI is also unusually well-resourced for an academic institution, so I'm able to do compute-heavy work that would've been much harder to do during my Ph.D.

Tell us about your research you’re working on. How does it contribute to solving real-world problems?

 After LLMs are trained, they are post-trained with real-world feedback. This feedback can be subjective (e.g., user preferences) or objective (e.g., whether the LLM-generated code passes some tests). There are a bunch of practical bottlenecks to getting this to scale, and I'm working on making this process more computationally efficient by drawing from findings in behavioral economics on how humans perceive probability.

 How do you think the experience at the AI Lab will advance your career goals?

The AI Lab is a great way to spend a year between your Ph.D. and faculty position to explore the blue-sky ideas that you've always wanted to explore. I've been using it to explore ideas that will lay the foundation of my work as a junior faculty member.

Do you have any advice for someone just starting a postdoc?

Postdocs can be greatly varied, so make sure you know what the expectations are, what resources are available, what you want out of the experience, etc.

What do you like to do outside of work?

Read, mostly. My appetite is fairly diverse, but I quite like magical realism that is grounded in history. I'm currently reading Midnight's Children by Salman Rushdie. 

What’s your favorite snack at the AI Lab?

I'm a ruminant so I'll graze on whatever is lying around. No strong preferences there.