Efficient AI
Training-free, plug-and-play methods. I would rather find a win you can bolt onto a system that already runs than one that only pays off after another round of training.
I'm a computer science senior at UNIST, working on efficient AI.
This September I'm staying on for my M.S./Ph.D. at the Ubiquitous AI Lab with Prof. Taesik Gong, where I've already been doing research as an undergrad since January. The part I like is getting a model to do more without paying for it twice: methods that need no extra training, that drop into a system already running, and that pass work between the device and the cloud when that helps.
Lately that has meant GUI agents. They can operate real software on their own, which is impressive right up until you notice every step is a full model call. I'm writing my first paper on cutting that cost while keeping what the agent can actually do.
A line I keep coming back to: “If you seek death you will live; if you seek life you will die.”