About
I am a Ph.D. candidate in Artificial Intelligence at KAIST, advised by Se-Young Yun and James Thorne, finishing a thesis on efficient long-context language models.
Context is what makes a modern model useful: the whole codebase, the whole conversation, a library of documents held at once. To hold it, the model re-reads everything: every token, at every step, in every layer and head. Does it need to? Mostly, no. My thesis lives in that no. SpotAttention teaches a small selector where a pretrained model already looks, so it can skip the rest; Attention-Span moves that choice into the model itself, declared as part of its own reasoning; and long-context speculative decoding is the third front, in progress, keeping the fast path fast when the context is not.
Before, I explored how much of a model’s work is actually necessary: pretraining models from scratch on multi-node TPU pods, folding long contexts into a handful of vectors, teaching models to say how sure they are. The exploration hardened into the thesis, and its machinery landed upstream in PyTorch/XLA and Hugging Face Transformers.
Not all of it was efficiency. The undergraduate teams I mentored asked a different question: how well text and image models represent cultures far from their training data. Their benchmarks made it measurable, from low-resource languages to image generation to what a model assumes from a face.
Systems
Research needs a place to exist. The lab’s GPU cluster began as scattered servers and became one system: one identity everywhere, storage that follows each user between nodes, a network that makes remote machines feel local, and a queue that stays fair in the week before a deadline. The same habit followed me home, where a smaller rack hosts everything I’d otherwise rent from the cloud.
Offline
A system built right does not need me around. The hours go to a motorbike, to calisthenics, to a camera, and to books: philosophy and anthropology and old novels, the kind that ask how humans and their systems work.