Active Aug 2026
How to Scale Your Model — 한국어판
Google DeepMind의 TPU/LLM 스케일링 교과서 'How to Scale Your Model'(Austin et al., 2025, MIT License) 전 12장을 한국어로 완역하는 프로젝트. 전문 용어는 영어 표기를 유지하고 그림·수식은 원본 그대로.
Project
Side quests
A free space for things built out of curiosity — away from paper deadlines, ablation tables, and anything that needs a review cycle. Some ship, some stay ideas; all of them are for fun (재미로).
Active Aug 2026
Paper Visualization
논문 하나를 완전히 이해할 때까지 뜯어보고 그리는 프로젝트. 같은 13개 질문으로 읽고 각 질문에 한 문장으로 답한 뒤, 문장으로 부족한 자리마다 인터랙티브 그림을 넣는다. 그림 속 숫자는 우리 A100에서 돌린 probe 실측값. 현재 DFlash(ICML 2026), GPTQ(ICLR 2023).
Active Jul 2026
This Website
The site you're reading — an Anthropic-inspired design system (ivory ground, ink text, one coral accent) on static Astro, plus an automated pipeline that turns a daily arXiv sweep into published paper reviews. The goal: a homepage that keeps growing even on weeks when I never touch it.
Idea Jul 2026
Market Flow Lab
Reading the market's tape the way I read a GPU profiler trace: regime detection on stock and crypto time series, order-flow and momentum signals, and backtests honest enough to hurt. If a signal can't survive transaction costs and a shuffled-label null test, it doesn't count as a signal.