Title : Scaling Laws and Transformers: Modern LLMs for the Physicist’s Mind
일시 : 2025년 11월 25일 (화), 17시
장소 : 아산이학관 433호
Speaker : 이준현(삼성 리서치)
Abstracts:
This talk introduces modern large language models (LLMs) through the lens of concepts familiar to physicists. We will explore how Transformers, the architecture behind models like GPT, leverage attention mechanisms to process and generate text, and how scaling laws — empirical regularities governing model performance as data and parameters grow — have driven the remarkable advances in AI capabilities over the past decade. Along the way, we will draw connections to ideas from statistical physics, high-dimensional geometry, and information theory, illustrating how a physicist’s intuition can shed light on why these models work and where they are headed.