From Co-reading to Co-creation: Hands-on LLMs and Localized Models at Twinkle AI Late-Night Study Session
- Time
- 2026-08-09 09:30 ~ 10:00
- Speaker
- Thomas Liang
- Room
- TR212
- Co-write
- https://hackmd.io/Bk61b4lIze
Abstract
As Large Language Models (LLMs) evolve rapidly, true understanding comes from hands-on coding rather than just reading papers. This session shares the technical insights and community experiences from Twinkle AI's "Late-Night Study Session." Using our open-source repository, TwinkleAI-LLM-Book-Club, based on "Hands-On Large Language Models," we will demonstrate how to transform a heavyweight book into interactive, community-driven materials.
We will walk through the mechanics of Tokenization and Embeddings, and showcase inferences using Twinkle AI's localized Traditional Chinese model (gemma-3-4B-T1-it). This session is a practical guide on lowering the barrier to entry through open-source collaboration and empowering the local AI ecosystem in Taiwan.
Speaker
Thomas Liang
Thomas 是一名跨足工程與設計領域的 ML Engineer，熱衷於探索機器學習底層技術。他目前是繁體中文開源語言模型社群 Twinkle AI 的活躍貢獻者，同時也是社群吉祥物「石虎」的創作者。作為「熬夜書坊（Late-Night Study Session）」的主持人，他致力於將艱澀的 AI 理論轉化為高重現性且易於上手的互動式教材。平時亦參與 Hugging Face Transformers、marimo 等開源專案，期望透過設計思維與「做中學（Hands-on）」，推動台灣在地生成式 AI 生態的發展。