HongYi Lee

• The Most Complete Breakdown of LLM Positional Encoding: Absolute, Relative, Rotary, and No Positional Encoding at All

• From Flash Attention to Speculative Decoding: The Most Comprehensive Guide to LLM Inference Acceleration

• Hung-yi Lee 2026: Context Engineering — The Core Technology That Lets AI Learn to 'Forget' and 'Focus'

• Three Experiments, One Core Question: Can AI Agents Truly 'Talk' to Each Other?

PM

• You Know the Tech, but Don't Know What Agent Product to Build? I Found a Framework from 3 Case Studies

Github

• Turn Any GitHub Repository Into a System Design Course: This Open Source Project Makes It Happen

Agent

• Learn Claude Code (Part 1): Write an Agent Loop from Scratch — Less Than 200 Lines to Run the Full Model + Tools Pipeline

interview questions

• An Interview Question from a Small Company: MiniMax M3 Model Deployment and Concurrency Estimation

• Transformer Interview Essential: Detailed Breakdown of Self-Attention + Why It's Better Than RNN for Long Sequences

• the-three-levels-of-handcrafting-self-attention

Paper

• Why Can a 13B Small Model Solve Geometry Problems That GPT Gets Wrong? — A Close Reading of AAAI-25 GNS

• LLMs Don't Lack More Knowledge, But Reasoning Structures: A Close Reading of Pri-TPG