Slides
English slides, Mandarin delivery — 1 of 14 weeks ready
The slides are written in English to accompany the Mandarin lectures. Each deck is a reveal.js page — open it in a browser, nothing to install.
提示Keyboard shortcuts
B chalkboard layer for writing on a slide · ESC slide overview · F full screen · add ?print-pdf to the URL to print
| Week | Topic | Slides |
|---|---|---|
| W1 | From n-grams to Seq2Seq: A Compressed History 從 n-gram 到 seq2seq：一條被壓縮的技術史 |
▶ Open |
| W2 | Tokenization: The World the Model Sees Tokenization：模型看到的世界，以及它看不到的 |
Not yet available |
| W3 | Transformer I: Attention Mechanics, Complexity, and Building It by Hand Transformer I：attention 的機制、複雜度與手刻 |
Not yet available |
| W4 | Transformer II: The Full Block, Parameter Accounting, and Training Dynamics Transformer II：完整 block、參數記帳與訓練動力學 |
Not yet available |
| W5 | Positional Encoding, Long Context, and the Limits of Context 位置編碼、長文本與 context 的極限 |
Not yet available |
| W6 | Architecture Routes: MoE, SSMs, and Linear Attention 架構路線：MoE、SSM 與 linear attention |
Not yet available |
| W7 | Pre-training: Scaling Laws and Data Curation Pre-training：scaling laws 與資料策展 |
Not yet available |
| W8 | Post-training I: SFT, PEFT, and Forgetting Post-training I：SFT、PEFT 與遺忘 |
Not yet available |
| W9 | Post-training II: Preference Alignment and Reinforcement Learning Post-training II：偏好對齊與強化學習 |
Not yet available |
| W10 | Reasoning and Test-Time Compute Reasoning 與 test-time compute |
Not yet available |
| W11 | Inference Efficiency: From FLOPs to Memory Bandwidth 推論效率：從 FLOPs 到 memory bandwidth |
Not yet available |
| W12 | RAG and Context Engineering RAG 與 context engineering |
Not yet available |
| W13 | Agentic Systems: Tools, Planning, Failure Modes, and Security Agentic systems：工具、規劃、失效與安全 |
Not yet available |
| W14 | Evaluation, Interpretability, Safety, and Open Problems 評估、可解釋性、安全，與 open problems |
Not yet available |
Three of the decks turn on a demo that only works if you have not seen it coming — a reward with a hole in it (W9), a hidden hint the chain of thought never mentions (W10), and an agent broken by prompt injection (W13). Those are posted after the lecture rather than before it.
回到頂端