Hung-yi Lee's Machine Learning / Generative AI Lectures: Review

Who’s writing this, and why

I’m Lin Zhou. I spent years as a full-stack developer in the Chinese internet industry before moving into AI application work, and I’m bilingual — I read papers in English, but I built my deep learning fundamentals in Chinese, largely through this course. I came to it in 2024, right when Professor Hung-yi Lee pivoted his NTU lectures toward generative AI, because I needed the mechanics behind LLMs explained properly and couldn’t face another stack of arXiv PDFs.

What this course actually is

Lee’s public lecture series is probably the most influential deep learning course in the Chinese-speaking world, and it reinvents itself every semester. Spring 2024 became “Introduction to Generative AI.” Since spring 2025 the “Machine Learning” title has covered LLM/GenAI topics outright, and fall 2025 added a dedicated intro course. As of August 2026, the spring term wrapped in June and the fall page isn’t up yet — normal cadence, not a cancellation. Everything is free on the official course pages and YouTube channel: videos, slides, code, assignments. The official FAQ says public learners get “almost all of the content”; only homework grading and credits belong to enrolled NTU students.

The part English readers need to hear first: the lectures are in Mandarin (Taiwanese accent), and the slides mix Traditional Chinese with English technical terms. Per-episode English subtitle coverage on YouTube is unverified; Bilibili copies are third-party re-uploads, not run by Lee’s team. If your Mandarin isn’t functional, skip to the alternatives.

What’s inside

The 2025–2026 syllabus reads like a current LLM survey: agent principles and RAG, understanding and training Transformers from scratch, fine-tuning and catastrophic forgetting, RLHF and test-time scaling, model editing and merging, diffusion and flow matching, spoken language models, safety and fast inference. Lee’s style is intuition-first — dense with diagrams and pop-culture examples, light on derivations. Assignments are built to run on free Colab GPUs.

My honest read: it teaches you to understand, not to derive and not to ship. In Chinese learner communities the comparison with Lin Hsuan-Tien’s more mathematical course is a permanent fixture, and the conventional wisdom holds — go there for rigor. Classical ML (SVMs, decision trees) got compressed hard after the pivot; the 2021/2022 archives are where that material lives now. And the public audience gets no grading, no Q&A, no certificate of any kind. The feedback loop is entirely on you.

Who should take it

Chinese-speaking beginners and career changers, plus anyone who wants to follow agents, RLHF, and diffusion without reading English papers. The fall 2025 intro course officially needs no prerequisites; the spring ML course goes smoother with basic Python, and official Colab/PyTorch primers are provided.

Is it worth it for an English reader? In exactly one scenario: you already follow Mandarin and want the friendliest current survey of the field. For everyone else the language barrier is real and the credential doesn’t exist. Andrew Ng’s Coursera specializations (Chinese subtitles available) give you structure, labs, and a certificate, though their content predates the LLM era. Datawhale’s happy-llm pairs well if you later want to hand-build a model.

Bottom line

If you understand Mandarin, this is the strongest free AI lecture series available anywhere — current, complete, famously easy to start. If you need math rigor, a certificate, or English instruction, it was never aimed at you.