Skip to content

Reading NTU Hung-yi Lee's Machine Learning 2026 Spring: An Agent-First Course That Is Open Except for Grading

Sep 30, 20261 min
Home›ai
TL;DRHung-yi Lee's Spring 2026 Machine Learning course at National Taiwan University opens with OpenClaw. The first half takes apart AI agents, context engineering, inference speed-ups, and positional embeddings. The second half covers harness engineering, self-correction, and self-improving AI. Slides and recordings for all 8 lectures, plus PDFs and Colab notebooks for all 10 assignments, are public, so it rates A3. What's missing is grading: JudgeBoi returned 502 on 2026-09-30, NTU COOL is campus-only, and the three guest talks have no materials at all.
#ntu-ml-2026#ntu#ai-course#machine-learning#ai-agent
Series: Reading NTU Hung-yi Lee Machine Learning 2026 Spring (0 / 21)
Dissecting the Lobster: Hung-yi Lee Takes OpenClaw Apart Until Only Next-Token Prediction and a Few .md Files Remain

🌏 中文版

Machine Learning 2026 Spring is this year's machine learning course from Hung-yi Lee in NTU's Department of Electrical Engineering. It does not start with gradient descent. The first lecture dissects a "little lobster": the open-source AI agent OpenClaw. The syllabus puts it plainly: this year AI "doesn't just talk, it has started to act," so the whole semester takes the AI-agent point of view and focuses on "how to influence and adjust model behavior."

This post is the series entry point. It covers the course structure, what outside readers can get, how assignments are graded, and where to start. Lecture content is left to the later posts. The course is taught in Mandarin, and all official materials linked here are in Chinese unless noted.

Sources: the course page, the policy slides, policy.pdf (22 slides), the NTU course catalog syllabus for 114-2, and the bonus assignment slides, bonus.pdf. I opened and checked all of them on 2026-09-30. The matching video is the ML 2026 course introduction.

The hard facts

  • Course number and credits: EE5184, 4 credits, elective, Fridays 14:20–18:20 in room 博理 112. The catalog notes it is co-taught with 吳沛遠.
  • Format: slide 9 of policy.pdf says Lee lectures for about an hour, then TAs walk through assignments. Everything is recorded and expected online the following Monday. End time varies.
  • Prerequisite: the syllabus requires you to watch the recordings of the Fall 2025 course Introduction to Generative AI and Machine Learning (playlist, ten lectures of about two hours each). It also says anything Lee already covered on YouTube won't be repeated. Slide 8 draws the two courses as a staircase: Fall 2025 is the introduction, Spring 2026 covers frontier techniques not taught before.
  • Programming: all assignments use Python, but the course doesn't teach the language. Colab is enough; you don't need your own hardware.
  • Status: the last news item on the course page is "6/1 HW10 released," and the last deadline was 06/18/2026. The semester is over.

The arc: the visible agent → the invisible model → how to educate a model

The course page's content table has 11 rows. Eight have slides and recordings:

DateUnitPost in this series
3/6AI Agent (1): dissecting the lobsterPart 1
3/13AI Agent (2): context engineering, agent-to-agent interaction, impact on academic researchPart 3, Part 4
3/20Speeding up generation: Flash Attention, KV CachePart 6, Part 7
3/27Handling very long inputs: Positional EmbeddingPart 9
4/10Educating the model (1): Harness EngineeringPart 11
4/24Educating the model (2): Self-CorrectionPart 13
5/8Self-improvement (1): Self-ImprovingPart 15
5/22Self-improvement (2): Self-Improving -2Part 18

The other three rows are guest talks, covered under gaps below.

The order is already a good learning path. You first see what an agent does on your computer (visible). Then you ask why its input is limited and why generation is slow (inside the model, invisible). Finally you come back to the harness, which makes a model stronger without touching its weights, and to whether a model can fix its own mistakes and improve itself. This series follows the official order, with each assignment placed after the lecture it was released with.

The plan from the first week doesn't fully match what happened. Slide 7 of policy.pdf scheduled a talk by Professor 陳暐 of NTU Agricultural Economics on 5/29 and "Unit 5: generation strategies (Flow Matching in detail)" on 6/05. The actual course page has the Spoken LM TALK on 5/29 and Professor 陳暐's talk on 6/05. Flow Matching never got a lecture; only HW9 remains. If you're reading old notes, trust the course page.

Access rating: A3, except the grading chain

The rating follows the definitions in the global AI/CS course map (A0 schedule visible, A1 syllabus visible, A2 materials partly open, A3 enough to self-study). This course is A3. All 8 lectures have pdf and pptx slides and full recordings. All 10 assignments have a problem PDF, a publicly accessible Colab starter, and a TA walkthrough video.

What outside readers can't get (each assignment post repeats this):

  1. JudgeBoi: ml.ee.ntu.edu.tw returned 502 on 2026-09-30. For assignments submitted to JudgeBoi (per policy.pdf: HW1, 2, 4, 5, 7, 10), you can't get a score or see the leaderboard and private baselines.
  2. NTU COOL requires an NTU account. The quizzes for HW3, 6, 8, and 9 are answered on COOL, and the HW10 PDF also says to submit on NTU COOL. You can read the questions in the PDFs but not the answers.
  3. The three guest talks have no materials: the Appier Research team talk on 5/15, the Spoken LM TALK by students 楊書文 and 楊智凱 on 5/29, and Professor 陳暐's talk on 6/05. The course page lists titles only, with no slides or recordings. This series doesn't give them posts.
  4. HTML comments don't count: the page source hides rows such as an AI Cup info session, a TA session on training large models across multiple GPUs, and "Reasoning." They don't render on the page, so this series doesn't treat them as public course material for this semester.

Slide 20 of policy.pdf has the key line: "The only difference between taking the course and auditing it is that TAs don't grade auditors' homework." A self-learner is essentially an auditor, except that now even the self-service submission route is closed.

Ten assignments: schedule and platforms

The table on slide 11 of policy.pdf marks each assignment with three properties: answered on NTU COOL, "graded by an AI TA" on JudgeBoi, and needs time to train a model. The dates below match the course page and each assignment PDF:

HWTopicReleasedDueNTU COOLJudgeBoiTrainingSeries post
HW1LLM Malicious Instruction Defense03/0603/26OPart 2
HW2AI Agent as an AI Engineer03/1304/02OOOPart 5
HW3LLM Fast Inference03/2004/09OPart 8
HW4Training Transformer03/2704/16OOOPart 10
HW5Finetuning without Forgetting04/1004/30OOPart 12
HW6Model Editing04/2405/14OPart 14
HW7Model Merging05/0805/28OOPart 16
HW8Test-Time Scaling05/1506/04OPart 17
HW9Flow Matching05/2206/11OOPart 19
HW10Spoken Language Model05/2906/18OOPart 20

All deadlines are 23:59 (UTC+8). HW10 is the one row that doesn't agree: policy.pdf marks JudgeBoi, while the course page's platform column and hw10.pdf both say submit on NTU COOL. Go with the assignment PDF. Slides 12–13 also warn you up front that some assignments may need hours of training, and that "anxiously waiting for training results and tuning hyperparameters in a daze is the true flavor of training AI."

One trap: the text layer of slides 10–11 swaps the HW4 and HW5 dates (HW4 04/16–04/30, HW5 03/27–04/10), but the table as rendered on the slide is correct. If you scrape it with pdftotext, you'll hit this. Trust the rendered slide and the course page.

Grading, auditing, and asking for help

  • Grading: slide 14 says "10 assignments × 10 points = 100 points," and no grade-adjustment requests to the instructor at the end of the semester.
  • Enrollment agreement (slides 17–18): you accept that an AI TA grades assignments, with appeals allowed for errors; you accept randomness, so your training results may differ from the TAs' even if you follow instructions; Colab has usage limits, but free resources are guaranteed to reach a passing grade (C-, 60 points); extra compute makes high scores easier, but the course doesn't provide it.
  • Add codes and auditing (slides 15–16): total enrollment, including students already registered, is capped at around 700. Students in the EECS college and related programs come first, then students who took or applied to the 2025 Introduction to Generative AI and Machine Learning. Auditors are welcome but must fill in the form too.
  • Help (slide 19): use the NTU COOL forum, TA office hours (twice a week), or email. Don't DM the professor or TAs.
  • Academic integrity (course page, homework section): no plagiarism, no hand-editing prediction files, no sharing code or predictions. A first violation multiplies your final grade by 0.9 and zeroes that assignment; repeated violations mean an F.

Three tutorials and the bonus

The top of the homework table has three tutorials released on 3/6, all reused from earlier years:

Bonus: Teaching Monster Arena (TA 許筠曼, walkthrough video; the bonus slides are in English). The task is to enter Teaching Monster, a competition run by NTU's AI Center of Research Excellence (NTU AI-CoRE): build a fully automated AI teaching system that receives a course_requirement and a student_persona over an API and returns a download link to a teaching video within 30 minutes. Topics are physics, biology, computer science, and math for ages 12–18, benchmarked against IB and AP. Videos are mainly in English, at most 30 minutes long, with no human scriptwriting, editing, or voiceover allowed.

Scoring in bonus.pdf is per team: +2 for participating, +10 for the top 30%, +20 for the top three, +30 for the champion. Points are split among team members and added directly to the semester grade. There are two versions of the deadline: the course page says 05/15/2026 19:59, bonus.pdf says 2026/5/15 23:59:59. The organizers also released a baseline repo. teaching.monster returned 521 on 2026-09-30, so the competition site is currently unreachable.

How to start

  1. Do the prerequisite first: if you haven't watched the Fall 2025 introduction, watch it. The syllabus requires it; it isn't a suggestion.
  2. Read in this series' order: one lecture, then one assignment. Assignment posts cover the task, the starter code structure, and which grading parts you can't reach from outside.
  3. Set your own acceptance criteria: with JudgeBoi gone, copy the baselines and grading rules from the assignment PDF before you start, and write a small evaluation script as your feedback loop.

Related on this site: the NTU AI/ML course guide compares this course with the rest of NTU's offerings, and the Hung-yi Lee section of Which AI Courses to Take in 2026 places it among other courses.

Next: Dissecting the Lobster: How AI Agents Work, Using OpenClaw

References

Related

Reading NTU ML 2026: Context Engineering — Compression, Filtering, On-Demand Loading, and Whether to Hand the Context to the LLM1 minReading NTU ML 2026: How AI Agents Interact and What They Do to Work — Collaboration Topologies, Werewolf, Moltbook, and AI Writing and Reviewing Papers1 minDissecting the Lobster: Hung-yi Lee Takes OpenClaw Apart Until Only Next-Token Prediction and a Few .md Files Remain1 min