Zhanbo Chen
Back to home

Relearning How to Learn

Learning Notes/AI Agent
2026-11-06
8 min read
0 reads

In brief

I was born in 2004. I grew up attending countless in-person classes and watching plenty of good online courses. By 2022, learning through video courses had become a habit. But that habit formed before we had LLMs like today's. Looking back now, starting with a video course may no longer be the best choice for me.

1. Why High-Quality Video Courses Are No Longer My First Choice

The idea for this article came to me a few days ago, when I wanted to learn PyTorch from scratch. Out of habit, I went to Bilibili to find a well-regarded video course. I opened Xiaotudui's PyTorch beginner tutorial, first published in November 2019. The first four videos spent nearly 45 minutes explaining how to set up a PyTorch environment, choose the right graphics drivers, and handle other setup details. In the past, dealing with these issues alone could easily take hours. Now, with AI agents, we can hand over almost all of that preparation. An agent can help choose compatible drivers and software versions based on our hardware, and help us judge which tasks benefit from a GPU and which can run perfectly well on a CPU. These used to be difficult decisions for beginners. That got me thinking: beyond having an agent set up my environment, could I also change how I acquire and understand knowledge?

I was born in 2004. I grew up attending countless in-person classes and watching plenty of good online courses. By 2022, learning through video courses had become a habit. But that habit formed before we had LLMs like today's. Looking back now, starting with a video course may no longer be the best choice for me.

2. How CS546 Made Me Rethink the Way I Learn

UIUC — CS 546 — Advanced Topics in Natural Language Processing is a course on advanced topics in NLP. The professor has extensive NLP research experience, and the course packs in material ranging from foundational knowledge to recent research. But I had never studied NLP systematically as an undergraduate. My background was limited to a single introductory AI course. There was a substantial gap between what I knew and what the course expected, and from the third lecture onward, I could barely keep up. I don't think language was the whole problem, though. The professor speaks with a Turkish accent, but I could clearly understand her suggestions in our project meetings. In lectures, by contrast, new concepts came so densely that we would often move on before I had understood the previous one.

That created a practical problem: the course had two closed-book midterms, and if I kept falling behind, I risked failing. Before each midterm, I spent about half a week studying the slides with an LLM for more than eight hours a day. My results—99/103 on the first midterm and 102/102 on the second—gave me some confirmation that I was absorbing the material. The learning process was somewhat painful, but I could feel my understanding of NLP improving.

Here is a quick account of how I used LLMs to study. First, I put all the slides covered by the exam into a folder and asked Claude Code or Codex to generate practice questions for each lecture. I would often get impatient and try the questions before reading the slides. In practice, that wasn't very helpful: without reading the slides first, I couldn't even understand the questions or the answer choices.

The slides themselves were packed with technical terms, and I lacked the background to make sense of them. So my first step was often to send an agent a screenshot of an entire slide and ask, "What is this slide even talking about?" That is a very vague question, but after reading the LLM's lengthy explanation, my brain would finally start to catch up with the slide. I would have an initial sense of what it meant.

Once I had that initial understanding, I would ask more specific questions. When reading a GRPO formula, for example, I would ask what each symbol meant and fill in the background knowledge I was missing. When studying dense retrieval and ANN, I would ask whether they were similar to database indexes. If a written explanation still wasn't intuitive enough, I would ask the AI to generate an interactive HTML demo of ANN. Once I could go back through the slides and felt that I basically understood them, I would write a short set of notes in my own words.

3. Why I Can Now Read Material I Used to Struggle With

I have always agreed with the idea that reading is the most efficient way for people to acquire information. But in the past, even though books and documentation let me control my reading speed, getting stuck still meant a lot of work: searching for explanations, sorting through sources, and filling gaps in my background knowledge. A well-organized video course, by comparison, saved me much of the effort of finding material and deciding what to learn first. That is why I used to rely more on online courses. Now, with an LLM, I can ask questions directly about the text, have it supply background knowledge, explain a formula, or try a different example. I can keep the freedom to read at my own pace without having to start another round of searching every time I get stuck.

Of course, starting with text doesn't mean looking only at text. Words on their own can still feel abstract. When reading papers and documentation, I also ask a multimodal LLM to explain the material alongside its figures. The illustrations in good papers and blog posts often contain a great deal of information themselves. If those still aren't intuitive enough, I ask the LLM to generate an interactive demo to help me understand the abstract concepts.

4. Two Important Conditions for This Approach to Work

Based on this experience, I think two things are especially important for learning efficiently with an LLM:

  1. Good learning materials. Part of the value of choosing a good course or teacher lies in how they select and organize the material. The CS546 slides, for example, helped me fill gaps in the basics while introducing research questions the professor considered worth paying attention to. They reflected her research taste. An LLM's ability to explain things does not replace that value.
  2. Self-discipline and confidence. I need to be willing to think about what I still don't understand, rather than treating a finished answer as something I have learned. I also need the confidence to ask very basic questions, keep asking when I am still confused, and have the AI check whether my own analogies actually hold up.

AI answered many of my questions during this process, but it couldn't always identify exactly where I was stuck or articulate the question I had in mind. To truly identify the gaps in their understanding, learners still need to think deeply for themselves.

5. What's Next: Applying This Approach to My Job Search

This approach worked well for studying CS546 and preparing for its exams. Next, I want to see whether it can help me apply what I learn to real projects.

Now that my goal is to produce something tangible, I also need to ask whether the learning paths laid out by popular online courses actually match my own goals. There are several ways to think about this, but I want to start with my personal development. In the AI era, I want the projects and experiences on my resume to reflect my interests, judgment, and abilities. If I simply follow tutorials to reproduce projects such as Heima Dianping or Cangqiong Takeout, I worry that it will be hard to show my ability to think independently and solve problems. Habit draws us toward whatever is popular, but no direction suits everyone, however popular it becomes. It is worth paying more attention to what we are good at, what we enjoy, and what sets us apart. Those interests may lead to more niche areas. For the specific questions I want to explore, I no longer want the existence of a ready-made video course to be a prerequisite for starting. I also want to try going directly to documentation, papers, and written tutorials.

I plan to prepare for summer internships and the fall recruiting season in 2027 by comparing my skills with the job descriptions for roles I am interested in. I will look at the practical problems those roles involve, find high-quality learning materials, and use LLMs to help me understand them. Then I will use project work and interview feedback to identify where I still fall short. I am also curious: without exams to define the scope of my learning and set deadlines for me, can I build an equally effective learning process on my own? That is what I want to explore next.

ZB

Zhanbo Chen

Java Backend & AI Agent Developer

Back to home
Comments