AI 放大的是输出，不是输入：如何用 /learn 流程深入学习一个技术领域

I'd like to chat with you about how I deeply learn a technical field in the AI era. Before AI, it was more about reading books, browsing blogs by well-known figures in the field both domestically and internationally, and then taking notes in a notebook. The speed was quite slow, but the joy of learning was immense. For example, when I was learning WebGL, understanding one thing would take about half a year of spare time—slow, but happy.

Even with AI, I still dislike the Internet's 'learn One Hundred Years of Solitude in 3 minutes' style, and I don't care for short dramas or watching shows at double speed. I prefer choosing quality content, and I'd rather be slow and truly understand than skim a pile of summaries that leave nothing in my head.

However, recently while writing the 'Claude Code You Don't Know' and Agent series, besides the parts I already understood, there were many unfamiliar areas. Fortunately, I had collected plenty of related materials beforehand, so I seized the chance to clear my backlog—understand everything thoroughly and then output it. I've always felt that how much you read, how much you hear, how much you input is not the most important. What matters more is how much you can output. Only what you can clearly articulate, write down, organize, and publish truly becomes yours.

Not long ago, I dug myself a deep hole—researching the training workflow of large language models, aiming to make it understandable even for non-professionals. It took two weeks of exploration. This experience is perfect for sharing, and the write-up is almost ready; it will be published soon.

I organize the entire learning process like writing code. Collecting high-quality materials is the first step: recent top papers, key tech blogs posted by major model vendors, articles written by model leads on X, relevant courses from Stanford and similar universities over the past two years, and classic 'DIY large model' code repositories—all are my sources. Then, using tools, I automate downloading, converting to Markdown, cleaning, and organizing, sorting them into a dedicated repository for this research project. Before I even start reading seriously, I first clean up the entire information environment.

Next comes reading and filtering. For the content I can understand, I read through it carefully: if it feels low-value, I delete it; if it's good, I keep it. For what I can't understand, I directly ask Claude to help me comprehend it, or for more complex material, translate it into Chinese first before reading. Code that can run locally gets executed; code that can't, I study its structure to grasp the core ideas. At this stage, I don't pursue mastering every detail—just developing genuine awareness of the field and understanding the technical principles is enough. Usually by this point, about half of the original content gets deleted. That's normal. Filtering is part of learning; keeping the right things matters more than reading more.

At this stage, having a rough understanding of the field, I can start outlining the article. I think through what I want to say, which source materials correspond to each part, the sequence of content, and what the reader should walk away with after reading. An article is written for others, so you need to understand the reader's cognitive level, where they might get stuck, and what level of explanation is needed. It's similar to giving a presentation—you're constantly thinking about the audience.

Then comes the grunt work, much like reviewing before a college exam: fill in each section piece by piece, add missing explanations, and make the whole piece coherent. This step usually yields a long, somewhat wordy first draft. This is where AI is extremely useful. Without changing the original meaning or tone, it can help me cut useless fluff, fix broken transitions, spot logical gaps, and identify what background knowledge still needs to be added. In this process, AI isn't writing for me; it's helping me tighten the structure, reduce noise, and expose loopholes—and I often end up learning things I originally missed.

This is also why I believe AI is most useful when you have genuine output. If you only use it to summarize, it's easy to feel like you've learned a lot, but your mind actually holds nothing solid. AI truly helps when you are seriously writing something, explaining a concept, or building a finished product. It amplifies what you are already doing.

After the draft is tidied up, I read through it once more myself—not letting AI read it. AI is just a tool. The moment you let it replace your judgment, the whole thing loses its purpose. As I read, I continue to revise and tune, much like the self-testing feel of writing code: constantly finding weak spots, smoothing rough edges, and fixing places where the reading doesn't flow. After reading it through twice, feeling it's basically there, I send it out for everyone to see.

This is why I built Waza—an open-source collection of Skills built around my actual workflow. One of them is called /learn, and it's specifically designed for this process: collecting materials, filtering, outlining, filling in content, AI-assisted optimization, and self-review before publishing—the entire process linked into a single thread.

I'm sure some folks worry that no one will read what they write, so they don't want to publish—or even write at all. I don't think this is a good reason. If the content is meaningful, it will naturally find readers. Not necessarily right away, not necessarily many, but meaningful things are rarely wasted. 'No one will read it' is mostly just an excuse to avoid writing.

This whole process has given me a clearer realization: in the AI era, learning speed has indeed increased a lot, but depth will never come automatically. AI can help you collect, translate, clean, organize, compare, and condense—industrializing the entire process—but true depth still depends on your judgment, your patience, your standards, and whether you're willing to turn input into output. That hasn't changed. In fact, it's more important now than ever.