Original video in Chinese.
Key Takeaway
- AI will become the standard for measuring personal ability, and the ability to use it will become a basic skill.
- The author shares an AI learning method: first deconstruct the topic, sort out initial judgments and questions, and provide AI with context.
- Use Gemini’s Deep Research feature to generate multiple detailed reports, then import them into NotebookLM through Google Docs for AI-assisted learning.
- Use Cursor together with Gemini 2.5 Pro to integrate, condense, and dehydrate all the materials, and finally output them as a Markdown file.
- Emphasize that Markdown is the most suitable file format in the AI era, making it easy for both humans and machines to understand and for long-term storage.
- The entire AI learning process significantly improves learning efficiency and widens the gap between people.
Full Content
Very soon, AI will become the standard for people.
What AI generates will be the standard of mediocrity. If what you produce—say, an article or a report—can’t beat AI, then you’re mediocre in that area.
The ability to use AI is the standard for a basic skill. If you can’t use AI in your work, it’s like not knowing how to use Office software today. If you can’t use AI in your learning, it’s like not knowing how to use a search engine to look up information today.
I’ve always felt that today’s AI is already strong enough. If we update one more generation along this trend—namely the GPT-5, Claude 4, Gemini 3.0 generation—AI will reach a mature state. Whether AGI can be achieved doesn’t matter, because what already exists is enough to profoundly change human society.
Most people still don’t realize this. That’s why I keep making videos—to filter out and gather the people who can see the future. In this episode, I’ll use a concrete example to share how I currently use AI to learn. If you feel something after watching, remember to join our community.
OK, let’s begin.
When I want to go deep on a certain topic, I don’t ask AI directly. I first deconstruct it myself.
For example, I have questions about AI PCs. If I just ask AI right away, it would most likely be: please generate a report on AI PC development trends.
Frankly speaking, that approach is very inefficient, ineffective, and very unskilled. I know without even looking at the result that it will definitely be a bland, watered-down report that plays it safe.
The right way is to first go through it in your own mind, deconstruct it, and sort out your initial judgments, general understanding, and the questions you most want answered. For example, regarding AI PCs, there are two things I especially want to know:
First, is AI PC a pseudo-proposition?
I know that on the most critical compute issue, AI PCs allocate and schedule CPU, NPU, and GPU resources. But is the NPU really reliable? Is it really not just dead weight? I need to put a question mark on that on my side.
So the first question is actually about the category itself, involving category definition and industry standards.
Second, can AI PC actually develop?
I know that at present, besides Intel pushing it, Qualcomm and AMD are also doing it. Although I still don’t know the specifics, based on experience in the field, there’s a high probability that these three companies each have their own routes, architectures, and toolchains. That will create a lot of optimization headaches for developers, because the standards aren’t unified.
So the second question is actually about the industry landscape and ecosystem development.
You see, this deconstruction process is actually a process of folding in your own thinking context. For any topic, everyone has different understanding and different priorities. If you don’t deconstruct it, you can’t distill it, and you can’t give AI more context. Then how could AI possibly generate what you want?
As I said in the community before:
In the AI era, the answers are all there—if only you can ask the right questions.
Once the deconstruction is done, I can use Deep Research to generate multiple reports, each with its own focus. Using the AI PC example, I asked Gemini to generate two reports for me.
Why two instead of combining them into one? Because a single report can be more detailed and more focused. And there will definitely be overlap between the two reports. They might even complement or verify each other.
After the reports are done, they can be exported to Google Docs. A couple of days ago I mentioned in the community that this is a capability and experience no one else has. Because once exported, they can be added to NotebookLM.
NotebookLM is currently the best AI learning tool. I’ve recommended it many times. It’s especially suitable for scenarios with textbooks and documents.
In the lower-left corner of this panel, we can load from Google Docs. The two Deep Research reports I just exported are in there.
Wait a few seconds and the model will finish parsing. Then you can start a conversation. For example, I would ask it: Is AI PC a pseudo-concept, or a tax on stupidity? Is the NPU dead weight? What are the route and architecture differences among Intel, Qualcomm, and AMD in developing AI PC?
If these answers are OK, you can pin them and turn them into notes.
I know some people may still complain that Deep Research results are not that good. But in my view, whether the feedback is good or bad, it’s still feedback, and it has value.
For example, Gemini must have searched hundreds of web pages to produce these two reports, covering all the public reporting on AI PC right now. So if these two reports aren’t that good, I’d still be happy—because that represents the current views of the media and self-media, and it means the current market consensus is problematic. And I’m going to dig out a more correct non-consensus view that is more likely to become the market consensus in the next stage.
If you have an investing or startup mindset, you’d probably get very excited when you encounter this kind of consensus deviation. Of course, ordinary people don’t; they just complain.
Alright, let’s not go off track. My AI learning process isn’t finished yet, so let’s continue.
Through the Q&A just now, with the help of NotebookLM, we’ve digested the materials pretty well and saved some notes. The next step is for me to let AI help me integrate all the materials into one piece, including the two reports originally generated, as well as the notes I saved during the discussion that interested me.
The main reason for doing this is that learning isn’t something you study once and you’re done—later you still need to review it, and you need to learn and then practice it from time to time. When needed, I’ll definitely come back and look through it again. So I need to integrate everything produced in this process into one complete thing. And I also need to make it concise and dehydrated, keeping only the most essential parts, so the next time I review it, it will be more efficient.
To achieve this, I use Cursor together with Gemini 2.5 Pro.
I also said in the community before that Gemini 2.5 Pro may be the most overlooked top-tier model right now. Its reasoning ability is very strong, it has all the skills it should have, and it also has a super-large context window. So it’s especially suitable for integrating reports and the like.
I downloaded the two reports just now from Google Docs to my local machine. Then I used Cursor to open the folder where they were stored. In the chat box, I told it: help me integrate these two reports into one, and be sure to make the logic clear and the content concise. Note that the core viewpoints and arguments must not be omitted. Output the final result in Markdown format.
Markdown is the most suitable file format in the AI era. Humans can read it, and AI can understand it too. I strongly recommend that from now on, everyone convert any valuable material you want to keep for a long time into Markdown format.
We can just watch the process. Cursor will complete it fully automatically. Compared with generating code, it’s just text work, which is absolutely no problem for Cursor and Gemini.
Finally, once I get the integrated Markdown file, I’ll store it in my Obsidian note repository. In this way, the AI learning for a topic is complete.
Let me recap the whole process.
First, I deconstruct the topic. This step is extremely important. Because we need to tell AI our understanding of the topic and the questions we’re interested in as context. Otherwise, AI can only give you ordinary answers.
Next, I give Gemini a prompt that includes my thinking context and ask it to help me do Deep Research. Recently Google updated the Deep Research model to Gemini 2.5 Pro, and this feature has become even better to use. Personally, I feel it has already caught up with OpenAI, with each having its own strengths.
Then, after getting the reports, I import them into NotebookLM through Google Docs so I can do AI-assisted learning.
Finally, I hand all the materials to Cursor and, with the help of Gemini 2.5 Pro’s super-large context, perform global understanding, complete the condensation and integration of the materials, form a Markdown file, and save it.
You can see that with AI’s support, the efficiency of the whole learning process is far, far higher than the traditional way. The same old line: this is how the gap between people gets widened.
OK, that’s it for this episode. If you want to discuss AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!