Original video in Chinese.
In 2026, if you want to build your own AI knowledge base, I strongly recommend that you adopt Karpathy’s approach: LLM Wiki.
First remember this conclusion of mine: the essence of this approach is upgrading AI from a “one-off retrieval generator” to a “continuous knowledge compiler.”
Let me walk you through it, and you’ll understand completely.
How did people do it before? Two approaches:
At first, they used RAG, which means splitting documents, vectorizing them, and then letting the large language model retrieve from them.
This approach has a high barrier to entry, and its limitations are quite large. So later, many people switched to tools like Claude Code, directly letting the agent traverse all documents.
Although these two approaches are completely different, their problem is actually the same: every time you ask a question, the AI has to read through the documents again.
“Having to rediscover knowledge every time” means token consumption will be huge, especially once your knowledge base gets bigger and bigger. And consistency will be very poor; those contradictions and cross-references have to be rediscovered every time, making it easy for hallucinations or omissions to appear.
So, in fact, the AI knowledge bases of the past two years can only be considered half-finished products. So what did Karpathy do right? He designed a system like this:
You have a material library that stores all the raw materials, such as articles, papers, notes, PDFs, and so on. AI will proactively “digest” these materials into a structured Wiki in Markdown format.
The content in the Wiki is not just simple copy-paste; it includes entity pages, topic summary pages, and so on. It is something generated after AI has completely understood all your content.
When you ask a question, AI will query the Wiki instead of searching through the raw materials every time. After AI answers, it can also save good answers back into the Wiki so knowledge continues to accumulate.
Later, whenever you add new materials to the knowledge base, AI will also proactively update the entire Wiki.
So, in your knowledge base, AI is no longer a secretary who has to check the materials every time it answers, but a knowledge base engineer. It compiles knowledge for you into a continuously evolving Wiki.
This is what an AI knowledge base should look like!
And it’s actually very easy to implement. Karpathy released a Gist. You give this link to AI, and then it knows what to do. I’ve already succeeded with it. In this video, let’s talk about the specific implementation method.
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can clearly explain both the Why and the How of AI. What I provide is far more valuable than tutorials. Remember to hit follow. If you want to connect with me, come to our newtype community. This community has been running for 700 days, and more than 2,000 friends have paid to join.
If you’re a domestic user, you can join through Knowledge Planet (Zhishi Xingqiu). If you’re an overseas user, you can join through Substack. My first course, daily Newsletter, and exclusive videos are all available in the community.
Back to today’s topic: LLM Wiki.
Karpathy’s approach applies whether you’re using Claude Code or another AI tool. For example, I’m using the newtype OS I built myself.
Because my knowledge base is already set up, I’m not trying to copy Karpathy’s structure directly, but to integrate the Wiki into it.
At the beginning, when I gave the link to the main agent — Chief — I asked it to make an assessment based on the current state of the knowledge base. The conclusion it gave was:
What Karpathy and I are doing is the same thing. For example, I already have an INPUT folder for storing raw materials; I have two Markdown documents, AGENTS and KNOWLEDGE, as guiding files. The only thing missing is the intermediate layer, which is the Wiki.
Although Chief suggested that I put the Wiki in the PROCESSING folder, I still decided to build the Wiki based on OUTPUT.
The OUTPUT folder stores my notes, articles, and video scripts. They are all finished products, written by me personally. Building the Wiki based on these materials is the most accurate approach.
So, although AI can help you handle many things, what this knowledge base should look like, the overall direction, must be determined by you. You have to use your brain and think it through carefully. Otherwise, it becomes the AI’s knowledge base, doesn’t it?
Once you’ve discussed it with AI and made everything clear, you don’t need to worry about the rest. AI will complete it automatically.
This process may take a long time, because AI first needs to digest all your content. I have three folders to process here. So Chief decided to split into three directions and create three background tasks for sub-agents to execute.
Once the comprehensive analysis from the three directions was done, Chief came up with a Wiki design plan. The whole structure was divided into several layers, such as the concept layer, which consists of concepts I originally created, repeatedly used, and fully argued.
After I confirmed the plan, Chief assigned the work and completed it in parallel. You can see this is the benefit of a multi-agent orchestration system.
I waited about 20 minutes for the whole process. It really was a big job. But in the end, it turned out pretty beautifully. Let me show you the finished result.
The things in the concept layer include various concepts I’ve mentioned before, such as AI OS, Code is intent, Everything is a file, and so on. Each concept’s Markdown document contains a very detailed, complete description. These were all extracted by AI from my past materials.
The things in the entity layer are the entities involved in various pieces of content. For example, Karpathy himself, and the Claude model, are both things I’ve mentioned often in the past.
The things in the thread layer are big themes that can be grouped together. For example, “The Three Acts of AI Engineering,” and so on. The content I wrote before really wasn’t independent topics; many of them can be grouped into one big theme or column. After AI fully understood them, it connected them together logically.
There are also several guiding files. For example, the Index document is structural, while the Log document is arranged by time. This kind of horizontal and vertical organization gives AI a very detailed understanding of the entire knowledge base, and it really is much stronger than simple search before.
The way the whole Wiki is used is also recorded in the AGENTS document at the lowest level of the knowledge base. From now on, AI will know when to read from or write to the Wiki. For example, when I save a new draft into the OUTPUT folder, AI will automatically process it and add it to the Wiki.
You see, after updating the knowledge base with Karpathy’s method, isn’t it much more powerful? The originally “dead” documents come alive after the Wiki is added. AI’s retrieval efficiency will be greatly improved. And it’s also very convenient for us to view in Obsidian.
So this is something extremely, extremely worth doing. Make sure to try it after watching the video.
OK, that’s all for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!