On LLM Wiki, compounding knowledge, and what it means to actually understand something

If, like me, you are something of a note taking nerd, then you are likely aware of the content collector's fallacy. This is the habit of collecting information but not actually using it, and it is not a new concern. Before the internet, the students who took the best notes in class were not always the ones with the highest test results. They diligently collected the information but could not apply it during an exam.

This same problem persists today in the Personal Knowledge Management community, where people have written endlessly about note creation approaches like atomic notes, PARA, and Zettelkasten. The rapid adoption of AI has only exacerbated the problem, and while it has never been easier to generate and organize your notes, it brings us no closer to answering the question: so what?

With that in mind, I was intrigued last week when I heard that Andrej Karpathy had posted about something called an LLM Wiki, which proposed automating the whole note creation process, highlighting the difference between storing knowledge and building it.

What an LLM Wiki actually is

The wiki is structured around three core components. Raw sources are your documents — articles, papers, transcripts, data files. They're immutable. The LLM reads them but never touches them. The wiki is a directory of LLM generated markdown files: summaries, entity pages, concept pages, comparisons, synthesis. The LLM owns this layer entirely. The schema is a file that tells the LLM how the wiki is structured and what workflows to follow. The wiki can be viewed through a markdown reader like Obsidian and queried through an agent.

When you add a new source, the LLM doesn't just index it. It reads it, extracts key information, and integrates it into the existing wiki — updating entity pages, revising topic summaries, noting where new data contradicts old claims. A single ingest might touch ten to fifteen pages. Cross-references are maintained automatically. Contradictions get flagged. The wiki grows over time.

How this differs from a second brain

I've been running an Obsidian-based PKM for a while now, using Readwise to collect highlights from my Kindle and Instapaper accounts. I create atomic notes based on these highlights as well as any other ideas I want to record. This is where I also journal and track my side projects. So my first question when I read the LLM Wiki proposal was: how does this compare to what I'm already doing?

The core benefit of a traditional second brain is that value comes from processing the source material yourself. Writing an atomic note is a cognitive event, not just storage. The extra step is meant to help you clarify your thinking as well as link your individual notes. In an LLM Wiki, the LLM does that processing for you. This is faster and more scalable. But it raises the question: does the knowledge compound in you, or just in the wiki?

What I'm building

I've been thinking about this as two separate vaults with a bridge between them.
My main Obsidian vault stays what it is: my mind. Atomic notes, my synthesis, my writing. A personal thinking layer.

The wiki becomes a compiled research layer. I drop documents in. The LLM structures and organizes them using the patterns above. I query it when I need depth on a topic without cluttering my personal vault with raw material.

The bridge runs one way intentionally. When a wiki synthesis genuinely shifts my thinking, I write an atomic note in my personal vault capturing my take. The wiki gave me the material; the note captures my understanding. There's a lot of reference material I want queryable but don't want to process personally. The wiki can hold that, but my vault holds what I actually think about it.

Getting to the So What Question

Collecting and organizing your own knowledge base takes time, and for people of a certain disposition it can be meaningful in its own right. But for many of us, we want our note taking efforts to serve some purpose, and I think we can borrow an idea from Richard Feynman.

He kept a list of twelve open questions. Not topics he was interested in, questions he actually needed answered. Problems he couldn't let go of. He updated the list slowly; these weren't passing curiosities but live concerns that shaped how he read, what he noticed, what he remembered. Whenever he encountered a new result, technique, or finding, he'd run it against the list. If it didn't connect to any of his twelve questions, he filed it away without much attention. If it connected he followed it hard. The questions acted as a filter that transformed information into material for thinking.

The result wasn't just that Feynman learned efficiently. It's that his knowledge compounded in a direction. He wasn't accumulating randomly. He was building toward something. This is what the synthesis layer of an LLM Wiki could actually do, if you set it up right.

In the schema, synthesis pages are stored answers: the wiki answers a question, you decide it's worth keeping, it becomes a page. That's useful but passive — you're querying against what's already there. Feynman's method suggests something more active: encode your open questions in the schema itself, and let them direct what the wiki synthesizes.

In practice, this might look like a section called "Active Questions" — a short list of problems you're genuinely working on. When the LLM ingests a source, it checks it against your questions. If there's a connection, it creates or updates a synthesis page explicitly tied to that question. The wiki doesn't just accumulate; it accumulates toward something.

The wiki becomes most valuable not as an archive but as a research instrument pointed at live problems. Feynman's twelve questions are the difference between a library and a laboratory.