Karpathy, LLM Wiki and My Notes - or Good Intuition in Obsidian

A little morning-coffee satisfaction

There I am, sipping my morning coffee, scrolling the feed, and suddenly I see the name: Andrej Karpathy. Co-founder of OpenAI, one of those people whose tweets pull half the AI industry along. I click. A gist. Title: “LLM Wiki”. I start reading.

And I read, and I read… and at some point I pause, because it starts to feel like I’ve read this somewhere before. And yeah — I have. On my own blog. Half a year earlier.

I’m not saying this to brag. There’s just something nice about it when an intuition you’ve been quietly developing on your own suddenly gets a stamp from one of the loudest names in AI. That little “hey, I was on the right track” feeling.

So what did Karpathy actually propose?

In short: instead of dumping the same files into ChatGPT every time and having it find everything from scratch (classic RAG), you build a persistent wiki in Markdown that the LLM itself maintains and expands.

The architecture is simple. Three layers:

1. Raw sources — your raw input. Articles, PDFs, notes. Immutable.
2. Wiki — a folder of .md files that the LLM creates, links, and updates.
3. Schema — a CLAUDE.md or AGENTS.md file that tells the LLM how to work.

And here’s the best part — Karpathy shows his setup: an LLM (Claude Code) on one side of the screen, Obsidian as the wiki browser on the other. He quotes himself: “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

Sound familiar? It did to me.

Why Did It Sound So Familiar?

Because I already have it. Not the whole wiki of course — but the foundation has been sitting there since December.

Take a look — on one side Karpathy’s theses from April 2026, on the other what I was writing on nomaid back in December 2025:

Bottom line? If you’ve got a decent Markdown note-taking setup — you’re already 80% of the way to an LLM Wiki. The rest is one CLAUDE.md file and a simple ingest script.

What does this change in practice?

My current workflow — editing the blog from my phone through Obsidian and Git — stops being just a convenient way to write on the train. It turns into a template for something much bigger.

Imagine this:

• You drop an interesting article into the raw/ folder.
• The LLM reads it. Updates the page about the author, the page about the technology, adds an entry to the index and to the log.
• A month later you ask: “what do I actually know about this topic?” — and you get a synthesis built from 15 different sources you’ve tossed in over the past weeks.

No RAG. No vector DB. Nothing.

And all of this on your files, on your disk, in your Obsidian. If Karpathy changes his mind tomorrow, if the LLMs disappear, if Obsidian falls apart — you still have your .md files. Because it’s still just plain text.

Is there no catch?

There is. More than one, actually. The discussion under Karpathy’s gist is heated, and it’s worth being honest:

• Scaling — with a few hundred pages the index.md stops fitting in the LLM’s context. You need to add BM25 or vector search.
• Knowledge drift — LLM-generated summaries are slightly lossy. Once the original drops off the radar, the compilation starts drifting from reality.
• Provenance — a generated wiki page looks identical to a hand-written one. It’s hard to tell at a glance which is the original and which is an LLM synthesis.

These are real problems. But for a single user working with ~100 sources? It’s still a whole class better than anything we’ve had so far.

Why this matters for you too

If you’ve been on the fence about whether it’s worth the time to learn Markdown or switch to Obsidian — this is the moment. Not because Karpathy says so. Because the whole AI industry is confirming that plain text + structure + your disk is the direction for the next decade.

Subscriptions will disappear. Apps will disappear. .md files will stay.

Links for the curious

• Karpathy’s “LLM Wiki” gist
• My earlier articles: Markdown — The Perfect Format · Obsidian — My Markdown Editor · Free Obsidian Sync

How about you — do you already have your own wiki in Markdown, or are your notes still scattered across five different apps?

Talk soon!