Karpathy's LLM Wiki Went Viral. Here's What It Means for Your Personal Knowledge Base.

The post that broke the timeline

On April 3, 2026, Andrej Karpathy posted something on X that traveled far beyond the AI crowd. He described a shift in how he uses LLMs, moving from generating code to generating knowledge structure. He showed a system where raw research materials go into a folder, an LLM compiles them into a structured wiki, and a single research topic had grown to roughly 100 articles and 400,000 words without Karpathy writing a single word of it directly.

Two days later, he followed up with a GitHub Gist he called an “idea file.” It describes a pattern for building personal knowledge bases using LLMs. Within days, the post had spawned a wave of community implementations and serious debate about whether this approach makes RAG pipelines obsolete for personal use.

One line from the gist explains why it went viral. “Humans abandon wikis because the maintenance burden grows faster than the value.”

Everyone who has ever built a wiki, personal or professional, recognizes that sentence. The creation feels productive, the maintenance feels like busywork, and eventually the busywork wins and the wiki rots.

Karpathy proposed a fix. Let the LLM do the busywork.

How the LLM wiki works

The architecture is simple on purpose. The foundation of this system is a two-tier directory structure with a raw/ directory and a compiled wiki. Raw sources like articles, papers, images, and data files are immutable. The LLM reads from them but never modifies them. This is your source of truth.

The LLM compiles those raw sources into a wiki directory of interlinked markdown files. The LLM owns this layer entirely. It creates pages, updates them when new sources arrive, maintains cross-references, and keeps everything consistent.

Instead of building a traditional RAG system, Karpathy’s approach treats the LLM as a compiler that reads raw source documents and produces a structured, interlinked wiki. The wiki itself becomes the knowledge base, with no embeddings or vector search needed at the scale of a personal knowledge base.

To keep things clean, Karpathy employs LLM “health checks,” automated passes that look for inconsistent data, fill in missing information using web search, and identify potential connections for new articles.

The human curates sources, directs the analysis, asks good questions, and thinks about what it all means. The LLM’s job is everything else.

Karpathy uses Obsidian as the front end. In practice, he has the LLM agent open on one side and Obsidian open on the other. The LLM makes edits based on their conversation, and he browses the results in real time, following links, checking the graph view, reading the updated pages.

The gist itself is worth paying attention to. Karpathy didn’t share a repo or an app. He called it an “idea file” and published the concept intentionally abstract, intentionally vague, so that anyone can hand it to their own agent and get a version built to their own situation.

As Glen Rhodes put it: “That is a genuinely different thing than what we’ve been doing for the last twenty years.”

The gist is a description of a concept, specific enough to be actionable, loose enough that an agent can fill in the implementation details based on your particular tools, preferences, and context. You hand it to your agent and say “build me this.”

The community responded fast. Within two days, there were multiple open-source implementations, including a Claude Code plugin implementing Karpathy’s LLM Wiki pattern with commands for ingesting, querying, and linting. Developer Farza built “Farzapedia,” a personal Wikipedia compiled from 2,500 entries across his diary, Apple Notes, and iMessages. The result is 400 articles covering research areas, people, projects, and ideas, all interlinked, all maintained by AI.

The maintenance problem is personal

The problem Karpathy described is one everyone faces, not just researchers. Anyone who has tried to maintain a personal wiki, a Notion database, an Obsidian vault, or even a well-organized notes app has hit the same wall.

You build it, stop maintaining it, and it rots.

The failure mode is structural. The tedious part of maintaining a knowledge base is the bookkeeping that involves updating cross-references, keeping summaries current, noting when new information contradicts old claims, and maintaining consistency across dozens of pages. None of it requires deep thinking, but all of it takes time, so it doesn’t get done.

Karpathy’s fix works because it offloads exactly that layer. The LLM handles the bookkeeping. The human handles the judgment.

Your personal knowledge base doesn’t have to rot

The pattern isn’t only for academic researchers with 400,000-word wikis. It applies to anyone trying to keep their knowledge organized, at work or in the rest of their life.

Your work notes. Research for a side project. An idea file you’ve been meaning to maintain for two years. A personal knowledge base covering everything you’ve learned about a domain. These all fail for the same reason Karpathy diagnosed. The maintenance burden outgrows the habit.

What makes Karpathy’s system different is that the wiki compounds over time instead of decaying. New sources go in, the LLM updates the compiled output, cross-references stay current. The knowledge base gets better as you use it, not worse.

Promptless is built around this same principle. You feed it your notes, documents, links, and research, and it maintains the compiled knowledge base for you. It handles the cross-references, the consistency checks, the updates you’d otherwise skip. You stay in the judgment seat.

The result is a personal wiki that doesn’t rot. Your idea file stays current. Your research compounds. Whether you’re organizing work knowledge or personal projects, the bookkeeping happens automatically.

Karpathy proved the pattern works. The question is whether you want to keep doing the maintenance manually.

← Back to Blog