Karpathy Proposes LLM-Wiki Framework to Revolutionize Knowledge Management

Saving does not equal ownership; highlighting does not equal understanding.

Those deeply insightful articles that excite you at 2 a.m., those densely interconnected bidirectional links in Obsidian, and those meticulously organized databases in Notion—all are merely "cyber mummies" lying dormant in note-taking apps.

The graph may appear impressive, but it has long been rotten.

This is a systemic failure in the age of information overload.

Karpathy, an engineer at Anthropic, former co-founder of OpenAI, and former AI director at Tesla, couldn't stand it anymore and dropped a bomb.

Gateway: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

He didn’t announce a new model or release a new framework—he simply said: treat your notes as immutable source code and let the LLM act as the compiler.

Two months later, this document has sparked a quiet yet intense migration across the Obsidian, Claude, and Cursor communities.

Someone has already expanded their wiki to hundreds of pages and hundreds of thousands of words.

Automation plugins are beginning to emerge. Academics, independent entrepreneurs, and lifelong learners are collectively shifting toward an entirely new relationship with knowledge production.

Before LLM-WIKI, the mainstream solution was RAG (Retrieval-Augmented Generation).

In simple terms, it’s like giving a large model a “research assistant”: when you ask a question, it searches through your notes for a few relevant snippets and pieces them together into an answer.

Sounds great, but anyone who's used it knows the gap between the seller's showcase and the buyer's reality.

It's just a courier: RAG can only handle local information and cannot understand the big picture.

It can tell you that the fifth note mentions A, but it cannot reveal the underlying logic common to all 500 notes.

It suffers from "split personality": if you believed A was correct six months ago but wrote notes yesterday refuting A, RAG often falls into self-contradiction, producing a jumble of illogical nonsense.

Graph rot: Manually maintained knowledge links are like code without automated cleanup—over time, broken links accumulate everywhere, and search efficiency declines exponentially.

Karpathy’s intuition is sharp: search and retrieval are signs of human limitation. What we need is "consensus," "structure," and "truth."

Karpathy's answer, from an action that programmers perform daily but never think about in terms of knowledge: compilation.

You write a piece of source code; the program does not reread the code each time it runs.

You compile it into a binary file; the compilation process is difficult this time, but every subsequent run is extremely fast. The cost of compilation is amortized over thousands or even millions of future uses.

Why can't knowledge be done this way?

Karpathy says to treat your raw notes as immutable source code and treat the LLM as a compiler, letting it compile this messy pile of material into a structured, interconnected wiki in one go.

Each time new material is added, the AI performs an integration: updating relevant entry pages, revising summaries, highlighting conflicts between new data and existing conclusions, and optionally reinforcing or challenging prior judgments.

The key difference is this: knowledge is compiled once and then continuously maintained, rather than being rebuilt temporarily with each query.

By the time you’re ready to ask questions, the cross-references are already there, the contradictions have already been flagged, and the review already reflects everything you’ve read.

You don't recompile the source code every time you run a program. So why should you make the AI reread your notes every time you ask a question?

Fundamental shift in the mode of cognitive production

In his LLM-WIKI framework, notes are no longer static text, but rather "source code."

Large models are no longer dictionary-style translators, but rather "compilers."

This architecture elegantly achieves three-layer decoupling:

1. Raw Layer (Raw Materials): This is your raw source of inspiration—your spontaneous insights, clipped articles, meeting notes. It is “immutable,” preserving the originality and unpolished nature of human input.

2. Schema Layer (Knowledge Constitution): This is the "code of conduct" you write for the AI. For example, you stipulate that each character entry must include "motivation, limitations, and key achievements"; each technology stack must outline "advantages and disadvantages."

3. Wiki Layer (Compiled Output): This is the area fully maintained by AI. It transforms your raw, disorganized data into structured, cross-linked, and logically consistent encyclopedia pages based on your schema.

Every day, there are just three actions:

1. Ingest: Drop in a new article, and the AI reads it, walks you through the key points, writes a summary, and updates related pages across the entire library—one source can trigger updates to over a dozen pages.

2. Query: Directly ask the compiled Wiki for answers with cited sources. The best part: a great answer can be directly turned into a new page—every exploration you make compounds over time.

3. Lint (Health Check): Regularly have AI self-audit like a code review—identify contradictions, outdated claims, orphaned pages with no links, and missing gaps. Clean them up early to prevent the library from decaying over time.

You are no longer a mere transporter of knowledge, but the architect of this empire of wisdom.

You are only responsible for input and final review; the AI handles all the "grunt work": organizing, aligning, cross-referencing, and conflict detection.

This is a fundamental shift in the relations of production.

This is not another chatbot. ChatGPT knows the internet; LLM-Wiki knows you—specifically, what you've taught it.

Each response comes with [wiki-links] back to your knowledge graph. Each reply is a starting point for an exploration path, not an endpoint.

By now, you might think this is just a smart workflow.

More than that.

At the end of the gist, Karpathy casually mentions Vannevar Bush and his 1945 essay, "As We May Think."

In 1945, shortly after World War II, this leading American scientist imagined a machine called "Memex":

A mechanical desk that can store all your books, records, and correspondence, and establish "associative pathways" between related items—connections between documents, as valuable as the documents themselves.

Does that sound familiar? It’s almost a word-for-word description of LLM-Wiki.

Bush's vision was actually closer to this than the later World Wide Web: a private, personally curated network of knowledge where connection equals value.

Then why wasn't Memex built in eighty years?

Because Bush was stuck on a problem he couldn't solve—who would maintain it?

Each association path must be manually created. Each cross-reference must be connected by someone.

Bush imagines there are dedicated "operators" who lay paths for you through knowledge.

In reality, no one can sustain this tedious, labor-intensive work at scale. Humans will abandon maintenance because the cost of maintenance always rises faster than the value it generates.

Karpathy’s statement is the key to the entire paradigm: the most exhausting part of maintaining a knowledge base has never been reading—it’s bookkeeping.

Update cross-references, keep summaries current, highlight conflicts between new data and previous conclusions, and ensure consistency across dozens of pages. This tedious work is enough to deter everyone.

Large models won't forget to update a single cross-reference and can modify 15 files in one go.

It never gets tired. Never gets frustrated. Never wears down in the middle of the night. Maintenance costs are reduced to nearly zero.

So, the machine that had stymied humanity for eighty years suddenly began to turn.

Looking back, LLM-Wiki is the third piece of Karpathy’s puzzle on “human-machine collaboration,” and also the most restrained.

First block, Vibe Coding (February 2025): Accept code written by AI without line-by-line review; trust the model and test the results.

Second part, Agentic Engineering (January 2026): Humans orchestrate AI agents instead of writing code themselves.

Third section, LLM Knowledge Bases (April 2026): AI is no longer managing just code, but knowledge itself.

In this new paradigm, what humans are relieved of are the tedious tasks no one enjoys—collecting, organizing, linking, and bookkeeping.

All that remains for humans are two things: deciding what to read, and figuring out what it all truly means. These are precisely the two things machines cannot do—and should never do for you.

This is the story of a tool that evolved to its extreme, ultimately circling back to return human attention to humanity itself.

That plain, almost annoyingly basic Markdown file didn't send any models or climb any leaderboards.

It simply offered a quiet reminder: your brain was never meant for bookkeeping.

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.