AI vibe coders have one more reason to be thankful Andrey Karpathythe originator of the term.
Tesla’s former AI director and co-founder of OpenAI, which now runs its own independent AI project, recently Published in X describing a "LLM Knowledge Bases" approach he uses to manage his various topics of research interest.
By building a sustainable, LLM-protected record of his projects, Karpathy addresses his primary frustration. "stateless" AI development: resetting the dreaded context limit.
Anyone with Vibe code can attest that hitting a usage limit or ending a session often feels like a lobotomy to your project. You are forced to spend precious tokens (and time) rebuilding the context for the AI. "remembers" the architectural nuances of your new build.
Karpathy offers something simpler and freer, mixed elegant than the typical enterprise solution of a vector database and RAG pipeline.
Instead, it describes a system where the LLM itself functions as a full-time "research librarian"— Actively compile, lint and link Markdown (.md) files, the most suitable and compact data format for LLM.
By diverting a significant part of it "token throughput" Karpathy laid out a plan for the next phase, moving into the manipulation of structured knowledge rather than code. "Second Brain"- self-healing, verifiable and completely human-readable.
Outside of RAG
Over the past three years, it has been the dominant paradigm for providing access to proprietary data to LLMs. Search-Augmented Generation (RAG).
In a standard RAG setup, documents are shredded arbitrarily "fabrics," converted into mathematical vectors (connections) and stored in a special database.
When a user asks a question, the system a "similarity search" Find the most suitable pieces and LLM them. feeding into what Karpathy called the approach LLM Knowledge Basesnegates the complexity of vector databases for medium-sized datasets.
Instead, it builds on the LLM’s increased ability to think about structured text.
System architecture as viewed by user X @himanshu Part of a wider response to Karpati’s mission, it operates in three distinct phases:
-
Data reception: The raw materials—research papers, GitHub repositories, databases, and web articles—are thrown together.
raw/directory. Karpathy uses Obsidian Web Clipper to convert web content to Markdown (.md) files, even images, to be stored locally so that LLM can reference them through its visualization capabilities. -
Compilation stage: This is a major innovation. Instead of just indexing files, LLM "compiles" they are. It reads raw data and writes a structured wiki. This includes creating summaries, identifying key concepts, creating encyclopedia-style articles, and most importantly, back links between related ideas.
-
Active Maintenance (Linting): The system is not static. Karpati describes running away "health checks" or "rolled" LLM goes through where it scans the wiki for inconsistencies, missing information, or new links. As a community member Charlie Wargnier observation, "It actually acts as a living AI knowledge base that heals itself."
Viewing as Markdown files "source of truth" Karpathy runs away "black box" vector placement problem. Each claim made by AI can be traced back to a specific claim .md a file that a person can read, edit, or delete.
Implications for the enterprise
It is currently described as the structure of Karpathy "incorrect collection of scripts," the implications for the enterprise are immediate.
As an entrepreneur Vamshi Reddy (@tammireddy) noted in response announcement: "Every business has a raw/catalog. Nobody compiled it. Here is the product."
Karpathy agreed and suggested that this methodology is a "amazing new product" category.
Currently, most companies "to drown" in unstructured data—Slack logs, internal wikis, and PDF reports that no one has time to synthesize.
A "Carpathian style" the enterprise level will not only search for these documents; this would be an active author a "Company Bible" updated in real time.
As an AI teacher and newsletter writer Ole Lehmann put it on X: "I think anyone who packages this for normal people is sitting on something massive. an app that syncs with tools you already use, bookmarks, read-later apps, podcast apps, saved topics."
Eugen Alpeza, AI enterprise agent founder and co-founder and CEO of orchestration startup Edra, X noted in his post that: "The transition from personal research wiki to enterprise operations is brutal. Thousands of employees, millions of notes, conflicting tribal knowledge between teams. Indeed, there is room for a new product, and we build it in-house."
As society investigates "Carpathian example," the focus now shifts from private research to multi-agent orchestration.
final architecture distribution by @jumperzFounder of AI agent creation platform Second mateshows this evolution through a "Swarm Knowledge Base" this extends the wiki workflow to 10 agency systems managed through OpenClaw.
The main problem of multi-agent swarm – where a hallucination can merge and "to infect" collective memory – where it is addressed by a specific person "Quality Door."
Using the Hermes model (trained by Nous Research for structured evaluation) as an independent monitor, each project article is evaluated and approved prior to evaluation. "to live" wiki.
It creates a system "Complex Loop": agents dump the raw outputs, the compiler arranges them, Hermes verifies the truth, and the validated briefs are sent to the agents at the beginning of each session. This herd never provides "wakes up empty" but instead begins each task with a filtered, high-fidelity briefing of everything the collective has learned.
Scaling and performance
A common criticism of non-vector approaches is scalability. However, Karpathy notes that at a scale of ~100 articles and ~400,000 words, LLM’s ability to move between abstracts and index files is more than enough.
For a department wiki or personal research project "ornate RAG" infrastructure is often more delayed and "search noise" more than it solves.
Tech podcaster Lex Friedman (@lexfridman) confirmed using a similar setup by adding a dynamic visualization layer:
"I often create dynamic html (with js) that allows me to sort/filter data and interact with visualizations. Another useful thing is that I have the system create a temporarily focused mini-knowledgebase…then I upload to the LLM for a long 7-10 mile voice mode interaction."
This "an ephemeral wiki" the concept suggests a future where users don’t just stay "have a chat" with AI; they create a group of agents to create a specific research environment for a specific task, and that environment is resolved after the report is written.
Licensing and file-over-app philosophy
Technically, Karpathy’s methodology is built on an open standard (Markdown), but viewed through a proprietary but extensible lens (note-taking and file-organization software). obsidian).
-
Markdown (.md): By choosing Markdown, Karpathy ensures that its knowledge base is not tied to a specific vendor. It is future proof; If Obsidian is gone, the files can be read by any text editor.
-
Obsidian: Although Obsidian is a special application, its "local-first" philosophy and the EULA (which allows free personal use and requires a license for commercial use) coincide with the developer’s desire for data sovereignty.
-
The "Vibe-coded" Tools: The search engines and CLI tools mentioned by Karpathy are custom scripts that bridge between LLM and the local file system, probably Python-based.
This "file on the program" philosophy is a direct challenge for SaaS-heavy models like Notion or Google Docs. In Karpathy’s model, the user owns the data and the AI is simply a highly sophisticated editor. "visits" files to get the job done.
Librarian and search engine
The AI community reacted with technical approval and confusion "vibe-coding" passion Debate has focused on whether the industry is over-indexing in Vector databases for problems of structure rather than mere similarity.
Jason Paul Michaels (@SpaceWelder314), a welder using Claude echoed the idea that simpler tools are often more robust:
"No vector basis. No embedding… Just tagging, FTS5 and grep… Every bugfix… indexed. Combinations of knowledge."
But the most significant praise came from him Steph Ango (@Kepanoco-creator of Obsidian, who emphasized a concept called "Reducing pollution."
It suggested that users keep their personal information private "warehouse" clear and let the agents play together "messy warehouse," only fetching useful artifacts after the agent-facing workflow has distilled them.
Which solution is right for your enterprise vibe coding projects?
|
Feature |
Vector DB / RAG |
Karpathy’s Markdown Wiki |
|
Data Format |
Opaque Vectors (Math) |
Human-readable Markdown |
|
Logic |
Semantic similarity (nearest neighbor) |
Backlinks/Indices |
|
Audit capability |
Bottom (Black Box) |
High (Live Tracking) |
|
Componentization |
Static (requires re-indexing) |
Active (self-healing through linting) |
|
Ideal Scale |
Millions of Documents |
100 – 10,000 High Signal Documents |
The "Vector DB" the approach is like a massive, disorganized warehouse with a very fast forklift driver. You can find anything, but you don’t know why it’s there or how it relates to the pallet next to it. Carpathian "Markdown Wiki" it’s like a curated library with a head librarian who is constantly writing new books to explain the old ones.
The next stage
Karpathy’s latest exploration points to the ultimate destination of this data: Synthetic Data Creation and Fine-Tuning.
As the wiki grows, so does the data "pure" through continuous LLM linting becomes a perfect training set.
Instead of LLM just reading the wiki "context window," the user can finally fine-tune the smaller, more efficient model within the wiki itself. This will enable the LLM "to know" the researcher’s personal knowledge base in its own weight, essentially turning a personal research project into a private, private exploration.
Conclusion: Karpathy didn’t just share a script; share a philosophy. By treating LLM as an active agent that maintains its own memory, it circumvents the limitations "one shot" AI interaction.
For the individual researcher, this means the end of the job "forgotten bookmark."
For an enterprise, this means a transition from a "raw material/data lake" a "compiled knowledge asset." As Karpati himself sums up: "You rarely write or edit a wiki by hand; LLM is the domain." We are entering the era of autonomous archives.
