On April 2, Andrej Karpathy posted a few paragraphs on X that shifted how a chunk of the tech industry thinks about AI. He didn’t announce a paper or a product. He described how he actually works now.

The post hit 17 million views. The GitHub gist that followed collected 5,000+ stars and 1,400+ forks in 48 hours.

The key line: “A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge.”

Most of the coverage zeroed in on mechanics: Obsidian as frontend, markdown files, folder structure. While the mechanics are interesting, the bigger takeaway is how Karpathy changed the AI’s role: from answering questions on demand to compiling and maintaining a living knowledge system.

What he actually built

Three folders. That’s the entire architecture.

raw/ holds source documents. Articles, papers, repos, datasets, images. The AI reads these but never modifies them.

wiki/ holds the compiled output. Markdown files with summaries, cross-references, backlinks between concepts, entity pages, an index, and a changelog. The AI owns this directory entirely. Karpathy rarely touches it directly.

A schema file defines the rules: how to organize content, what conventions to follow, what to do when ingesting a new source.

His result on a single research topic: roughly 100 articles and 400,000 words. He didn’t write any of it. The LLM read the sources, compiled the wiki, linked everything together, maintained the index, and flagged contradictions between sources.

He skipped vector databases, embeddings, and RAG pipelines entirely. Just folders and text files.

His summary: “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

This echoes Vannevar Bush’s 1945 Memex: a personal device that would store documents with “associative trails” linking related ideas. Bush’s concept was right. The part nobody could solve for eighty years was who does the maintenance. The LLM does.

The reactions that matter for business

The post went viral with developers, but the responses from industry leaders tell a more useful story.

Lex Fridman confirmed he runs a similar system. He generates interactive HTML visualizations with JavaScript and creates temporary “mini knowledge bases” that he loads into voice mode during 7-10 mile walks.

Steph Ango, CEO of Obsidian, gave the most practical advice in the entire thread: vault separation. Keep your human-curated knowledge clean. Let agents compile in a separate space. Only bring over useful artifacts once the AI-generated content has been distilled and validated.

Elvis Saravia from DAIR.AI published a full analysis and confirmed he uses a similar approach to curate AI research. His verdict: “The most personalized research system I’ve ever built, and it’s not even close.”

But the reaction that should matter most to anyone running a B2B company came from Vamshi Reddy: “Every business has a raw/ directory. Nobody’s ever compiled it.” Karpathy agreed.

Within a week: seven+ GitHub implementations, Farza building “Farzapedia” (a personal Wikipedia from 2,500 life entries compiled into 400 articles), and Edra — founded by former Palantir executives — announcing a $30M Series A led by Sequoia to solve the enterprise version.

The knowledge rot problem

I’ve been in computers and business for 35 years. I’ve watched a lot of trends come and go. But what Karpathy describes points at a problem B2B companies have carried for decades.

McKinsey estimates employees spend 1.8 hours per day searching for internal information. That’s 25% of the workday. IDC puts it at 2.5 hours. While estimates vary, the underlying productivity drain is obvious.

That knowledge currently sits scattered across Confluence, SharePoint, Slack threads, email chains, Google Docs, and a huge volume of tribal knowledge trapped in the heads of three people who’ve been at the company for a decade and nobody has sat down to interview.

Now look at how most companies “solve” this with AI: upload documents to ChatGPT, ask a question, get an answer, close the session. Next time someone has the same question, the process starts from zero. Nothing accumulates between sessions. Every query starts from scratch.

Karpathy’s approach flips this entirely. Because new sources and user questions are constantly filed back into the wiki, the system actually grows richer over time rather than resetting.

That’s the gap between a chatbot and knowledge infrastructure.

Enterprise limitations to consider

Scaling this exact architecture to a company presents concrete technical obstacles.

Context windows limit how many documents the LLM can process without losing information. Karpathy works with ~100 articles. A mid-size company has thousands; a large one has millions. At that scale, the model navigates self-maintained indexes and can easily miss relevant information buried deep in the wiki.

Standard markdown links also lack the associative power of vector-based retrieval. Scott Rippey, an engineer whose comment got 67 reactions on a LinkedIn thread, put it directly: “This IS hand-built Graph RAG, not a replacement for it.” Embeddings discover connections across differently-worded text that manual linking simply cannot.

Without role-based access control, audit trails for regulatory compliance, or data exfiltration protections, this stays a personal tool. A chmod isn’t a security strategy. Git history isn’t an enterprise audit log. For any regulated industry, these aren’t nice-to-haves.

But the risk that concerns me most is error compounding. The LLM writes a wiki page with a subtle mistake. Someone queries against that page. The mistake enters the answer. The answer gets filed back. Now two pages reinforce the same error. Without external verification, errors accumulate just like correct knowledge does.

Nilou Salehi from Across AI framed it well: “The real unlock isn’t how do we store knowledge. It’s how do we make knowledge executable.”

What B2B companies should actually do

Here’s what you can do about it this week.

Before building anything, map where your company’s scattered knowledge currently lives. Where do product decisions get recorded? Where’s the process documentation? Which critical knowledge exists only in specific people’s heads?

Once you have that map, isolate a single domain to compile. Competitive intelligence, technical documentation for a specific system, or client onboarding all work well. Trying everything at once is a recipe for never finishing. Build the knowledge base for one vertical, validate the pattern, then expand.

As Steph Ango suggested, keep AI compilation spaces separate from your official knowledge base until a human has validated the artifacts. Mixing the two contaminates your source of truth with hallucinations that look like facts.

There’s a metric nobody tracks that tells you whether this approach is worth your time: documentation decay rate. How fast does your documentation go stale? If the answer is “weeks” and nobody updates it because the work always gets deprioritized, you’ve found the exact use case where an LLM absorbs the maintenance burden.

One last thing: Karpathy himself called his system “a hacky collection of scripts.” The pattern is sound. The enterprise implementation needs governance, access control, and a context layer that models don’t ship with by default.

Knowledge isn’t a chat

Karpathy’s method completely changes how organizations interact with their own knowledge.

For years, companies tried to fix this with more advanced search tools, complex tagging systems, and strict folder hierarchies. None of it solved the problem because the bottleneck was never search. It was maintenance. Nobody wants to be the person updating the Confluence wiki at 4 PM on a Friday.

LLMs change the maintenance equation. For the first time, you have an agent that doesn’t get bored updating cross-references, that can touch 15 documents in a single pass without complaining, and that improves the system with every interaction instead of degrading it with every month of neglect.

Without a maintenance plan, corporate knowledge inevitably goes stale.

At IQ Source, we help B2B companies build knowledge systems that get more valuable with every interaction, with the governance layers that a viral X post doesn’t include but your operations require.

If you want to map where your organization’s raw/ directory lives, let’s start there.