What If AI Organised Everything You Know?

The most useful thing AI could do for most knowledge workers this year has nothing to do with writing emails or cleaning up slides. It is organising the information you have already collected — the research notes, meeting transcripts, half-read PDFs, Slack exports, and voice memos sitting in folders you never open again.

That sounds mundane. It is not. A research analyst who can ask "what did we learn about the German supplier last quarter?" and get a direct answer — not a search result, not a summary, but a synthesis across eight different documents — has compounded every hour she has ever spent reading anything. Most of us have not.

Last week, Andrej Karpathy published a system that makes this kind of automatic organisation possible for ordinary people without writing any code.

What Karpathy built

The idea is deceptively simple. You drop raw materials — papers, notes, articles, transcripts — into a folder. An LLM reads everything, identifies concepts, writes one wiki-style article per concept, and links them to each other. When you add new material, the wiki updates itself.

VentureBeat reported that Karpathy's single research topic grew to roughly 100 articles and 400,000 words — none of which he wrote directly. His framing is the part worth stealing: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."

In plain English: the place you store knowledge is a folder of text files, the thing writing into it is an AI, and what you end up with is a living document of everything you know that someone else maintains for you.

Why this matters more than another chatbot trick

Almost everything AI has been sold to professionals so far is one-shot. Draft this email. Summarise that report. Clean up these bullets. Useful — but nothing compounds. You drafted an email on Tuesday and by Friday the work has evaporated.

A self-building knowledge base compounds. Every meeting you record, every article you read, every briefing you receive becomes part of a growing structure you can ask questions of later. For a management consultant moving between client engagements, that is the difference between walking into Monday's call with half-remembered notes and walking in with a tight synthesis of everything your firm has said about the client's sector in the last three years.

What you can actually do this week

You do not need Karpathy's full setup to get most of the benefit. The underlying idea generalises three ways any professional can use now.

One — structure one existing pile. Pick a single source of accumulated mess. Your meeting notes folder. Your "save for later" bookmarks. The research doc for a project you are on now. Give AI the whole thing and ask it to produce a structured outline with a one-paragraph summary per topic and a list of the key open questions. A legal associate who does this with her case-file notes once a month will pull ahead of colleagues still scrolling PDFs by hand.

Two — define the shape of the knowledge before you feed it in. Karpathy's wiki works because the AI knows what "an article" is supposed to look like. You can do the same thing in a regular chat by telling it what format the output must take. A product manager researching competitor features might specify: each competitor gets a card with positioning, pricing, three example customers, and two weaknesses. Consistent structure is what makes knowledge reusable three months later, when you actually need it.

Three — keep it living. A knowledge base that updates itself is worth ten times a snapshot. Each week, take the new material you collected — articles, call transcripts, one recorded customer interview — and feed it into the same structure with the same instructions. A recruiter maintaining a candidate-market map this way will know things about their niche that nobody else at their firm knows.

The trap to avoid

The obvious failure mode is letting the wiki become a dumping ground. Karpathy's system works because it is scoped — one topic, deeply. A consultant who tries to build a wiki of "everything I have learned about business" will get 400,000 words of beige. A consultant who builds a wiki on "the European battery supply chain, 2024 to 2026" will build something rare.

Narrow beats broad. Specific beats comprehensive. The skill here is not feeding AI more material. It is choosing the boundary of the knowledge base so what grows inside is worth something.

Monday morning

Before your first meeting next week, pick one topic you care about professionally and one folder of existing material about it. Spend thirty minutes asking AI to organise that material into a structured outline with cross-references and an index of the open questions. You will not have Karpathy's full system. You will have something more valuable: the first version of a second brain that compounds, week after week, while you do everything else.