Andrej Karpathy went viral with his “LLM wiki” idea — a way to make AI compound knowledge over time rather than start from scratch. I recognized it immediately. Not because I follow AI research, but because I’ve been teaching the same three-part structure in education for two decades.
The Idea
Karpathy’s insight is deceptively simple: instead of asking AI to search your documents every time you have a question (RAG), you ask AI to read your documents and build a structured, interlinked knowledge base from them. The knowledge base grows as you add new material. It compounds. The AI becomes not a search engine but a librarian — maintaining a living wiki on your behalf.
The internet responded the way it usually does: dozens of tutorials, YouTube explainers, GitHub repos.
I responded differently. I read the description and thought: I know this structure. I’ve been drawing it on whiteboards for twenty years.
The Recognition
Here is Karpathy’s three-part pattern, stripped to its bones:
- A specification — rules that tell the system what to do and what not to do
- An editable surface — the knowledge base the AI maintains
- An evaluator — a rubric for checking whether the system is actually improving
If you have spent any time in instructional design, curriculum development, or educational research, you just felt something click. That’s because this is the same structure we use to design learning systems:
- Learning objectives — the specification
- Student work — the editable surface
- Assessment rubric — the evaluator
Karpathy formalized for AI research systems what educators have known for decades: a system without an evaluator is just a process with no way to improve. The spec tells the agent what to do. The surface is where the work happens. The rubric is how you know whether it’s working.
This is not a coincidence. It’s a pattern that shows up anywhere you need a system to learn and improve over time — whether you’re teaching a student, training a model, or governing an AI agent inside your knowledge garden.
Why I Needed to Rebuild Anyway
I’ve been writing about my digital garden and knowledge system for a couple of years now. The original setup was built around a maturity model: Seeds, Plants, Evergreens, Groves, Forests. A note started as a rough idea and graduated toward something polished and reusable as it developed. It was a good model for one person moving ideas forward intentionally.
But two things broke it.
First, the volume. Once I started using AI tools to clip, summarize, and capture material at scale, my inbox became a fire hose. The maturity model assumed I had time to tend each note. I didn’t. The garden got weeds.
Second, the blending. I care a lot about keeping my voice. The research I’ve done on AI use in education — particularly on agency and friction — showed me clearly that low-agency AI use is defined by one thing: accepting the default. Taking the first draft. Letting the tool structure the thinking. I had started to do that in my own notes without noticing.
I needed a system that could handle volume and protect the human layer.
Consume, Curate, Create
The rebuild organized around a simple idea: different layers of the knowledge system serve different purposes, and AI should be present in some of them and absent in others.
01 Consume is the intake layer. Everything comes in here — web clippings, podcast notes, paper summaries, videos, AI chat logs, quotes. This is a source layer. Material arrives with minimal processing. I add to it constantly, and the rule is: sources stay sources. They don’t get rewritten, merged, or synthesized here. They are preserved with their provenance intact.
02 Curate is the wiki layer. This is where AI does its primary work. A set of agents — I call them the Inbox Sorter, the Consume Custodian, and the Librarian — routes incoming material, maintains metadata, and updates a structured knowledge wiki based on what comes in through Consume. I am largely hands-off here. The agents work from explicit rules I wrote. The wiki grows automatically as new material arrives.
03 Create is where I show up again. Blog posts, newsletter issues, evergreen essays, courses — all of it is produced in this layer. AI may assist with drafting or editing, but the initiating voice, the argument, the choices about what matters — those are mine. I pick up the pieces that the curate layer has organized and synthesize them into something worth publishing.
The Firewall
Here is the thing I keep coming back to when I explain this to people: the layers are not just organizational. They represent a firewall between AI-maintained content and human-authored content.
In Curate, the agents work. They route, classify, synthesize, and link. But they operate inside hard rules. The Librarian cannot touch the Create layer. The Inbox Sorter cannot invent content — it can only route and flag. Every run is evaluated against a rubric before the outputs are accepted.
In Create, the human works. The blog post you are reading right now is not AI-drafted and Ian-approved. It is Ian-initiated, Ian-structured, Ian-written. AI may have helped me pull a quote from a note or check a connection — the same way I would use a search tool. But the thinking, the argument, the voice — that’s the work that stays mine.
This matters because it means the system has explicit design friction built into it. In my research on student AI use, I found that high-agency users were defined by their willingness to refuse the AI’s first output, to push back, to re-prompt with constraints. I am not exempt from that dynamic in my own work. The firewall is how I engineer refusal into the architecture rather than relying on my willpower in the moment.
The Sovereignty Argument
There is one more layer to this that goes beyond productivity.
The entire system runs locally. Notes are Markdown files. The vault is backed by Git. No SaaS subscription controls my archive. No platform change can lock me out of fifteen years of writing and thinking.
I have written before about platform capture — the way knowledge technically owned by you becomes functionally trapped inside systems you don’t control. The Consume/Curate/Create architecture is a live argument against platform capture. Karpathy’s original design made the same choice: files as the universal interface. Local-first. Portable.
For me, that is not just a technical preference. It is a values statement that I try to model for the people I teach and write for. You cannot teach digital sovereignty if your own knowledge system is tenant farming on someone else’s infrastructure.
What This Has to Do With You
I do this publicly because I believe showing the process is part of the work. My positioning — helping people become AI-literate through practical skills and critical thinking — is hollow if I only describe the ideas and never demonstrate them. The rebuild of this knowledge system is the demonstration.
Part 2 of this series gets into the mechanics: how material actually flows through the system, what each agent does, how the evaluation rubrics work, and what a typical day looks like for a knowledge system that runs itself during the week and waits for me to pick up on the weekend.
But the conceptual core is here: the pattern Karpathy made viral is the same pattern that underlies good instructional design. Specification. Editable surface. Evaluator. Build those three things right, and the system can learn. Build them wrong — or skip the evaluator — and you get noise that compounds instead of knowledge.
I was not surprised by Karpathy’s idea. I was relieved that someone with his reach had finally said it in terms the AI world could hear.