Andrej Karpathy’s LLM-wiki idea (gist) caused a stir in my circle. I’ve been doing something similar for a few years now. It used to barely work; the quality has grown to the point where it would be strange not to use it.
For me an LLM-wiki isn’t about accumulating knowledge. It’s about how I read.
Technical articles aren’t poetry — we don’t read them for pleasure, we read them for what we can take away. It’s enough to answer one question: what’s new here for me? I feed the article to an agent and it lays it out across the wiki: new notes, updates to existing ones, links between them. I read the diffs and ask follow-up questions. Obsidian plus the agent keeps a live snapshot of what I understand, growing with every source.
The other use case is projects. Each project or activity — a work task, a hobby — gets its own space with its own schema. New information shows up; the agent is the entry point and figures out where it belongs. It makes mistakes sometimes, but those are an excuse to talk through why it made that call and ask it to update the schema.
I used to let knowledge pile up in messages, files, slide decks — a growing data backlog where I was the bottleneck: if I didn’t write it down, part of it was lost, and the loss accumulated like tech debt. Now it’s different: each project is managed by the agent, and I manage it through the agent. I’m not the one doing the structuring — I steer the flow, read what comes out, push back when something feels off. Many people do this through cowork tools; I’m more at home in Obsidian plus an agent.
The agent gets things wrong, sometimes extracts the wrong slice — but the quality is already high enough that the value is obvious. If you only have one project and you’re good at staying focused, this is overengineering. If, like me, you suffer from context-switching, this keeps you from drowning in the stream.