The deal was always the same: capture everything, tag it well, link it obsessively, and eventually your notes would become a second brain. A personal knowledge base that would surface the right idea at the right moment. Roam, Notion, Obsidian, Logseq — the tool changed, but the promise didn't.
And it never quite delivered. Not because the tools were bad, but because the model was wrong. You weren't building a second brain. You were hiring yourself as a full-time librarian for your own thoughts.
Every note needed a home. Every idea needed tags. Every concept needed backlinks to the three other concepts it touched. The system grew, and so did the maintenance burden. The more disciplined you were, the more time you spent on upkeep instead of thinking. The notes became the work.
"The problem was never Obsidian. The problem was that you became the maintenance worker for your own knowledge."
Andrej Karpathy's LLM wiki idea cuts through this directly. The insight isn't just "use AI to search your notes." It's more fundamental than that: what if the knowledge system could maintain itself? What if you could just talk to your accumulated thinking — in plain language, without tags, without a folder structure, without a weekly review ritual that you always skip?
For me, the ops background reframes this clearly. In infrastructure, we learned early that any system requiring constant manual intervention is a system waiting to fail. Toil is the enemy. If a process can't run without a human babysitting it, it's not a system — it's a job. PKM tools, however beautifully designed, were toil by another name.
An LLM-backed knowledge base inverts the model. You write when you have something to say. The system handles retrieval, synthesis, and connection. You stop being the librarian and go back to being the person who reads the books.
That's the shift worth paying attention to — not the specific implementation, but the ownership of maintenance moving from human to machine. It's the same reason CI/CD replaced manual deployments. Not because manual deployments were impossible. Because they were the wrong thing to spend human attention on.