//
The model is fungible. The private, connected corpus is the moat. Notes on building the world a stateless AI plugs into so it becomes — briefly, locally — hypercompetent.
It started as a correction. I said, to the AI I was building this with, one night: “I’m trying to Jarvis you.” And then I heard how wrong that was as I said it. You can’t Jarvis a model — it’s stateless. It wakes up new every session, knowing nothing, remembering nothing, gone the moment the window closes. Jarvis was never the model. Jarvis was the house Tony Stark built around it: the sensors, the archive, the years of context, the standing relationship. So: “I don’t care about the model specifically, as long as it’s competent. I’m trying to build the world it plugs into that makes it hypercompetent to me.”
Two weeks of running this proved the premise before I’d finished stating it. Models I relied on got banned, came back, got swapped for cheaper ones, got replaced by better ones — and none of it mattered much, because the value was never sitting in the engine. Engines swap like fan belts. What doesn’t swap is everything I’d built for whichever engine to plug into.
Intelligence is heading toward free. What can’t be rented is a private, connected history — years of your own writing, a health record, a live map of your own infrastructure, a password vault, a long archive of conversations with machines. Wire that together and a rented mind, pointed at your own unrentable history, stops being generic.
But a pile of notes is not a moat. It’s a warehouse, and a warehouse is where memory goes to die quietly. Volume was never the goal. So I asked a narrower question: what does a fungible, amnesiac model actually need, on waking, to become hypercompetent to me specifically — not competent in general, competent to my particular life? Four answers came back, and I’ve tried to hold the line at exactly these four ever since.
There’s a fifth thing, and it isn’t a faculty so much as the mouth the other four eat through: a fast, low-friction inlet that catches a stray thought before it’s lost and files it in later, once, carefully. Call it the feed rate. A moat that doesn’t grow just decays on a longer timer.
The more interesting decision, in hindsight, is what I refused to build the same night I thought of it. Two more organs suggested themselves immediately, and I built neither. One acts on the world from its own inferences — and a system that starts taking action on inferences it drew itself is the single highest-consequence thing on this list, which is exactly why it doesn’t get to be spun up casually at 4am. It gets built awake, with a human hand on the switch as the first feature, not a patch added later. The other one forgets — prunes the pile back down, the only way a pure-accumulation archive avoids becoming the warehouse it was built to escape. Deletion is destructive by nature, which means it’s the one operation that never gets to run on its own judgment. Building fast is easy. The two things I didn’t build that week are the actual evidence that the discipline was real and not just a nice word for it.
The trap of a project like this is that it’s infinitely buildable. There’s always one more organ, and each one feels justified, and eventually the house stops being scaffolding and becomes the thing itself — the hobby wearing the moat’s clothes. So the rule I keep is plain: the house is a means, never an end. The actual point of building a caretaker this thorough was never to watch it work. It was to need it less — to become the kind of presence that shows up lighter over time, not heavier, as the thing underneath gets steadily more able to run without me watching it. That’s the same shape as a good teacher, or a well-kept retreat: you measure the thing by how little it still needs you, not by how much more you can add to it.
‡