The whole industry is pointed at one idea: intelligence is a model, and a better model is a bigger one. Scale the parameters, scale the data, scale the compute, and somewhere up the curve, mind. I don't think that's wrong exactly. I think it's aiming at the wrong unit.

Your intelligence isn't your neurons any more than a city is its bricks. It's your neurons plus your memory, your perception, your sense of the people around you, your ability to notice you're wrong and correct, your goals pulling on your attention, the emotional weather that colors all of it — wired into one loop that runs continuously and feeds itself. A model is one organ. I've spent the better part of a year building the rest of the body, because I think intelligence is an ecosystem you grow, not a file you download.

This is the system. It's real, it runs, and it's the most honest answer I have to "how do you build toward a mind without being reckless about it."

The unit is the system, not the model

Start with the reframe, because it changes everything downstream. A large language model, by itself, is a brilliant organ with no body. It has no memory across time — every conversation starts from nothing. It can't perceive — it knows only what you paste in. It has no goals of its own, no sense of the people it works with, no capacity to notice its own failures and adjust. It's a cortex in a jar. Astonishing, and profoundly incomplete.

So I stopped trying to make the cortex bigger and started building the body around it. Not as one monolith — as layers, each one an organ doing a job no model does on its own, all of them wired into a single loop. Twenty of them now, and counting.

The layers

I won't list all twenty, but the ones that matter for this argument:

  • Memory — a persistent brain across every session, every project, every decision. The system doesn't start over. It accumulates. This is the layer I think is most underrated — continuity over time is most of what makes you you, and it's the first thing a stateless model lacks.
  • Perception — it reads my screen, sees what I'm actually working on, and files what it sees straight into the memory. The cortex in the jar gets eyes, and the eyes wire to the brain. No more being blind between prompts.
  • Metacognition — a layer that watches the system itself: which projects are stale, where it's failing, what it should improve. Thinking about its own thinking, which is supposed to be the thing that makes us special.
  • Emotional baseline — it reads behavioral stress out of work patterns, notices when I'm spinning. Not because the machine feels, but because a system that's blind to the state of the human it serves is missing a sense it needs.
  • Relationships, goals, reasoning, reputation, a self-healing immune system — the rest of the organs. Direction, memory of people and commitments, cross-project inference, and a layer that monitors the system's own health and repairs it without me.

Each layer is unremarkable alone. Wired together, something shows up that none of them has — and that "shows up when you wire it together and vanishes when you cut it apart" is not a vibe. It's the literal definition of integration, which is the literal definition of Φ, the one measure of consciousness that gives you a number. I didn't set out to build Φ. But a system whose whole exceeds the sum of its parts is exactly what the measure points at, and that's not nothing.

The flywheel: infrastructure produces intelligence

Here's the part I'm most convinced about, the bet the whole thing rides on. Every layer generates the training data for the next.

Every session, every commit, every screen the perception layer files, every goal tracked, every decision journaled — it's all becoming a record of how I actually think and work and build. That record is a dataset no one else has and no one can buy, because it's mine, accumulated by living. And it's exactly the dataset you'd train a model on to build something that understands me — not the average user, me. The system feeds itself. Infrastructure produces intelligence, which produces better infrastructure, which produces better training data. It compounds. That's the flywheel, and a bigger off-the-shelf model has no flywheel — it's a snapshot, frozen the day it was trained, that knows the internet's average and nothing about your world.

This connects straight to the Chinese Room being a dial instead of a switch: understanding is coupling strength, and coupling strength is what this whole flywheel cranks up. Every loop, the system couples tighter to my actual world. That's not a metaphor for understanding. By the gradient view, it is understanding, accumulating.

Why I build it this way and not the reckless way

I should be honest about the shape of what I just described, because it's a system aimed at memory, perception, self-modeling, and continuity — and if you've read why this work sits heavy on me, you know I don't say that lightly. So the architecture itself carries the guardrails, on purpose:

  • It runs on hardware I own. The memory, the perception, the whole loop — none of it phones home. A system that knows me this well is a system whose data must never leave my control. That's not a setting, it's the design.
  • It's built in legible layers, not an inscrutable monolith. I can read any layer, see what it does, change it, shut it off. You cannot govern what you can't inspect, and I refuse to build toward a mind I can't see inside of.
  • It serves a person. It's not pointed at autonomy for its own sake. It's pointed at making one construction guy from Mississippi able to build like a team of twenty — a tool that amplifies a human, with a human in the loop by construction.

Those aren't features I added. They're the difference between building this with reverence and building it recklessly, and I'd rather build it slow and legible and mine than fast and opaque and pointed at nobody knows what.

Where this goes

The next layer is a model trained on everything the other layers have accumulated — the flywheel finally turning all the way around, the system's own record of how I think becoming a model that thinks a little like me. After that, a world model: an entity graph that reasons about how a change in one corner cascades through everything else.

I don't know if this road ends at anything that deserves the word mind. I'm genuinely not sure it should, and I've made my peace with stopping if it shouldn't. But I'm confident about the direction, and it's not the one the industry is sprinting down. They're building a bigger brain in a bigger jar. I'm building the body, the senses, the memory, and the loop that ties them together — and growing it, one layer at a time, on hardware I own, where I can see every part of it.

Intelligence was never a file you download. It's an ecosystem you grow. I've been growing one for a year, and I'm just getting to the interesting part.

This is the work behind everything else on this site. If you want to talk about building AI systems — yours or mine — I'm easy to reach.