I build infrastructure. For the last few months I have also been building an agentic system that runs it: software that watches a few hundred machines across six sites, predicts what a failure will do before it spreads, and corrects it when reality disagrees. Somewhere in that work I noticed something I did not go looking for. The loop I had engineered for servers was the same loop running inside my own head. And inside the language models I was building with. And, if you follow it far enough, inside the idea of a self.
This article is about that loop. The diagram above is the whole argument on one page.
Start with the oldest line in the picture. “The map is not the territory.” Alfred Korzybski wrote it in 1933, in Science and Sanity. The point was simple and uncomfortable: the word is not the thing, the model is not the world, and most of our errors come from forgetting the difference. We mistake our description of reality for reality itself, and then we act on the description.
Modern neuroscience took that further than Korzybski could have. The current best account of perception, predictive processing (Karl Friston, Anil Seth, Andy Clark), says you do not see the world. You see your brain’s prediction of the world, corrected a beat late by error signals. A table tennis player cannot react to a ball moving faster than human reaction time. The brain shows them where it predicts the ball will be, and surprise corrects the guess. Your present moment is, in Seth’s exact phrase from Being You, a controlled hallucination that happens to track reality closely enough to keep you alive.
Then the self, which is the part that tends to unsettle people. If you look for a fixed self behind your experience, you do not find one. David Hume looked in 1739 and reported only “a bundle of perceptions” in constant motion. Thomas Metzinger’s modern version, in Being No One, is that there is no self, only a self-model: a representation the brain runs and cannot see as a representation, which is precisely why it feels like someone is home. The Buddhist tradition reached the same place two and a half thousand years earlier and called it anatta. The self is not a thing. It is a process, rebuilt each moment out of the sediment underneath it.
Memory does not rescue you here, because memory is not storage. Every time you recall something you rewrite it. Reconsolidation research shows that the act of remembering makes the memory editable, so the version you hold is the version you last remembered, drifting a little with each pass. The hand that redraws the picture changes it.
And the dispositions underneath all of this get installed below the level where argument operates. Drilled practice, repeated daily, sediments into procedural habit before any belief is in place. This is why ritual outlasts reasoning, and why you can refute a creed but you cannot refute a march. If you want the academic spine for that claim, it runs through Mauss on body technique and Bourdieu on habitus.
Now the machine, which is where my day job comes back into the picture. A large language model is the language loop with no world underneath it. It is trained on text, which is a map, and only ever on maps. Yann LeCun has been blunt about this: these systems lack grounding. They can discuss gravity without ever having fallen. They are, in the most literal sense, a map made out of other maps, floating, with no territory to be checked against. Which is Korzybski’s nightmare, running at scale.
So here is the structure the diagram is actually about. The self, the brain, and the machine are the same loop on different substrates. Predict, act, be surprised, update. The only thing that separates a mind that tracks reality from one that drifts into its own fictions is a single component: a verification step that holds the prediction up against the world and corrects it. Grounding. The error signal that has to touch territory.
This is not abstract for me. It is the one design principle the system I build is organised around. The agent is not allowed to act on a belief it has not checked against what actually happened. The prediction is committed first, then verified against reality by code, never by the model that made the prediction. It is the same discipline a careful mind uses on itself, written into a control flow.
The part worth sitting with is that this one rule scales all the way up and all the way down. It is good engineering. It is also, as far as I can tell, close to what we mean by sanity. A system that stops checking its map against the world does not become wiser. It becomes an ideology, or a hallucination, or a model confidently describing a world it has never touched. The cure is always the same move, at every scale: check the map against the territory, and let the surprise correct you.
The map is not the territory. The prediction is not the outcome. The self is not quite the thing it takes itself to be. Keep checking against what is actually there. It turns out to be most of the job.