To predict and to minimize prediction error are not two activities. They are one activity described from two directions. Everything else in this arc is built on that identity, so it’s worth saying slowly.
Navigation: ‹ The Beetle and the Tunnel · Why Next-Token Prediction Is Prediction-Error Minimization (And Where It Isn't) ›
A prediction is not just a sentence about the future.
“I think it will rain tomorrow” is a prediction because tomorrow can answer back. If the sky stays clear, the prediction has failed. If rain falls, it has survived one test. The important thing is not that the sentence pointed forward in time. The important thing is that the world was allowed to disagree with it.
That is what turns a guess into a prediction: it can be scored.
Once a prediction can be scored, getting better at predicting just means making the score come out less badly. You expect one thing, something happens, and the gap between the expectation and the event is the error. To predict well is to make that gap smaller. To minimise prediction error is to predict well. These are not two different achievements. They are the same achievement seen from opposite sides.
From the inside, it feels like expectation: the mind, body, model, or creature leans toward what it thinks is about to happen. From the outside, it looks like correction: the system changes so that the next lean is less wrong.
This is the identity the arc needs. Prediction is error minimization in motion.
2. The beetle who was a map
Once there was a beetle who carried a map of its world. Not a paper map — beetles have no pockets. This map lived inside the beetle (in fact, it was the beetle’s whole body, its bodymap), and it contained a record of everything that had ever happened to the beetle.
On the map, under bark meant sometimes grubs. Under stones meant cool moss, and once, a very rude centipede. And tunnels — the beetle had known exactly two tunnels in its whole life. The first had led to a rotten log absolutely full of dinner. The second had gotten narrower, and narrower, and narrower still, and then stopped — and the long, slow, backward crawl out was the worst afternoon the beetle could remember. So on the beetle’s map, tunnels were marked: sometimes wonderful. Sometimes terrible.
One morning, the beetle found a hole in a bank of earth, and the hole opened into a long dark tunnel — a tunnel exactly one beetle wide. Too narrow to turn around in. Too tight to crawl backward out of, except very slowly, and the beetle (or, you could say, the beetle’s bodymap) remembered exactly what that felt like.
The beetle peered in. It could not see the end. So it began to make guesses instead, based on its world map. Maybe food down there. Maybe a great big cave to explore. Maybe it gets narrow, and narrower, and stops. The beetle didn’t know which guess was true. Nobody who stands at the mouth of a tunnel ever does.
The beetle’s map is not a picture hidden inside the beetle. It is the beetle’s body, nervous system, history, sensitivities, habits, scars, appetites, and readinesses arranged so that the world can mean something to it.
That is why the same tunnel is not the same tunnel to every creature. To a stone, it is only a hole beside another stone. To a paper map, it is a line or a shaded mark. To this beetle, it is food-or-danger, shelter-or-trap, maybe-wonderful, maybe-terrible.
A world model is not a copy of the world. It is a compressed history of how the world has mattered so far.