Ask not whether a system predicts, but who suffers when it predicts badly. If the answer is somebody else, you are looking at a tool. If the answer is the system itself, and suffering enough means ceasing to exist, you are looking at something with a stake.
Navigation: ‹ An LLM Has a Loop But Not a Life · Movement II — The Boundary ›
Movement I began with a simple identity: to predict well is to minimise prediction error. Then it made the uncomfortable admission that language models really do belong inside that identity. They are not fake predictors. They are not merely pretending to minimise surprise. They are trained by being surprised less.
So the easy escape is closed.
If we want to distinguish a language model from a creature, we cannot do it by saying that one predicts and the other does not. We cannot do it by saying that one has a loop and the other does not. We cannot do it by pointing at sophistication, fluency, or scale, because scale moves. Models get larger. Scaffolds get better. Interfaces become more convincing.
The right question is not “does it predict?” or “does it loop?”
The right question is: who owns the loss function?
Who chose what counts as error? Who pays when the prediction fails? Who has to reorganise, act, repair, flee, learn, or die because the map did not fit the world?
That is the hinge of the movement.
5. Thinking costs crumbs
I should either explore the hole or go look for food somewhere else, it thought.
Either. Or. Not both. Whichever one the beetle chose, it was paying for with the other one.
And standing there deciding was not free either. Wondering costs crumbs — the very same crumbs that were keeping the beetle alive — so the longer the beetle wondered, the less beetle there was to wonder with.
Which is why a beetle can never wait until it is completely sure. A beetle who waits to be completely sure is a beetle who starves very carefully.
A language model has a loss function because somebody gave it one.
This is not a criticism. It is how tools are made. A training system needs a target. Engineers choose data, define the scoring rule, run the optimisation process, stop it when it has gone far enough, and then place the resulting model inside some larger human purpose. Even when that process becomes complicated — preference tuning, safety layers, tool use, memory, scaffolding, agents — the basic orientation remains: the model's problem is set by something outside the model.