What intelligence is. One mechanism — minimising prediction error — runs from a beetle in a tunnel to a language model. What separates them turns out not to be sophistication, but ownership.
Navigation: ‹ Overture — Let Our Conjectures Die In Our Stead · The Beetle and the Tunnel ›
What kind of thing is intelligence?
The answer begins smaller than a person. A beetle at the mouth of a tunnel is already doing the essential work: guessing what might be true, checking the world, and updating its map when the world answers. Intelligence begins there, not with language, self-reflection, or human cleverness.
But that same structure also appears in a language model. A model predicts the next token, is scored against what actually came next, and improves by reducing surprise. That is not a metaphor for prediction-error minimisation. It is a real instance of it.
So the easy distinction fails. We cannot say organisms predict and models merely calculate. We also cannot say a model is a creature simply because it predicts. The question has to move one level deeper.
Who owns the error?
Prediction is error minimisation; next-token prediction is genuinely prediction; but a creature differs from a model because the creature's errors matter to the boundary that keeps it alive.
That is the hinge of Movement I.
A language model is scored. A beetle is at stake. If the model's prediction is wrong, the training process registers an error. If the beetle's prediction is wrong badly enough, the beetle may not come back out of the tunnel.
The movement does not use that difference to declare that machines can never matter. It uses the difference to state a criterion:
No self-maintaining boundary, no perspective.
That is a narrower, more useful position than either familiar shortcut: it’s just matrix multiplication, so never, or it talks, so probably.