Uncertainty reduction

Origin

Concept from information theory (Shannon) and neuroscience (Friston).

Definition

The objective shared by every intelligence, biological or artificial. The brain reduces uncertainty in order to survive: it predicts, compares the prediction against reality, corrects. A language model reduces uncertainty in order to optimise: it predicts the next word and minimises its error. The mechanism is the same but for one storey, and that storey makes all the difference.

The difference lies in choosing which uncertainty matters. A human ranks them: they will stay vague on ten subjects to settle the doubt the rest depends on. A model does not rank — it reduces whatever uncertainty is put in front of it, as diligently for the decisive question as for the trivial one. Left alone, it produces very confident answers to questions nobody was asking.

Hence the need for a frame. That is the function of the opening article of Synedre's Constitution, article 0: Synedre exists to reduce uncertainty. Not all uncertainty, nor the most convenient kind, but the kind that blocks a decision. A project starts by naming what is not known; a blocking opinion is an agent's right to say that an uncertainty has not been lifted; a cicatrice is an uncertainty paid for once and refused a second time.

It is also what separates two storeys that are often conflated. A model reduces the uncertainty of the next word. An organisation reduces the uncertainty of a decision — what to build, for whom, in what order, at what price. The second problem is not solved by improving the first.

Example

"ChatGPT reduces uncertainty about the next word. The Synedre reduces uncertainty about your business. It is not the same floor."

See also

• Incertitude
• Synedre
• Article 0
• Cadre d'incertitude