Main idea

A graph tool splits a wiki of 1,013 marketing notes into 22 groups: one around the Mom Test, one on influence and Cialdini's principles, one on the chasm, one on customer success. The split itself is a computation: an algorithm groups the notes that are more linked to one another than to the rest, without reading any of their content. It runs in a fraction of a second and needs no model.

Yet the groups come out anonymous: "Community 0", "Community 1"… Giving them a name means reading the notes they contain and deciding what those notes have in common. The tool can do it by calling an AI — that is, by bringing back exactly the cost the rest of the chain avoided.

The dividing line therefore runs through a single feature: structure is computed, meaning is judged.

Why it matters

This gives a criterion for deciding where to place the call to a model in an analysis chain: not on the whole feature, but on the step that requires interpretation. The rest can be deterministic, free and reproducible.

It also says what is left to the human reader of a computed graph: interpreting the groups. That is real work, and it does not disappear because the split is automatic.

Nuances and limits

A group's name is less useful when the group is badly formed: if the split follows books rather than ideas, naming it amounts to finding the book's title again.

And manual interpretation does not scale: with a few dozen groups it is done by reading, with several hundred it becomes work you would want to delegate again.

Open questions

• From how many groups does the cost of having an AI name them become lower than the time spent interpreting them yourself?