Main idea

When you work up a product opportunity with an AI, what it needs amounts to very little: the mission file, the facts already established, the next steps to validate. A light, chosen context whose contents you know.

Open a borderless network of notes to it instead, and the behavior changes in nature. Every note opened suggests others, every link is a legitimate invitation, and nothing in the graph says where to stop. The model follows the connections, fills its context window with neighboring material, and hunts for the relevant information in an ocean of links most of which has nothing to do with the question asked.

This is not a steering flaw that a better prompt would correct. A borderless graph contains no stop signal: it is the structure of the material, not the instruction, that decides how far the exploration goes.

Why it matters

This makes the boundary an element of design, on the same footing as the content. What you make accessible to an agent determines its behavior more surely than what you ask of it.

It also gives a simple reading of a frequent symptom — answers that go wide of the mark on a subject that is nonetheless well documented: the problem is not that information is missing, it is that there is too much of it and no limit says which piece counts.

Nuances and limits

A selective retrieval mechanism — a search that returns only a few passages — bounds the exploration without bounding the base itself, and then moves the difficulty onto the quality of the selection.

And broad exploration has its moment: when what you are looking for is precisely what you didn't know to look for, wandering is the function, not the flaw.

Open questions

• What, in a corpus, can serve as a stop signal without having to carve out the scope by hand?