Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
Delegating the construction of a knowledge base to a machine does not remove the human work: it shifts to three places where nobody had budgeted for it. Upstream, the choice of domain decides everything, because only knowledge fixed by standards can stand being consolidated in advance — applied to opportunity material, the same machine propagates errors instead of capitalizing. In the middle, the tuning of the ingestion rules, where you choose which way the system will err, because no threshold takes in everything that counts without taking in something else. Downstream, the validation of what comes out, which presupposes already knowing the domain — failing which you build on sand, without seeing it. Extraction, for its part, works on the first try: what has become easy gets noticed, what has become decisive does not.
Synthesis
The starting point is a very concrete lack. A setup that recomposes its answer at every question leaves nothing behind: the cross-references between documents are redone every time, under pressure, and never reread. What is missing is not the model's power, it is the existence of an intermediate object — a synthesis, a link between two concepts, a contradiction spotted — which must exist before the question or will never exist at all. The answer to that lack is a maintained artifact, and the tipping it brings about is less the automation of a piece of work than its redistribution.
The first shift is invisible because it precedes the project. A network of linked notes propagates everything it contains, noise included; whether it consolidates or contaminates does not depend on how well it is built but on the lifespan of the material put into it. The same practitioner can therefore abandon the method for product opportunities and take it up again for standards without contradicting themselves — and it is that sorting, done before a single line of instruction, that decides the result.
The second is visible in use and surprises every time. Extraction works straight away; calibration takes weeks. A permissive version produces noise, a hardened version loses relevant notes without leaving any trace of what it lost — hence the need for a procedure that makes that loss legible, by setting the hardened version against the previous one's output. The consequence for what has to be kept is clear: the notes can be remade, the body of rules cannot. The artifact that counts is not the one you look at.
The third does not automate at all. A base built by a machine is only worth something for someone able to judge what it has extracted, and that competence is acquired through the same reading work as before — books, standards, the field. The setup structures and capitalizes knowledge; it does not give it. What makes the difference between two practitioners equipped with the same tool therefore remains what they already knew, which is why the promise of a shortcut for non-specialists turns against itself: the better formed the output, the less detectable any falsity in it is by someone who lacks the level to check it.
What the three shifts have in common lies in their form. None produces a deliverable, none can be counted, and all three sit outside the operation you thought you were buying. That is the ordinary configuration of the work an organization stops funding because it does not see it.
Tensions / contradictions
Two of the angle's requirements get in each other's way. The notes are said to be derived, therefore destructible and regenerable at will — which is what allows hardening the rules without fear. But tuning those same rules requires keeping the previous version's output as a control set. The disposable artifact is also the only available yardstick, and nothing says which of the two properties should give way.
A second tension, about neutrality. The base is presented as factual and unopinionated, as against the personal system of thought that takes positions. Yet it is the ingestion rules that decided what deserved to be retained. The base therefore has a point of view — that of its calibration — which its notes do not own, and which nobody reads.
Questions
- What, within a domain, signals that it is stable enough to deserve prior consolidation, before you have paid for the ingestion to find out?
- Can the competence to validate be passed on to someone who did not build the base, or does it stay attached to whoever tuned its rules?
- Does a body of rules calibrated on one domain keep anything useful when carried over to another, or does everything have to be redone?