Evidence compounds. Intelligence does not.
If the gain arrives for everybody at once, it is not an edge. It is weather. The only line that rises for you alone is the one built out of the times somebody told the machine it was wrong.
A model release is a strange kind of event. Every competitor you have gets better on the same morning you do, by the same amount, for the same price. You did nothing. They did nothing. The floor moved.
I was excited about that for about two years before I noticed what it actually means. If the gain arrives for everyone at once, it is not an edge. It is weather.
The two lines
The dashed line is the rented gain. It rises, it will keep rising, and it costs you a subscription. Nothing on it is yours.
The solid line is what a system does when it keeps its own record. It takes the same rented lift everybody else gets, and then it steps above it every time a human corrects it and the correction holds. Those steps are the only part of that drawing you own.
The gap between the two lines is the whole business.
Why almost nobody has the second line
Because the second line is made out of boring, unglamorous, slightly embarrassing work. It is built from the times you told the machine it was wrong.
Every one of those moments is the most concentrated signal in your company. Somebody who knows the domain looked at a confident recommendation and said no, not that one, and here is why. That is your best person’s judgment, in writing, attached to a specific case, with an outcome you can go check later.
There is no other way to get that. You cannot buy it and a bigger model will not produce it, because it is not knowledge about the world. It is knowledge about your world.
A better model raises the floor for everyone. Only your own corrections raise yours.
The swap test
Ask what happens to your setup if the model you use gets swapped tomorrow for a different one from a different company.
If the answer is that you lose everything, you did not build anything. You picked a vendor.
If the answer is that the new model arrives and immediately inherits everything your team already figured out, every correction, every outcome, every case where the obvious call turned out to be wrong, then you built something. The model is a part you replace. The record is the machine.
This is why I stopped needing to bet the company on which lab is ahead this quarter. Not because the race stopped mattering. I care a great deal about which one is best at a given job, and I grade them. What I will not do is let the company memory live inside the answer.
The last part is what happens once the record gets good enough that the system can act on it by itself: when the loop closes, it starts acting.