ChromaticsHow it works
How it works.
Six drawings and six short arguments: what a model can do, what ownership changes, why size is a cost decision, what routing settles, what training costs, and why an unmeasured system decays.
What a model can and cannot do.
A model generalises from the examples it is given. Inside that range it is reliable. Outside it, it is confidently wrong.
Most of the engineering is knowing where that line sits, and deciding what the system does when it crosses it.
Rented. He reads your papers. You cannot train him.
Owned. He moves in. You train him.
Rented or owned.
A rented model cannot be trained on your data, frozen while you certify around it, or taken with you when the terms change. You pay per request, indefinitely.
Open-weight models are close behind and can be trained on your examples. The weights and the evaluation sets are then yours, and cost stops tracking volume.
Size is a cost decision, not a quality one.
The largest models are general. They cost more per request and need hardware you would rather not put beside a production line.
For one well-defined job, a small model trained on your data matches them, answers faster, and keeps working when the connection does not.
Routing.
A router decides which model answers each request. That single decision drives most of what the system costs to run.
Routine requests stay on your hardware. Hard ones escalate under rules you write, some data never leaves, and the local path still answers when escalation is unavailable.
Training is a data problem.
Training means fitting a model to your examples: your parts, your sounds, your manuals, your exceptions. The method is well understood.
The cost sits in finding, labelling and cleaning those examples. That is about half the work, and it needs your people as well as ours.
Conditions drift.
Cameras foul, suppliers change parts, vocabulary moves. A model that is not measured against fresh data is quietly getting worse, and nothing announces it.
Measurement and retraining are ongoing work with a real cost. They are the part most business cases leave out.
Why this is possible now, and was not two years ago.
Open models are now a few months behind the best paid ones, not years.
Running your own model is cheaper than renting once you use it all day.
In many buildings the data is not allowed to leave. Now it does not have to.
Send this to a colleague.
The six drawings and explanations as a four-page PDF.
We send the file and nothing else unless you ask.
Looking for the technical version? Read Under the hood.
Start with a conversation.
45 minutes, no slides, an honest answer.