For most of the last few years, the interesting question about artificial intelligence was whether it could do a thing at all. Write the email. Read the contract. Summarise the meeting. That question is largely settled, and it settled faster than almost anyone expected.
Which means the interesting question has moved. It is no longer can it. It is what does it know.
Inference became a commodity
Inference is the part you see: you ask, it answers. It is what every assistant, chatbot and copilot does, and it is now available from a dozen providers at prices that fall every few months. That is a genuine achievement, and it has a consequence that is easy to miss.
When everyone can buy the same capability at roughly the same price, the capability stops being an advantage. If your assistant and your competitor's assistant are both answering from the same public internet, they will give roughly the same answers, because they are the same thing wearing different paint.
This is the position most businesses and most households are in right now. They have access to something remarkable, and it knows nothing whatsoever about them.
What is actually scarce
The scarce thing is not a model. It is a model that knows something nobody else knows.
Consider what you already hold. Years of correspondence. The paperwork on the house, the practice, the family. Contracts, invoices, notes to yourself, the running record of decisions and why they were made. Most of it is not written down anywhere useful — it sits in folders, inboxes and drawers, and its value is entirely locked up in the fact that finding anything in it takes longer than remembering it yourself.
That material is the asset. It is the thing a general model cannot reach, cannot buy, and cannot replicate. And the moment an assistant can actually use it, the quality of what it tells you changes in kind, not degree — because it stops answering from the internet and starts answering from your own record.
Which is why training moves
Here is the difficulty, and it is the reason this has not already happened everywhere.
To train on material, something has to read all of it. Every document, every message, every note. For a general provider that means the material has to travel to them, sit on their infrastructure, and be processed under their terms. For a household archive or a firm's client files, that is not a small ask. For some professions it is not an ask that can lawfully be granted at all.
So there is a real tension: the material that would make an assistant genuinely useful is exactly the material you are least willing to hand over. Most people resolve that tension by not resolving it. They use the general assistant for general things, keep the important material out of it, and accept that the assistant stays a stranger.
The other resolution is to move the training rather than the material. Put the machine where the records already are, train there, and let nothing leave the building. The material never travels. The model that results is trained on your world and is useful only to you, because it is a model of your world and nobody else has one.
What that changes in practice
The distinction matters most in the ordinary moments rather than the impressive ones.
A general assistant asked what you spent on the guest house last year will tell you how to find out. An assistant trained on your own records will tell you the number, what most of it was, and that the same repair appears twice. The first is a search engine with better manners. The second is closer to someone who has worked for you for a decade.
Nothing about that requires a larger model. It requires a model that has read your papers — which is a completely different problem, solved in a completely different place.
The honest part
Two things are worth saying plainly, because this subject attracts a great deal of overstatement.
First: training on your own material does not make an assistant better at everything. It makes it better at you. For a question about the world — a legal principle, a piece of history, a language — a large general model remains better, and any honest system should be able to reach for one when you say so, rather than pretending it knows.
Second: this is work. Archives are messy. Some of what you hold is scanned badly, duplicated, or wrong. Training on material that contradicts itself produces an assistant that contradicts itself. The preparation is the unglamorous majority of the job, and anyone who tells you otherwise has not done it.
Where this goes
The next few years will not be a race to a cleverer answer. Cleverness is becoming cheap. They will be a race to assistants that actually know the person or the business they serve — and the determining factor will be whether the material could be used at all, which is a question about where the training happens.
That is the shift worth preparing for. Not a better model. A model that has read your world, kept in the building where that world already lives.
What that looks like day to day →
Start with a written scope of your own material →
Local AI. Private data. Local training.
$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?