On the matter of SuperIntelligence

Those three lines are easy to read as modesty, or as a small company making a virtue of what it cannot afford. They are neither. They are a position on where the useful work in AI is right now, who it is for, and what a European company should be building.

This post sets out that position, and what it commits us to.

The race we are not running

A handful of laboratories are trying to build systems more capable than people at nearly everything. Some of them take the risks of that seriously, and the safety and interpretability work they publish is among the most important research being done anywhere. We read it closely, and we are grateful for it.

But a race to the frontier asks for a particular kind of company: one that can spend without limit, ahead of any revenue, on the bet that the prize at the end repays everything. We do not think every AI company should be that company. We do not think Europe needs one more entrant in that race to have a serious place in this technology.

Most of the value AI will add in the next decade will come from models that are good enough, cheap enough and trusted enough to be used every day.

Not the model that tops a leaderboard for a month. The model that drafts the reply, reads the contract, checks the invoice and fixes the build, a thousand times a day, for a price that makes sense on the thousandth time.

What we build instead

We build models for work, and we judge them by the work. The question we ask of a new model is not how it scores on a benchmark, but whether it finishes the task a person gave it, how often it needs correcting, and what that finished task cost.

Our default models, Janus 1.0 Flash and Janus 1.0 Pro, are what we run Kaer on, and what we use for our own day-to-day work. If a model is not good enough for our own work, it is not good enough to sell, and we find out first because we are the ones it lets down.

That discipline keeps us honest about cost, too. A model that is impressive but expensive per task is not a product for most people. It is a demonstration. We would rather ship something smaller that a small business, a school or a public office can afford to run all day.

Cost is how access is decided

It is tempting to treat price as a commercial detail and capability as the real story. We think that gets it backwards. The price of a useful answer decides who gets to use AI at all, and how often. When the best tools cost frontier prices, the benefit goes to the organisations that were already ahead.

So efficiency is not a compromise we make on the way to something better. For us it is the research problem. Every gain in work done per euro widens the set of people for whom this technology is useful rather than merely interesting.

Sovereignty is a property you can check

"European sovereign AI" is a phrase that can mean almost anything. We mean something specific, and checkable.

• Janus 1.0 Flash and Janus 1.0 Pro, our default models, run in European data centres.
• Kaer's own servers and primary storage are in the EU.
• Our Data Processing Addendum under GDPR Article 28 applies to every business customer, on every plan, with nothing to sign.

We will also say what that does not cover. If you choose a different model inside Kaer, it is run by the provider our Privacy Policy lists for it, with that company's country, and some of those are outside the EEA. Sovereignty that is claimed for everything is usually true of nothing, so we would rather draw the line where it actually is. The details are on our EU inference and Kaer for Europe pages.

Sovereignty matters because dependence is a risk, not a slogan. A continent whose companies, public services and research all run on models it cannot inspect, host or price for itself has handed away a decision it will want back. Building capable, affordable models that run here, under European law, is one practical way of keeping that decision.

Your data is not the product

We are paid for the work our models do: plans, credits and the API. Not for what we learn about the people who use them. That shapes every choice we make about data.

We never use customer data to train third-party foundation models. Each person has a setting, "Help improve our services", which lets eligible content, with secrets removed, help improve Kaer's own models. It is on by default, and we want to say so plainly rather than bury it. Anyone can switch it off in Settings, which deletes the samples kept. Private chats are never collected. Organisations that need more can agree zero data retention in writing.

We think that is the honest version of "not in the business of collecting your data": a company whose income does not depend on it, with a switch each person controls and a written promise for those who need one.

For use, and for the public

The phrase we keep coming back to is AI that matters for scale, for use and for the public. Scale, because a model only changes anything when it is cheap enough to run everywhere. Use, because a model is measured by what it gets done. And the public, because the institutions that most need dependable, affordable AI, the clinic, the council, the small firm, the classroom, are the ones least able to pay frontier prices or audit a black box held somewhere else.

We do not claim to have solved this. We are a small team, and the problems are large. But we think it is the right direction, and the one a European company is well placed to take: building for the many uses that already exist, under rules people can read, at a price people can pay.

What this commits us to

• We run Kaer on our own models first, and ship only what is good enough for our own work.
• We measure models by finished work and its cost, not by leaderboards alone.
• Our default models run in Europe, and we say clearly where that stops.
• We are paid for work, not data, and each person controls whether their content helps improve our models.
• We will keep writing here about what we learn, including what does not work.

Superintelligence may come, and the people working carefully on it deserve support. In the meantime there is a great deal of ordinary, valuable work to be done, and most of it is waiting for a model that is affordable, useful and trustworthy. That is the model we are building.

Frequently asked questions