A mind of our own

Some milestones do not look like anything. No new button, no banner — just a key removed from a configuration file and, with it, the last line that still passed your work through a borrowed mind. As of this week, every piece of artificial intelligence inside Spaces — the model that reads a document, the one that makes it findable, the one that picks out the people and places and dates — is one we run ourselves, on machines we control, in Europe. None of it is forwarded to a name in San Francisco any longer. It seemed worth explaining why we bothered, because the reasons are not quite the ones you might expect.

The honest first reason is not cost, though renting frontier intelligence by the call adds up quickly and owning it does not. The real reason is that a rented mind is one you can only ever gesture at. You send it your work and it sends back an answer, and between those two moments sits a sealed box you are not allowed to open. You cannot change how it thinks. You cannot teach it the difference between your client’s name and a brand it has never heard of. You cannot plane down the parts that do not fit or sand the edges that catch. It is finished furniture, delivered, and you live with it as it arrived. A model you own is a different material altogether — it is a piece of wood. You can work it: train it on the way your studios actually write, smooth the places it gets things wrong, shape it slowly to the grain of the thing you are making, until it stops feeling like a tool you reached for and starts feeling like a thing you made. There is a particular satisfaction in that which no API key has ever given anyone.

The second reason is quieter, and it matters more to the people who trust us with their files. When the intelligence is rented, your contracts and your invoices and your unreleased work are read, however briefly, by a system that answers to somebody else — read on terms you did not write, perhaps folded into the next model’s training, an asset on a balance sheet that is not yours. When the intelligence is ours, none of that holds. The model that reads your documents answers only to us. Nothing trains on your work. Nothing is mined from it or sold on. The reading happens, and then it is over, and the only thing kept is the answer you asked for.

The third reason is the one you only appreciate after it has bitten someone. When the mind inside your product belongs to another company, they hold the switch. They can change the price overnight, rewrite the terms, retire the model you built on, or simply decide — for reasons of their own, on a morning that is not yours — to turn the access off. It has happened to businesses that woke to find the thing their whole product depended on had been withdrawn or repriced beyond reach, with nothing to be done but wear it. A studio’s archive is meant to outlast the tools that read it. Ours can, because the tools are ours, and there is no one who can reach in from the outside and take them away.

We try to be precise about the claim, because the precision is the whole point. The true sentence is this: the models that read your files are open, are ours, and run on infrastructure we control in the EU, with no third-party AI company anywhere in the loop. It is a smaller sentence than the ones reached for elsewhere, and a sturdier one — the kind you can stand behind when a client asks you to account for it. Building it this way cost us more time than renting would have — and the deepest reason we did it anyway is the one that is hardest to say.

The companies building the largest models are, more and more and quite openly, selling them into defence and national security; the usage policies that once ruled military work out have been quietly rewritten, and the partnerships have followed the change in wording. We are wary of easy accusation, and we do not claim to trace where any given system finally ends up. But a tool can be a genuine marvel and still be turned, somewhere far down the chain, toward the targeting of a building with people inside it — and once you have let yourself see that, you cannot quite unsee it.

Let me put it more plainly than a company usually allows itself to. I have two young daughters. The people on the wrong end of a war are children, and the mothers and fathers around them, and the generations after them; I am against it without qualification, and for the slower, harder things — talking, listening, the patient work of two sides coming to understand one another. I cannot square that with quietly building on tools that, in the wrong hands, are turned to terrible ends — and the wrong hands are not always the ones you would picture, because a government can be the wrong hands too. A national flag gives no air cover, and neither does the phrase “national security”; both have draped, in the past and in the present, things that go uncalled as crimes and are crimes all the same.

They are crimes if you believe what I believe: that no one should be bombed out of their house, or chosen for death on a spreadsheet, or — worse, and now within reach — picked out by an AI and killed for looking different, for speaking another language, for living on the far side of someone’s refusal to understand. This does not make us pure; the open models we run were trained by others, and I will not pretend our hands are spotless. But where there is a choice, and here there was one, I would rather own what we build and be able to answer for it than rent it and look away.

And I do not intend to watch this from the sidelines, because I think it will become the question the whole field is judged on. The dazzle of these tools buys patience, not forgiveness; if the companies with the most power over them will not lead with a conscience, people will, in the end, refuse to follow them at all — and the wonder could be lost along with them, squandered on the uses a few of its builders allowed. That would be a tragedy of our own making, and the avoidable kind is the only kind worth getting angry about.