Who Owns Your AI?

Download an open model, run it on your own machine, switch on airplane mode: it keeps answering. It feels like independence.

It is borrowed. Whether there is a better open model next year, under what license and with what capabilities, is decided by a handful of organizations — nearly all of them in the United States or China — and by the governments with authority over them. Self-hosting does not make the dependency disappear. It only moves it to where you no longer see it.

ownAI's promise is "the AI that belongs to you." One half of that promise is already true: conversations, memory, tools, self-written programs — everything stays on your device, and nothing leaves your machine except the requests to the model you chose. The other half no one has delivered yet, including us: the intelligence itself comes, here as everywhere, from the hands of a few makers. This text is about how we intend to change that.

A single point of failure

The dependency can tighten at any time. Makers have already hardened licenses and stopped publishing open weights; states already restrict, citing security, which models may be freely released. A system whose whole purpose is autonomy on your own device must not have a single point of failure at its most important link.

But insurance is the weaker reason. The stronger one is what becomes possible.

The guild principle

A guild was an association of craftspeople who got better together without any master handing over their secrets. Federated training makes exactly this form possible again — for intelligence.

It works like this: every member trains on their own machine, on material they chose themselves. Not a single document is shared — only a mathematical distillate of what was learned, an adapter update smaller than a photo. The distillates flow together into a shared model; the data stays home.

A legal-aid association would get a model that drafts in precisely its house style, without a single client document ever leaving a member's machine. A network of repair shops would get a model that knows thirty years of fixes no manual ever recorded: the noise the error code does not explain; the workaround for the part that is no longer made.

Nobody joins a guild out of idealism. You join because the exchange pays: give one distillate of your own experience, receive the distillates of everyone else's. The resulting model is better — at what the group actually cares about — than anything you could rent. And it belongs to the people who built it.

The road there

We move in steps, and every step has to be worth it on its own.

Today: ownAI runs entirely on your device. Open models locally, or any provider you choose; through its memory, your instance gets to know you over time.

Next: the personal model. One person refines an open model on their own machine, with their own material — and feels the difference in their daily work. This is what we are working on right now. The results will appear here, including the failures.

After that: the first guild. A small, hand-picked group merges what its members trained. Every contribution is voluntary, holds for one round, and can be inspected before it is shared.

In the end: decentralized training of entire base models over ordinary internet connections. This is active research; models in the tens of billions of parameters have already been trained this way, still far from the frontier. If that research matures, guilds are how its results reach ordinary people. And should the day come when no one may publish open weights anymore, communities that can train together are the only way open models come to exist at all.

Why now

A race of closed, national AI systems leads to a dead end. The alternative is open systems, built across borders, owned by the people who build them. You can wait for them — or start. An AI that belongs to you is not one you rent. It is one you help build.

Does that resonate with you? Anyone can use ownAI, build with us on Codeberg, or follow the work on Mastodon and Bluesky.