Our position

Why we build this way

For most of the past decade the interesting ethical question in AI has been what a model should say. We think the question that matters now is who gets to build with one.

Access is the upstream question

Capability concentrates where compute concentrates. When the machines able to run serious models sit inside a small number of organizations, the ability to shape what those models do concentrates in the same place. The people best positioned to check that work become the people who cannot afford to run the experiment.

Open models are a real and underappreciated shift. But a published model only reaches as far as the hardware available to load it, and weights without machines are permission rather than capability.

You cannot audit what you cannot execute.

Local-first is the part we can build

We are not in a position to change how compute is distributed across an industry. We are in a position to decide where our own software runs. Every product we build assumes the sensitive part of the work stays on a machine the person controls.

That is a structural commitment rather than a policy one, and the difference matters. A promise not to read your work and an architecture that cannot read your work fail in completely different ways, and only one of them fails safely. Privacy that rests on a terms-of-service document can be revised next quarter. Privacy that rests on where the data physically sits cannot.

What that means in practice

Safety and freedom are not opposed

A common framing treats user freedom and system safety as a trade, where more of one buys less of the other. We do not think the evidence supports it. Real safety comes from transparency, robustness and human oversight.

Consider what it takes to establish that a system performs as intended. That is an empirical claim. You test it, reproduce the result, and let independent parties probe the system and report what they find. Every one of those activities requires access. Opaque filtering that pre-decides what a person may encounter produces reassurance instead of safety, and it fails quietly, which is the worst way for a safety measure to fail.

The test we apply

When we make a design decision, the question we ask is whether it leaves the person more informed and more capable, or less. Software that runs where the work lives answers that in the user's favour by default, and that is the reason we build the way we do.

← Back to Server9