On-device AI is further along than the skeptics think and less magical than the demos suggest. Both things are true, and the honest answer lives in the gap between them.
What the phone handles well
A modern phone can run a small, quantized model quickly enough to feel instant for the everyday work of a companion:
- Drafting and rewriting short text.
- Classifying, tagging, and routing your own notes.
- Answering questions grounded in the context you’ve already granted.
Because it runs locally, that work never leaves the device — which is exactly the kind of task you want kept close.
Where the sealed cloud picks up
Some requests genuinely need a larger model. When that happens the work goes to the sealed cloud — processed without being retained or readable after the fact.
The goal isn’t “never use the cloud.” It’s that the cloud never becomes a quiet second copy of your life.
The routing itself is boring by design:
const route = fitsOnDevice(request) ? 'local' : 'sealed-cloud';Small and local by default; larger and sealed when it has to be. You shouldn’t have to think about which — but you should always be able to find out.