Now / Case study
A voice interface for an AI that stays close.
I am exploring what it feels like to talk to a capable model running locally: no dashboard between the person and the system, no cloud dependency required for the first conversation.
The question
Can a local model feel present without becoming opaque?
The experiment starts with a voice interface. Speaking is immediate; the interesting work is everything around the response: when to listen, when to stop, what the system remembers, and how a person stays in control.
What the tests show
The current local runtime has been exercised on Apple Silicon with Metal acceleration, unified memory, and a small quantized model loaded on the machine. The runtime can distribute work across the GPU and CPU, which makes the experiment possible on personal hardware rather than a remote cluster.
The point is not to claim that every local model is fast enough for every task. The point is to make the boundary visible: what can run locally, what the interface should communicate, and where a human should remain in the loop.
The interface idea
A good voice interface should make the system feel understandable. It should expose state without forcing the user into a terminal, preserve the option to inspect what happened, and fail in ways a person can recover from.
What I am not publishing
No proprietary code, private paths, credentials, model hashes, internal names, or closed implementation details. This page is about the concept and the lessons from the experiment, not about revealing the system itself.