Almost everything we have built into machines imitates the fovea: a narrow, high-resolution, deliberate stare. Biology spends most of its budget elsewhere.
Ask a perception system what it sees and it will answer about the centre of the frame. Ask a person walking through a room and the answer is stranger — a sense of volume, of something having moved, of a surface being wrong before it is identified. The periphery does not report objects. It reports change, and it does so cheaply.
Cheapness is the feature
A model that attends everywhere at full resolution is not more capable; it is more expensive, and expense buys latency. Our working assumption is the inverse of the usual one: resolution should be requested, not maintained. A room should run on a coarse, continuous, almost indifferent read of itself and escalate only when the coarse read becomes interesting.
The goal is not a system that watches. It is a system that notices, and otherwise leaves you alone.
This is also where the ethics live. Continuous high-fidelity capture creates an archive whether or not anyone wanted one. A peripheral system, built correctly, cannot produce that archive — there is nothing in the buffer to subpoena.
What we are testing
- Event-first sensing: change as the primary signal, image as an exception
- Budgeted attention, with resolution allocated per second rather than per frame
- On-device summarisation, so that raw frames never leave the sensor plane
- Legible off-states — an instrument that can be seen to be idle
None of this is settled. We publish it because the framing has already changed what we build, and because we would rather argue about it now than defend it later.