Our first ambient devices had a single design constraint: they were not allowed to ask for attention.
No wake lights for their own sake. No ambient glow that said “I am thinking.” No notification vocabulary borrowed from phones. If the object needed a person, it had to earn that need with something the person would have wanted anyway.
How much intelligence fits
Less than we expected, and more than is useful to advertise. The useful ceiling was not model capability. It was how much inference could run without changing the object’s presence in a room. Once the device felt “on,” people treated it like a screen — which defeated the point.
We kept three behaviors and retired two prototypes that performed well in demos and poorly on a Tuesday morning. The survivors share a pattern: they act late, explain only when asked, and look unfinished until you need them.
Attention cost as a budget
We started measuring attention cost the way teams measure battery: as a budget that must not be spent lightly. A correct answer that costs a glance is sometimes worse than a slightly later answer that costs nothing. That trade-off is uncomfortable in a culture that equates responsiveness with quality.
These notes are incomplete on purpose. Quiet hardware is still early. We are sharing the constraint set because it keeps us honest more than any roadmap slide does.