What happens when I ask smart glasses a question?
A typical voice request passes through several stages: microphone capture, speech recognition, application logic, answer generation and output. Different products divide those stages among the glasses, a phone and remote services. The frame is only one possible place for the computation.
Follow one request through the system
Imagine asking for the next appointment. Speech recognition turns your words into text; the application retrieves the relevant calendar information; the response is formatted for speech or a display. A general language model alone cannot know about a calendar it has not been given access to. We like this example because it separates understanding the request from obtaining the information needed to answer it.
A useful architecture drawing labels each stage with the device that performs it. If internet disappears, you can then identify which stage stops. That is more informative than describing the entire product as either online or offline when different features take different routes.