You're right, though - it's a bit weird because ostensibly the interface is via the GPT system and you've gotta work through those interactions (same problem Alexa had), but given how early both the platform and the product category are, they'd benefit big from letting early adopters build capabilities for them.
https://www.openinterpreter.com/
even their hardware is (apparently) open-sourced
https://brilliant.xyz/products/frame
They have shipped products before, and they include a bunch of code - today on GitHub to start hacking with.
The Rabbit R1 looks like it would be the perfect device to play with.
As is I just assumed Rabbit was sending off pictures and stuff to a Chat GPT API or something. I never assumed the models ran on device