[cle]

Because it's faster and works with spotty/no reception?

(I'm just guessing because I can't find details on this stuff by wading through the feed.)

Also, just because it's doing local inference doesn't mean it won't be uploading data back to Google periodically. As GP said, they have a strong incentive to do this, so it's reasonable to just assume they will do this until they can prove otherwise.

[monocasa]

And doing the inference locally, and just uploading transcripts for the most part is way more cost effective. They don't have to pay for the power on these endpoints.