Hacker News Reader

There's Gemini Nano which runs on-device

No, it's not all just calling to a cloud backend. The newer Pixels have hardware to do some stuff on-board, including non-trivial stuff like realtime audio cleanup (that works amazingly btw). STT and TTS are done mostly locally as well (with some cloud processing used when network connection is available), but can work all offline. The Gemini Advanced stuff is going to the cloud though.

Yeah I was primarily speaking about the LLM part, as apple is doing that (smaller LLM models will handle simpler tasks on device and hand over to cloud when necessary + apps will be able to efficiently deliver fine-tuned LORAs for the already-available local models).

I do remember a lot of other typical "AI" stuff being done locally and well when I had a pixel a couple years ago.

I fail to see what is much better about a system where you don't know when it sends your data to the cloud or not (and is likely to do it 90% of the time).