I think its possible for apple to allow LLM siri to work on for the new HomePod mini.
However it will come in one of three ways.
1. Most likely piggy back off of your Apple TV 4K (A17 pro only)
More likely would be piggy back off a smarter HomePod-family device. The HomePad probably would have
Apple TV like performance since it is driving a larger screen ( than other Pods) anyway.
The Mini coverts the voice to text and text shipped to HomePad hub and smarter stuff comes back.
If both the tv and new mini get Bluetooth 6 and Wi-Fi 7 they have the means of strong communication between devices.
Wi-Fi 7 doesn't necessarily increase range.
Ask Siri a question it gets fired to the TV that does all the LLM work on device and fires it back to the HomePod to reply.
If LLM work is larger than bounded range of the A17 Pro it is going to the cloud anyway. Dispatching exclusively to a local hub isn't necessarily the best , low latency path from the Mini. If it fits within the bounds of local processing then it is better as more private and more energy efficient.
However my 1st gen HomePod mini’s have always been unreliable and adding this into the mix is just another point of failure in my eyes.
there connection is poor (not necessarily Mini's fault ) or their ability to preprocess the info before passing to the internet is poor. Or the backend they were connected to is poor (in this case wrong blame assignment).
2. Make the new mini completely independent which would require a specialist SOC (no re-using Apple Watch chips)
Quite unlikely. If wanted more expensive Pods then could put the same SoC that was in the AppleTV above in them. The problem is that the price doesn't work for the mini. A more expensive mini runs into the same problems that the more expensive HomePod ran into (fewer buyers ).
The HomePad will present a screen so likely it will come with a higher price. If it is AppleTV+MiniPod priced ( i.e., > $250) then the system has enough revenue to cover a faster chip. In short, need to add some value besides "better LLM chatting" to justify higher price.
A 'specialized' chip isn't going to make the RAM any cheaper to do more complicated LLM workloads. Most of the voice to text can be done local but that is not outside the bounds of a Watch chip. (that technically isn't LLM work. LLMs only work
AFTER you have tokenized what is said. )
The HomePod 2 was essentially re-engineered to be cheaper to produce because no one could justify the cost.
right.
3. Have the new HomePod mini rely fully on cloud. No processing on any device apart from pinging the cloud to send questions and receive answers.
Amazon lost billions on shipping everything to the cloud. It is a dubious idea at large scale.
As stated above LLMs do not work on 'voice" data. They work on tokens. Really short command sentences a limited set of common Apple features don't really need a huge LLM to figure out. 'Set a timer for xx minutes' , 'pause play' ' resume play' , 'skip this track' , etc.
Of course this would be a step back on apples privacy commitments.
Not necessarily. If the Mini is running as a proxy to an AppleID account then have the basis for setting up a
Private Cloud Compute dispatch.
Yes private cloud compute is a thing but Apple have always praised themselves on only sending complex commands to the cloud and doing basics on device wherever possible.
It wasn't "complex commands" . It was "too complex to do locally commands" . It was always an estimated workload filter as how to alter the dispatch of the require. Someone with a Mac Studio Ultra Maximum memory would not dispatch as much stuff as a iPhone. It wasn't going to be the same for every single Apple ecosystem device.
The Mini would just 'punt' more than other stuff. That has implications as to whether Apple has enough 'back end' to handle all of those. "Set a timer for 10 mins" is bit much to fire up a PCC session , drag in local and user specific context , and then immediately flush it all out again.