I believe it's because it's at least a magnitude more expensive to run an LLM instead of a basic voice assistant ML model. Especially when considering anyone who has purchased a Google Nest Mini can use it, for free, forever.
run "set timer" and "what's the weather" and "turn on the lights" through some regexes as a first pass and save the LLM for the complicated stuff. actually, with some caching they can probably cut costs pretty heavily there too.
I’m guessing they are already doing this. Adding an LLM is asking to add a very expensive thing to the long tail of queries which likely cannot be cached.
And when you consider these voice assistants already feel at risk for further investment due to subpar results, it makes sense. Bit of a catch-22 though as many of us have given up on expecting much out of our existing ones
But a lot of the question /responses could be trivial cached. No need to run expensive LLM every time for the same basic "how are you today?" prompts, it only has to be cached once.
Caching static requests alone is hard enough. With all the ways you can ask this question, welcome to the most complicated caching backend ever. Caching exact matches would also not help much because of this.
Then you’re kind of defeating the purpose of an llm.
Fixed responses for common queries is what we have now.
Not to mention that LLMs tend to be very wordy right now. I’d hate to way 20 seconds to hear my phone say “As a voice assistant I’m not aware of the exact menu of the Thai restaurant on 2nd, but I have opened a google search for it and found the following results.
