If I had to guess, I'd say they're coming up against the limits of LLMs where LLMs are ... stupid. There's a quality issue that is difficult to get around when you try to make an LLM do something out of its comfort zone and it can suggest you to do dangerous things very easily.

I wish they would forget the really advanced features and just focus on little ones. Like particularly with Alexa back when I used to use that, one of the major frustrations was figuring out the magical word order that would get it to do what I wanted, it would only do a thing if I said it in a certain way otherwise it would get condfused or do the wrong thing. An LLM would be great at making that not happen.

But instead of just that little bit, they're trying to do Star Trek levels of computer voice control, missing out on incremental improvements.