Transformers can fit about 4000 tokens in the context window. That's a lot of dialogue rounds. The problem is these models are too expensive to offer at scale.
We almost never get to see the latest AI in big corporation products. It is always AI from a couple generations ago that is cheaper to run. When we get the latest and greatest, like GPT-3 and DallE, then it's super expensive to use, not free.
Wonder what's blocking that from being applied to systems like Alexa.
You (Amazon) don't want your precious little bot say random things that land you in the press. And oh, GPT-3 will.
Point two is that models that can be goaded to a specific place. So you'll need to account for that too, or... oh look, press.
Point 3 is that they're trained on the Internet at large, and you don't want that in your house :)
It's a lot easier to converse gracefully than to add useful info to a conversation.
This is essentially what has made past intelligent assistants so dysfunctional (like Microsoft's Clippy). They couldn't model cognitive concepts like 'relevant', 'interesting', or 'necessary' relative to the active topic.
Until NLP can create a short-term memory model to retain info on prior utterances, Alexa's comprehension will be limited to self-contained single phrases -- spells and incantations.