Impressive demo, but like half the interactions were "hello" "hi how are you doing" "great thanks, what can I help you with" etc.
The benchmark for human-computer interaction should be "tea, earl gray, hot", not awkward and pointless smalltalk.
Seriously. I've had to spell out that it should just answer in twelve different ways with examples in the custom instructions to make it at least somewhat usable. And it still "forgets" sometimes.