I'm a native speaker and this was my experience as well. I had better luck manually sending the message with the "push to hold" button.
I know this is a serious conversation, but when the presenters had to cut it off, I got flashbacks to Data in Star Trek TNG!! And 3PO in Star Wars!
Human: "Shut up"
Robot: "Shutting up sir"
Turns out rambling AI was an accurate prediction!
When you tell Siri to shut up, it either apologizes or complains about your behaviour. When you tell Alexa to shut up, it immediately goes silent.
I prefer the latter when it comes to computers.
It also tangentially reminds me of an excellent video I re-watched recently called The Tragedy of Droids[1]. I highly recommend it. It raises interesting moral questions about the nature of droids in the star wars universe.
I do my work in english and talk to people just fine, but with machines it's usually awkward for me.
Also on your other note (demo seems to be very rambly), it bothered me as well. I don't want the AI to continue speaking, while having nothing to say until I interrupt it. Be brief. That can be solved through prompts at least.
That's been a major issue for me with LLMs this whole time. They can't just give me an answer, they have to spout a whole preamble that usually includes spitting my query back at me and trying to impress me with how many words it's saying, like it's a requirement. You can tell it e.g. "don't give me anything other than the list" but it's annoying to have to ask it every time.
Every AI chat needs a "no yap mode"
What I notice most is that almost always repeats verbatim unnaturally long parts of my requests.
This might be more useful to people that do lazy prompting. My nature compels me to be clear and specific in all written text.
I'm guessing not quite possible now, just because I'm guessing patiently waiting is a different band of information that they haven't implemented. But I really don't know.
It shouldn't be too hard to make this work though. If you make the AI start by emitting either a "my turn to talk" or "still listening" token it should be able to listen patiently. If trained correctly.
Stop talking for x sec = process response.
I bet the bot would wholeheartedly agree that it would definitely wait for you to finish talking, then just not do it. It doesn’t know anything about the app it’s “in.” At least at a deep level.
I agree that all this is impressive, but with odd, unclear bounds that sometimes confuse users.
SEO spam got into the training data.
It's probably related to GPT's more general sycophant inclinations. Acting like a doormat is apparently easier to teach than nuanced politeness -- much in the same way that repeating yourself ad nauseum is easier than intuiting specific points of emphasis.
I tried using the voice chat in their app previously and was disappointed. The big UX problem was that it didn't try to understand when I had finished speaking. English is a second language and I paused a bit too long thinking of a word and it just started responding to my obviously half spoken sentence. Trying again it just became stressful as I had to rush my words out to avoid an annoying response to an unfinished thought.
I didn't try interrupting it but judging by the comments here it was not possible.
It was very surprising to me to be so overtly exposed to the nuances of real conversation. Just this one thing of not understanding when it's your turn to talk made the interaction very unpleasant, more than I would have expected.
On that note, I noticed that the AI in the demo seems to be very rambly. It almost always just kept talking and many statements were reiterations of previous ones. It reminded me of a type of youtuber that uses a lot of filler phrases like "let's go ahead and ...", just to be more verbose and lessen silences.
Most of the statements by the guy doing the demo were interrupting the AI.
It's still extremely impressive but I found this interesting enough to share. It will be exciting to see how hard it is to reproduce these abilities in the open, and to solve this issue.