As much as it'll be "interesting" to see how models behave in real world examples (presumably similarly to how the demos went), I'm not convinced this is a premade recording like what seems to be implied.

I'm imagining this is an incomplete flow within a software prototype that may have jumped steps and lacks sufficient multi-modal capability to correct.

It could also be staged recordings. But, I don't think it really matters. Models are easily capable of working with the setup and flow they have for the demo. It's real world accuracy, latency, convenience, and other factors that will impact actual users the most.

What's the reliability and latency needed for these to be a useful tool?

For example, I can't imagine many people wanting to use the gesture writing tools for most messages. It's cool, I like that it was developed, but I doubt it'll see substantial adoption with what's currently being pitched.

Yea the behavior of the AI read to me more like a hard coded demo but still very much "live". I suspect him cutting it off was poorly timed and that timing could have amplified due to WiFi? Who knows. I wasn't there. I didn't build it.