Went out of their way to show actual usage of these features on actual devices in actual people’s hands

Heh, I noticed the same thing, after the DaringFireball callout last year about the normal product demo progression. It looks "real" this time, but the question is how far along we are: will the journalists have a chance to play with it at the event?

It’s in the developer beta, people are posting videos of it, e.g.: https://x.com/iupdate/status/2064078761856037112?s=46

I noticed that, but I also noticed that it is ridiculously slow too which prompts the person to keep talking while the response was being generated.

After getting it in my hands, it's the same. At least 4 times slower for similar basic Siri responses. My guess is they are doing less local and more server-side generation to start as the on-device models might not be good enough yet.

You could see jitter in the reflection on the MacBook as the guy typed, so that all looked great.

The responses came in very fast though, so I’m sceptical that the latency is representative (or that they didn’t cherry pick results, but they looked LLM generated). We shall see though.

Even for live shows they'd cherry pick specific scenarios that were known to work, and those would sometimes fail. IN a heavily produced pre-arranged pre-determined marketing video? It was as polished and made as smooth as possible.

I think that’s mostly okay, I just worry about the times. It was like 2-3 seconds to search and pull context then generate the answer.

I’m writing AI apps these days, and even pulling Gemini 3.5 flash on Google Cloud takes longer to get a multi-step response.

Obviously the video is not representative, and there are fast models on fast hardware. But if this takes 2 minutes it’s not very compelling to users.