[ManuelKiessling]

I think there is an important distinction to make. Siri is actually two things — there are voice informers (“Please explain how photosynthesis works”) and voice assistants (“Remind me to call Dave when I reach my work location”).

The former is really straightforward to implement with an LLM — it’s basically what an LLM is.

The latter is a whole different story.

[wkat4242]

It's not really. I've done the latter.

You just prompt it with "If I'm asking you to call someone, please output only "<CALL>" and the name of the person". Then capture that keyword.

It works fine like that.

[kaba0]

I mean, the latter is basically OpenAI’s assistant, that has some APIs it can autonomously call.

The open problem here is making it on-device, and privacy preserving. Though I’m optimistic about this, as Apple has bought up a huge number of AI startups in the last couple of years, so they are probably onto something.