I think there's a line worth drawing here: Pre-LLM voice interfaces required you to guess the command(s) the designer of the thinking were having in mind for the action you want to perform. With LLMs you can be 10ft into human-level vaguery and metaphorism and your intent might still survive.

So the difference wrt discovery is that you only have to gesture at what you wanna do and, if a matching action exists, there is a chance it will be understood.

I'd wager we'll see a renaissance of voice assistants with LLMs, especially once the good-enough ones can run on device.

When you put it that way I am reminded of my youthful efforts at solving puzzles in Sierra Online text adventures by guessing the right prompt. I guess I got early training for successfully interacting with LLMs :)

it reminds me of all the old text games like Zorg or whatever where they gave you a cursor leaving you to figure out what you could ask/say/do. which is fine for a game. it sux for a "tool"

> like Zorg or whatever

Ow, right in the AARP flyer.

There are obviously cases where CUIs work well: while driving if you don’t have a physical button to do the thing you want, the CUI is better than a screen UI, and if it’s a complex thing, you would never have a physical button for it anyways. In the home when you want to control devices in the morning “open all the shades” or “open the north shades”, unlock the door and start the car, etc…

Not really conversations though, more like accessible commands. I don’t think we get conversations until the tech improves and the latency goes way down, meaning on device processing of speech at the very least.

I’ve had exceptionally bad experiences with car CUIs - all the problems of trying to figure out exactly how to tell Alexa what to do (and then figure out what Alexa actually _did_) while going 70mph in traffic. Just give me physical buttons.