I've had the latter with Gemini in Android Auto. In fact, my scenario was even worse.

Me: "Send a message to ".

Gemini: "OK what's the message?"

Me: "Can you bring Tylenol I'm having a headache?"

Gemini: "I'm sorry, I cannot help with that. If you need medical advice..." <-- or something to that extent.

Up to this point, I thought when Gemini would prompt for a message, it would simply accept whatever I said essentially as plain text - and maybe with the new Rambler feature on Pixel phones it may cleanup my message slightly.

But now I can see scenarios where Gemini would possibly need to parse my input. And it should have been obvious since I would often say things like "cancel" and I never questioned it.

It’s really pathetic. I would think even the lowest quality LLMs could figure out tool use like this. It has to figure out the tool and most of the time, it’s stated in the command. Like remind me or put this in notes, so why can’t it do this more easily. I’m just having trouble figuring out why. I’m curious if they still have tons of rules in the background in a way that is muddying the pure LLM experience.

I don't use voice often (I'm also not quite sure what you're asking with 'three fifty', but if I just type '50F' into the new Siri it immediately shows '10C'. Hopefully the voice recognition continues to improve.

Three fifty = 350?

It’s not an unusual way to save effort when speaking.

Ah, doh! I just tested it and it worked. Siri clearly can answer the question, now it seems more like a speech to text issue. I wonder if everything is using the new models or if stt is still using the old ones?

“Then preheat the oven to three fifty, and give that spoon a lick …”

No high hopes for how that recipe turns out with Siri in the kitchen then