>I'm less certain about Amazon but unless (insert AI company) wants to take on all the business risk of hosting governments and corporations and hospitals on a cloud platform I think Amazon can just publish their own models, buy someone else's, or integrate with multiple leading AI model publishers.

Amazon is capturing massive amounts of the value in AI via AWS. They'll be fine. But for real I don't see a reason why Alexa is not using a good LLM now. Could just be infinitely better...

▲

taeric 6 days ago | parent | next [-]

I'm not clear on what would be better about my Echos having an LLM interface. I want to like the idea, but I'm also growing increasingly annoyed at how things that used to work are beginning to not.

If they could gate it behind a "start chat session" or something, I would be more excited. Doing it by cannibalizing how well basic "play radio/start time/read from audible" worked for the longest time, everything they do that causes friction there is frustrating, to the extreme.

▲

vonneumannstan 5 days ago | parent [-]

>Doing it by cannibalizing how well basic "play radio/start time/read from audible" worked for the longest time, everything they do that causes friction there is frustrating, to the extreme.

Theres absolutely no reason why plugging in an LLM would break any of those features but asking generic questions would be 100x better than "Searching the web for a shitty Quora or Alexa answers question."

	▲	mjmas 5 days ago \| parent \| next [-]
		Pluhging in an LLM breaks it when those features require it to be worded a certain way. Agree that it shouldn't, but that seems to be the way it is often done. For example, I tried Google's Gemini a while ago instead of Google Assistant on my phone and it was unable to do basic things like 'open the Signal app' and would instead go on a big tangent about how it can't open the Signal app for me, but that I could open the Signal app by finding it on my home screen or if I don't have it installed I can download it from the play store etc.
	▲	taeric 5 days ago \| parent \| prev [-]
		In principle, I agree with you. In practice, it keeps getting worse. I also don't typically ask generic questions. Ever, that I can remember. Again, I don't want to dislike the idea. If people are really getting value from it, I would like them to continue to do so. But it seems to be a more expensive way to service use cases that were working just fine.

▲

xena 6 days ago | parent | prev | next [-]

> But for real I don't see a reason why Alexa is not using a good LLM now.

Large language models are too slow to use as real-time voice assistants. ChatGPT voice only barely works because they have to use a much worse (but faster) model to do it.

	▲	coredog64 5 days ago \| parent \| next [-]
		Amazon has a commercial Speech-to-Text model (Nova Sonic) that is passable. I used it to create a post-sales call assistant and was surprised that the underlying model was able to do a bunch of stuff I thought I was going to have to use Claude for.
	▲	vonneumannstan 5 days ago \| parent \| prev [-]
		At least on paper OpenAI claims the Voice models are actually the ones you are picking i.e. GPT 4o, 5. In any case even a GPT 3.5 would be superior to current alexa...

▲

qcnguy 6 days ago | parent | prev | next [-]

Alexa was a huge money loser for Amazon even before LLMs. They can't afford it.

▲

aleph_minus_one 6 days ago | parent | prev [-]

> But for real I don't see a reason why Alexa is not using a good LLM now. Could just be infinitely better...

Alexa would "a higher order infinity" better if it wasn't spying on you ...