Hacker News Reader

Comment by oezi | Hacker News Reader

> 4. ChatGPT is clearly better at answering questions.

It is stunning that Amazon wasn't the forerunner in using ChatGPT/LLM-based tech for Alexa. A mistake in the tens, if not hundred billion dollars when looking at the price MS paid for OpenAI.

Why aren't they hooking up a LLM right now?

View on HN

> It is stunning that Amazon wasn't the forerunner in using ChatGPT/LLM-based tech for Alexa

I'm a current scientist at Amazon. Amazon has little ability to invest in longer term research. While you have scope to pursue projects without immediate product application to some extent (though far less than you would at Alphabet/Meta/Microsoft), that is less valued than at the other Big Tech companies. There's also a far more bureaucratic hurdles to jump through to publish at Amazon. The lack of care for non-product research shows at every instance. Amazon also has extremely low -- or I should say non-existent standards for hiring scientists. If you're putting a person with a decent and non-existent pub record on the same level, that is not going to yield great results.

Luckily for Amazon, their bread and butter -- Ads via retail and AWS -- are not immediately threatened by LLMs.

Honestly if they had taken the billions they burnt on Alexa and put it in research they would be far far far better off today. But try explaining that to the bean counter mentality of executives that Bezos cultivated...

This is a good point. For an outsider it always seems that Amazon (or at least AWS) excels at deploying microservice solutions which plug into an ever growing ecosystem of interconnected offerings. I would have imagined that Alexa org is setup similarly and allows teams to deploy a solution such as an LLM and then given success be rewarded with growth.

Because LLMs are so unreliable. Alexa misinterpreting and giving a nonsensical/unrelated answer is annoying, but fine. If it was an LLM, when it was wrong it would be lying confidently and unless you're double checking everything it outputs (in which case just search it that way in the first place) then that means it'll be misinforming you without you having a clue, as opposed to the current system where it's obvious when it gets it wrong. That's clearly just worse.

And that's ignoring the fact that all the publicly accessible LLMs have been "jailbroken" and you can get them to say all manner of wild things, and Amazon is definitely aware that all it takes is one or two viral clips of Alexa saying something unhinged for their sales to start dropping, and the tech isn't currently sufficiently understood for them to avoid that.

I think you are missing the point that using LLMs to improve Alexa would certainly be hard, but it would have been massively worthwhile if it worked. Microsoft and Google are well aware that LLMs have the potential to upend the search dominance of Google.

I am also much less concerned about jailbroken LLMs, because Alexa is already so broken it can't get much worse. Normal users don't seem to care much about potential misuses (compare with search autocomplete biases of Google) as long as a service works for them. But unless you widely deploy a tool, you won't uncover how it is abused (see Sidney's launch).

> But unless you widely deploy a tool, you won't uncover how it is abused (see Sidney's launch).

ChatGPT was, while less widely deployed, being abused (very publicly) in very similar manners to the way Sydney was; Microsoft could have learned about those abuses by paying attention to the public information about ChatGPT.

I don't think you can learn about it much unless you have the tech deployed yourself. Microsoft obviously believed that their own prompt design and the learnings at OpenAI would prevent this, but was shown quickly otherwise. I believe they will use the beta to iterate this quickly to a point where we come to accept the residual level of abuse potential.

>I think you are missing the point that using LLMs to improve Alexa would certainly be hard, but it would have been massively worthwhile if it worked.

I really struggle to think how it would have worthwhile. My problem with smart speakers isn't that they aren't smart enough - it's that for anything slightly more complex, my phone has 100x better UX and information density. An Alexa LLM sounds cool in a sci-fi sort of way, but still feels like it would be useless.

Alexa could be your therapist? It’s always helpful to talk to your house when you are feeling alone.

“Alexa, be my friend and talk with me for a while, ok?”

I saw a demo the other day of a guy who had hooked Chat GPT up to some kind of voice assistant. The main benefit was that it supported chaining commands so you could say "turn on the light, then set the volume to 50% and then tell me the weather" and the device could parse it all and do the things you wanted in sequence. I can't find the video now but it was super impressive.

Google Assistant does that already. I know because I often tell it to "set volume to 1 and turn off the bedroom light"

Of course I only need to do this because I might listen to loud music during the day and if I forget to turn it down it'll blast "OK TURNING OFF THE MASTER BEDROOM LIGHT" at max volume right as the family is going to bed.

There isn't much of a need to hook up a LLM to Alexa's output. "Turn on the lights" doesn't need a bunch of text in response.

It's the input side of these things that needs a ton of improvements and the use of an LLM could be basically invisible to the user, just "summarize what the user said into a command"; my most common issues with voice assistants are if I say things out of the expected order or change my mind. "Alexa turn on the bedroom lights actually turn on the office lights" or somesuch, or something like "Remind me to make a reservation for Friday tomorrow" not distinguishing that "make a reservation for Friday" is the text of the reminder and "tomorrow" is the desired time.

Reply on news.ycombinator.com