Amazon blew Alexa's shot to dominate AI, according to employees (2024)

We ended up "closing down the project with enthusiasm".

All these years how did I live without a text to crappy image generator?

To me Perplexity Pro is clearly winning in terms of quality when I use it for this purpose. I have tried the 20 dollars Chatgpt and the 20 dollars google gemini as well. Chatgpt is pretty good and Google Gemini is pretty bad even with the 20 dollars a month but Perplexity Pro is by far the best AI for me. It's gotten even better with the recent Deepseek integration that they have done.

Anyway when I compare Perplexity and Chatgpt with the things coming out of the Big Tech companies I can't help but shake my head. Hasn't Amazon spent billions and billions on Alexa? Even then Amazon is clearly behind...

I’ve seen this said a few times lately, yet I’ve not really seen much of an explanation of exactly how it’s regressed.

Maybe it’s just a difference in how we use ours, but it still has worked the same as it always has.

Out uses are fairly basic, but I’m not sure how much more complex it needs to be:

- tell us the current/forecasted weather

- pick random numbers/flip a coin (playing board games)

- ask it for information about X

- play music on Spotify

- carry out actions in our home devices

All of these have always worked, and continue to work just fine. For from “lost all of their functionality” (this feels exaggerated).

Can you/others go into more details on how Google Home has become broken lately?

Small nit: the initial rollout of Gemini was a dumpster fire. It is now quite good. For my use cases, I can't get any other current LLM to give me better results than Gemini 2.0 Flash. It's also free

But even that kind of proves your point, right? Pretend you 100% believe me without verifying. That means in a year or two the winner has transferred between ~3 companies. This is not a cheap mantle to keep passing around. The AI wars are going to get heated over the next year or two

I’d love to know what the people in Amazon who came up with that thought they were doing. Did they not use the product themselves? Surely they knew their users mostly didn’t want that. Why not add a way to turn that off?

we tried the Alexa/Home lighting solutions this last year and it gets so annoying to ask every morning to turn the lights on. even worse when one of the bulbs drops from the network or the wifi is down.

I use and talk to chatGPT while driving to get things done and as a knowledgebase .... i.e. wanted get my car towed and junked and get best deal. GPT told me nearest junkyards and all details needed like need title & sign over title to them but it didnt interface with local junkyard's AI agent to schedule it all for me & send me an email confirmation or text. I wished it did.

That's because voice assistants are and have always been terrible, and could only play music, tell you the weather, set an alarm, and set a timer with any advantage over any other input method. People who liked using them for these things liked them. If you didn't need them for these things, there was no good reason to put a direct open line to Amazon, Google, or Apple in your house just to read you what wikipedia says about something.

People will love real assistants to death, especially local ones that learn from their personal habits. During the wave of voice assistants, they were promised that futuristic secretary through fraudulent advertising and endless industry marketing, and instead they got talking clock radios connected to advertising companies.

If anything, I'd love ditching the cloud for the basics.

I feel like this is the feeling at every big tech company

Being on the outside looking in it seems like these companies rapidly innovate

But when you actually work at one, it feels like everything moves at a glacial pace and the “right” answers are never the ones implemented

A good friend of mine has worked at Nvidia for years and even he was complaining that “nothing gets done” there

I would love a way to ask recipes and have it look up in my NYT Cooking subscription and give me step by step.

IMHO the basic modalities have to change to make use of LLMs in a voice assistant. Alexa and other assistants are built around short utterances that map to a command or simple question answering. There's not much room to make that smarter; things like chained commands aren't just hard to parse (except with LLMs), they are hard to compose.

There's a lot of household use cases they could tackle, but it's almost like building a whole new product. I doubt they have the organization capacity to do that.

There's probably another category of development in making it more socially aware, also building on its position in a household. But it would have to change some of its privacy expectations to be "aware" beyond its wakeword. They have pretty good speaker identification, and I don't know what all they could do with their mic arrays and a sort of audio/spacial awareness, but I bet some interesting stuff is possible. By socially aware I mean adapting length of response, follow-ups, giving more continuity over time, and so on. That's something LLMs do well, but involves hidden state and theories of interaction that these big companies are very reluctant to implement. They are only comfortable implementing it at the training stage where it ends up in opaque model weights and no one takes any responsibility for the result.

From the article: "Research scientists who worked on the LLM said Amazon does not have enough data or access to the specialized computer chips needed to run LLMs to compete with rival efforts at companies like OpenAI."

To me this sounds like folks who don't know what they are trying to achieve, and coming up with excuses. (I thought there must be a term for this, and asked ChatGPT, and am quite delighted with all the options it gives: Strategic self-handicapping, Pretext paralysis, Deferral delusion, Obstruction illusion, Motivated impossibility – all of these seem to be hallucinated terms but focus on a different aspect.)

But it's true could also go deep with something like GPT's Advanced Voice Mode. This does actually does seem like a big technical lift. It's a change in the product, but besides the technical difficulties of better turn-taking, speech generation, memory, etc., it would otherwise be possible to drop it in to the product. It would change the relationship of Alexa with the household, but it's pretty clearly compelling.

Imagine if that box were able to (somewhat) intelligently answer questions.

I'm sure everyone here has asked Alexa some random type of questions .. sometimes you can get an answer from wikipedia, but mostly it's just "sorry i don't know that", and people settle for a subset of standard voice commands to play music or dim the lights.

Amazon had 10,000 people working on this product for a very long time, and had the lead in market share of a device that people talk to regularly.

It's one of the biggest technology balls to ever be dropped.