They don't really need OpenAI. There's literally a dozen LLM providers that would be sufficient for what an LLM-powered Alexa would do. Amazon has their own LLM

They don't really need OpenAI. There's literally a dozen LLM providers that would be sufficient for what an LLM-powered Alexa would do. Amazon has their own LLMs too. The underlying LLM technology is hardly the issue, at least now. (I'm sure it seemed like it was a year ago.) 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.