The article buries the lede on the only point that matters with these "AI" hardware devices. They need to solve a problem their customers have, and all the devices these companies have released so far don't do anything that a smartphone can't easily do.
An "AI Hardware" device that needs a network connection, and makes API calls to the mother ship, in order to accomplish anything, is not interesting. It's a Raspberry Pi.
On the other hand, a device that could work offline is interesting. One that could work in a zombie apocalypse is even more interesting. Especially if it was solar powered and contained the knowledge needed to rebuild society.
> An "AI Hardware" device that needs a network connection, and makes API calls to the mother ship, in order to accomplish anything, is not interesting. It's a Raspberry Pi.
Kind of an unnecessary dig at the Raspberry Pi, no? Modern Pis and SBCs in general are good at lots of things. I use mine for self-hosting some apps I use, and I've definitely seen them used in little compute clusters for AI inference.
> On the other hand, a device that could work offline is interesting. One that could work in a zombie apocalypse is even more interesting. Especially if it was solar powered and contained the knowledge needed to rebuild society.
This is kind of interesting in an abstract sense; it's fun to imagine burying a solar-powered oracle in a hardshell case in a bunker somewhere so that some hypothetical person can use it to bootstrap civilization after the end, but that's all it really is: hypothetical. Fun to imagine. A project for hackers or maybe a non-profit. It certainly fails the "toothbrush test" mentioned in the article; no one will be consulting their doomsday box once or twice a day (absent a doomsday, anyway).
If I can be really reductionist for a second, I think there's a lot of AI cart-before-horse happening with these hardware products. Smartphones changed the world a decade and a half ago because they took something that people wanted--the internet, but mobile--and finally made it work. Since then they've dramatically changed the landscape of the internet and social media etc, but the idea--that people already had the internet but wanted to interact with it in a different way--should probably be the foundation for how we think about AI hardware products. What can they do for people better than what they already have? We should not need the benefit of hindsight to see why something like the Human AI pin, that doesn't really do anything and does it badly, failed.
RPis just aren't interesting. It's a full fat computer on a small board that does everything a normal computer does, just small.
PCs just aren't interesting because they're all fundamentally the same thing and are capable of the same set of tasks.
RPis aren't interesting because 99.9999% of projects they're put in are better served by a microcontroller and not an entire linux system. If not just a 555. It's boring to throw an entire linux computer into a project. You've utterly given up on the hardware and have assumed you can do everything with software.
I don't think that passes the toothbrush test. I don't need to survive a zombie apocalypse every morning.
Your market is going to be doomsday preppers. Can you imagine starting a business in that market? I bet the trade shows are filled with bunker developers, underground infrastructure dealers and arms dealers.
I'm imagining a stratified market with two distinct customer personas - very rich and paranoid, and very poor paranoid.
Have fun!
What problem? All these AI devices seem like a solution in search of a problem.
The problem is that consumers aren't paying enough for AI nor are they providing enough sweet, sweet monetizable data to mine.
For me the interesting part was that OpenAI's only viable option for a mobile OS is Android, while Google is their main competitor.
Why is that interesting in why would it matter. Many competitors to Google use Android on their devices. Why would this be different?
Well, I wouldn't put it past that. It could literally feed Android into an llm context window and probably redesign it to be even better and it already is.
Android still has weird laggy jumps and just is not that smooth. Even on the new pixel devices.
Is this... a joke?
Which part are you referring to? Creating a new OS from an LLM or that Android still isn't smooth in 2025?
Feeding the 250M lines of Android source code into an LLM's context window. From your response, though, I'm guessing you were serious about that.
Just wait until we can just feed a serialization of our "coworker's" mind into the context and get an improved coworker who has more practical skill in applying LLMs!
Yeah, that's the kind of things monopolies create.