This isn’t a joke, though it almost seems like one. It uses Llama 3.1, and supposedly the conversation data stays on the device and gets forgotten over time (through what the founder calls a rolling “context window”).
The implementation is interesting, and you can see the founder talking about earlier prototypes and project goals in interviews from several months ago.
iOS only, for now.
All-on-device AIs that could run on an iPhone would be terrible. Its sending tokens somewhere I guaran-fucking-tee.
false advertising law: hello?
Anyone can test if it’s sending with a firewall. If it’s not connected to the internet, it ain’t sending. Don’t forget that iPhone chips have been the Moore’s law executors.
Do you even have a slight idea how processing intense even 5 year old LLM models are?
And iPhones aren’t magically immune to thermodynamic.
The FAQ says that it requires an Internet connection.
It also mentions e2ee, which isn’t too reassuring when one of the ends is their servers.
Exactly, there are ways to make the tokens unreadable even in a server hosted LLM but I know for a fact that’s not what’s going to happen here.
And I fully expect all of our engagement data to be used in the 2028 election to target us.
This was my big concern as well. E2EE only matters if you control each end. That’s why I’ll let the YouTubers and security analysts dissect it first.
Check my edit for instructions on how to build your own. It’s even called “Friend,” so it’s probably the same thing tweaked for a different LLM.