That's not what it does. Most implementations are so insecure that it fully exposes your data without your knowledge. But maybe there's one out there that I can actually self-compile that does implement sensible security... I'll tag you when I find one.
Worse: all your data gets sent to the LLM too. With a rented model, unless you use Maple or maybe Venice (I haven't been able to verify their claim though) you are sharing everything that you share with the bot with a 3rd party.
I was reading about homomorphic encryption, isn't that where the ai takes your data scrambled and understands it but never sees it, tech is moving so fast it's scary
I haven't seen any explanation on how we would do tokenization on FHE models if they aren't trained with FHE. Especially I would like to know how something like Anthropic can do this over lattices without sharing their secret key with their customers, haha.
That's not what it does. Most implementations are so insecure that it fully exposes your data without your knowledge. But maybe there's one out there that I can actually self-compile that does implement sensible security... I'll tag you when I find one.
Worse: all your data gets sent to the LLM too. With a rented model, unless you use Maple or maybe Venice (I haven't been able to verify their claim though) you are sharing everything that you share with the bot with a 3rd party.