That's why I say: if it works in actuality the way they present it in theory, I'm fine with it.
Even if they used half the techniques they described in this post it would already be a huge privacy improvement over the way people are using ai today.
I use my own llms on ollama.ai . 100% private and open source.