Yep, much better business model than Google/Meta but they still need the data to make their expensive products work well for everyone. ML/data is the only way to do that at scale.

Actually, It looks like it’s a public protocol: https://datatracker.ietf.org/doc/draft-ietf-ppm-dap/

This paper describes it in more details. It’s using zero knowledge proofs to reject invalid vectors without seeing them: https://arxiv.org/abs/2307.15017

Differential privacy is pretty old and as far as I know nobody really broke it, so if they didn’t put any backdoors in it, this looks legit.

They should open source all the code. And maybe use this to fix the damn keyboard:-)