THose who are claiming that they are running Machine learning at the edge, think about this:
how on earth can they trust the data coming from it?
We all know that ML is hard, we also know that to get accurate classification requires serious horsepower.
None of this can be run on phone hardware without people noticing.
If you were to design a system for handling reported data, who's hard constraint is _human_ eyeball time, would you run ML at the edge? No.
You'd make it so the "report this message" sent the last n lines and the attachment, and then run the ML optimised for accuracy (not speed or memory) on your hardware that you control and trust explicitly.
I would have thought all of this is obvious, given that facebook is all about avoiding spending money on human moderators. Edge ML is not going to help you do that.
https://twitter.com/WABetaInfo/status/1435221936888483847
The pro publica article is pure click bait
It seems like whatever the AI decides is questionable gets sent to moderators automatically, also revealing the thread history (at least partially) as well.