Why Do Box Permissions Complicate Personal AI?

Aaron Levie says the high degree of complexity in permissions, access controls, and data at Box tends to be the rub when training a model around one person's world.

A researcher can imagine having access to everything, so a model trained only on that personal world sounds useful. A lawyer has a tiny access point, the five projects they are working on, because somebody one door over is working on the competitive project for another company. The lawyer cannot have any overlap with what the person next door sees or knows, and not a single document can pass between the two walls. The barriers have to be hard. A model can still be trained just for that one user. What stays unresolved is what happens if every single day the user gets added to or removed from something that adds important context.

Breakthroughs in continual learning will probably resolve the conflict, in his view. Five years ago there was no way to pull the approach off.