The paper shows a three-way no-win situation: an AI society cannot be closed off, keep learning forever, and stay perfectly safe for humans all at the same time.
This paper argues that the fastest and safest path to super-smart AI is for humans and AIs to improve together, not for AI to improve alone.