Nick Bostrom has talked about some of the issues around attempting to control super intelligent AIs, and the closely related issue of how to determine if we are AIs inside a virtual environment (closely related because containment is one approach to controlling AIs that we don't understand)

http://www.techthefuture.com/technology/can-we-control-super...

http://www.nickbostrom.com/

We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. Elon Musk is surely right in saying that AI could be more dangerous than nukes. With the advancement in Machine Learning techniques we are moving to that point.

Elon Musk is over rated. He's just repeating what we already know.