Here’s the tl;dr and it’s awesome:

“A system that can simulate itself is to some degree self-aware... We said, “Let’s take all that infrastructure that people have made to help robots learn about the world, and we’re going to turn it inside, on itself.””