Thanks for the pointer to lesswrong. The relevant article on this topic appears to be here: https://www.lesswrong.com/posts/pQYEPitFqztcRvBsS/openai-s-u...

Just glanced at it and it looks pretty good!