The reddit discussion is
https://old.reddit.com/r/MachineLearning/comments/a0nnp7/r_m...
Some skepticism in the comments, since the algorithm assumes you can "return to a previous state" (e.g. with a simulator). So could be interesting for robotics, games..