--------------------------------------------
 AI plays itself and exploits bugs
 2019.11.11 17:52:35 CET
 --------------------------------------------

 A colleague shared with me a link of a video presentation based on
 paper written about AI agents playing hide and seek [0]. The amazi
 aspect of the experiments is how the AI agents exploited bugs or
 unexpected features of the game and physics engine to win. If you
 haven't watched it, please do.

 I'm on the same side of the AI debate as Elon Musk and others. I
 think we're racing to develop better and better AI agents and will
 soon be surprised about the unexpected side-effects which might be
 disastrous to humans.

 The system won't be built to cause destruction. It will find a
 shortcut towards its harmless goal - a shortcut not envisioned by 
 developers. The shortcut will have unplanned side-effects.

 That will be an amazing post-mortem to write and study. Hopefully
 humans will be around to do so.

 [0] https://youtu.be/Lu56xVlZ40M