I should've mentioned the AI effect isn't my theory, it's just a part inside of my theory haha.

>I wondered why the Gato paper (from what I read of it) didn't try any cross domain exercises. e.g. get a robot arm to play an atari game.

I believe this is being done by something with google, and likely others. I'm not sure why specifically they didn't do it with Gato a while back now, but it is definitely being done with other models.