This list by @vkrakovna (and the paper that supplied a bunch of them) is one of my favorite things about machine learning. I'm collecting another unofficial list of unpublished examples.
Like this one by @_joelsimon, who evolved new floorplans for an elementary school. It produced a cavelike warren of bubble-rooms because nobody had told it to do straight walls or windows. joelsimon.net/evo_floorplans… 
And this ongoing thread by @sterlingcrispin whose robot dogs have learned to tap-dance, flop on the ground, do pushups, and even hack the simulation physics to glide - basically, anything but walk.
And these robots by @DougBlank
And the algorithm @citizen_of_now trained to bet on horse races, which discovered that the winning strategy was to place zero bets.
And the video game AI @forgek trained that, when it sensed it was about to lose, would button-mash furiously enough to crash the game, taking down the whole Matrix rather than die.
Another algorithm that would reconfigure the Matrix to avoid bumping into walls. It would end up walking straight while the entire world continuously reformed around it.
And @nickstenning's unicycles, that were rewarded for keeping the seat above the ground, so they learned to hack their simulation physics for flying superpowers
And this example that @smingleigh just posted:
1/2 
1/2 
