(A little more) Super Mario World
The above video shows a neural network solving a very simple Super Mario World level. So far, so already done, but this neural network had absolutely no training whatsoever: it is from the first generation of a population which was going to be evolved by a genetic algorithm. But since it's the first generation, no evolution actually took place -- the successful run of the level was generated entirely randomly. The entire process, from generating the initial population of random genomes, through evaluating them (a population of 200), to writing the successful genome to disk, took 23 seconds.*
I'm a curious mixture of pleased, impressed, and embarrassed.
* Concurrency was 5 though. With just one emulator running it would have taken about two minutes.