Use genetic algorithms to train flappy bird
Train a neural network using genetic evolution to play flappy bird! For a tutorial on how to get started with your own genetic algorithms, take a look here.
Initially all the models would do the "same" wrong thing. So they would all die out quickly.
After some time however, they would start to show more variation, but still perform the wrong moves. This gives us a spread of flappy birds throughout the screen (lengthwise).
After a bit of training (~1hr) the spread decreases and is more concentrated at the height where there is a hole in the pipes. They start performing a lot better since they now understand when to flap and when not to.
The model with the maximum fitness can be considered as a trained model and it would perform much better than the average human.