Programmer Seth Bling is no stranger to modifying games. In addition to creating a working Atari 2600 emulator inside of Minecraft that plays ROMs like Donkey Kong and Space Invaders (albeit at 60 frames every four hours), he also injected Flappy Bird code into Super Mario World. It appears he's at it again by training a neural network to play Super Mario Kart.
Called MariFlow, Seth utilized the neural network by feeding it 15 hours of Super Mario Kart videos and directly modifying some of its behavior.
As a primer, a neural network is able to learn things by example. The more data you feed it, the more it learns...similar to how a human brain works (hence the word neural). Seth uses about four layers of computation that allow MariFlow to predict what moves it thinks an actual player might make. In fact, Seth can program it to behave like certain people (he used his father as an example).
MariFlow is technically a recurrent neural network, meaning it's capable of remembering information. Whenever the program would get stuck on walls, Seth would manually "take the wheel" and show the computer what the correct response should be. The program would then learn from his example.
MariFlow has been able to earn gold medals in the Mushroom and Flower Cups and the silver medal in the Star Cup. To be fair, it was done on the lowest difficulty, 50cc, but it's an impressive feat nonetheless. Seth wrote during a Twitch stream of MariFlow that the goal was simply to see if it could get medals in each cup. Mission accomplished.
Seth has actually done something like this before. Back in 2015, he trained a neural network called MarI/O to beat Super Mario World. MariFlow is different from MarI/O in that MariFlow's goal is not to win the game, but rather to predict the controller inputs that Seth would use in any given situation.
Check out the Twitch stream of MariFlow beating the Mushroom Cup and prepare to submit to your future neural net overlords.