The Grand Theft Auto video game series, despite ostensibly being about stealing and driving cars, isn't exactly known for precision driving dynamics and realistic scenarios. Anyone that's ever played the game can attest to it. That said, it does put you in more real-world scenarios than games like Forza Motorsport or Gran Turismo. Charles, a convolutional neural network, is learning to drive through deep learning. The entire thing is being streamed live on Twitch. Based on what we've seen, the machine uprising will involve lots of head on collisions and trampled pedestrians.
Watch live video from sentdex on www.twitch.tv
In typical GTA fashion, Charles simply goes and goes, leaving destruction and devastation in its path. According to the description, Charles "learns and takes all actions based on single frames at a time, and bases his decisions on just pixel data." It can only see on screen exactly what we see, meaning it can't predict the game's reactions to its driving. The creator hopes that in the future, Charles will have some sort of memory that actually improves its skills.
Charles is driving in the GTA universe, meaning that killing pedestrians has a tendency to irk police. We witnessed a police chase and a shootout that ended with Charles stuck between a wall and a garbage truck. When that happens, the car resets and Charles goes about his merry way. Watch it for a few. It's a reminder that, as virtual Gods, we create AI in our own image.