Artificial intelligence is no longer satisfied with speed-running "Super Mario Bros." — it's learning how to build the game now, too.
Researchers at Georgia Tech wanted to see if a neural network could reconstruct a video game engine just by watching it being played. They gave the network a catalog of game graphics, some basic rules and about two minutes of footage.
From that, the network was able to break down the footage and use that info to build a system that could more or less emulate the properties of the original game.
Games have long been used to train artificial intelligence to accomplish new feats: It's easier for an AI to grasp the artificial rules of a game than the complex rules of the real world, after all.
But the tech might one day transcend the gaming world. One of the researchers told The Verge: "A future version of this could [analyze] limited domains of reality."