Machines just took aim at video-game development—from the '80s. AIs have been able to learn to play games like Space Invaders  by watching them for a while. But now, Georgia Tech researchers have written a paper describing how AI can actually build the underlying game engine of Super Mario Bros. just by spectating.

The approach, first reported by the Verge, works by analyzing thousands of frames of game play to see what happens as everyone’s favorite mustachioed plumber moves through the game. The AI looks at what changes between one frame and the next, and tries to link cause to effect—what happens when Mario, say, touches a coin, or lands on an evil sentient mushroom (oh, okay, then: a Goomba).

Over time, the researchers say, the AI can build up rules into a rudimentary version of the game engine. The Verge’s James Vincent calls the results “glitchy, but passable” and notes that the tool is limited to simple 2-D platform games like Super Mario Bros. and Mega Man at the moment.

Speaking to the Verge, one of the researchers says that “a future version of this could [analyze] limited domains of reality.” That’s a nice idea, but as we’ve explained before, making sense of the world is one of the biggest challenges facing AI right now—and re-creating Super Mario Bros. is only a very small jump toward cracking it.