A quick YouTube search reveals countless Super Mario World speedruns by some of the game’s most skilled players. And MarI/O can keep up with the best of them, even though it’s actually just a piece of advanced learning software made up of virtual neural networks.
Created by Seth Bling, a developer who also happens to be an insanely skilled Super Mario World speedrunner, MarI/O works a lot like the human brain does when it comes to learning. Instead of starting with a crash course in how Super Mario World is played, what everything on-screen means, or what perils need to be avoided, it had to learn them all by scratch using trial and error. Seth calls it neural evolution, and the biggest difference between humans and the software is that no controllers were thrown across the room by the AI while it built up its skillset.
And instead of needing weeks to master the nuances of Super Mario World, MarI/O only required about 24 hours to get to the point where it could breeze through a level without dying, and without the need for a single power-up. Imagine how stressful playing would be without using even a mushroom.
The source code for MarI/O has been posted here if you want to try it for yourself, because it should work just as well for other side-scrolling games too. If there’s a level on Sonic the Hedgehog you’ve never been able to beat, this might finally be your key to victory.