This artificial intelligence has been learning to master Super Mario Bros 1-2 for 17 days
There are some who would argue that those two days could have been spent better. To them, I defensively point out that it could be worse. Right now, on YouTube, you can see an AI learning to play the original Super Mario Bros. Or more precisely, the first two levels of Super Mario Bros – an endeavour it has been working on for 17 days and 18 hours at the time of writing.
You’d think that in that time it would have gotten pretty good at mastering the route necessary to complete these early training levels. You would be mistaken. Watching the live stream for the last 20 minutes – along with 86 other anonymous gluttons for punishment — I’ve seen Mario jump off ledges and walk straight into turtles, but by far the most common death is from natural causes. In this context, “natural causes” refers to a hard reset caused by Mario walking into a wall indefinitely, refusing to jump over the line.
It’s strange because at times it looks like the AI would give the world’s best speedrunners a run for their coins, leaping over piranha plants at reckless angles and falling deftly into the path of rising platforms. Then it goes and spoils it by doing something stupid like walking into the steps at the end of the level indefinitely.
The AI can’t claim it didn’t learn from the best, either. The video is being live-streamed from YouTuber SethBling’s account – and he currently holds the record for speedruns of Super Mario World. In the video below, you can see him introducing the Marl/O AI, and explaining how he trained it on footage of him playing the game.
Eventually, the AI will theoretically become a Mario master, but watching it unfold in real-time is a handy reminder that, while super-smart people such as Elon Musk and Stephen Hawking fear the repercussions of a rampant AI without controls, at least when it comes to Mario our speedrun records look safe for immediate future.
Image: clry2 used under Creative Commons