Robin Baumgarten, a PhD student at Imperial College, London, author of the AI Panic blog, and fellow Friendly AI supporter, recently got some nice blog coverage for creating an AI (script, really) that plays Super Mario more effectively than any human. Check it:

At this point I must brag that I have beaten Lost Levels. The tactics that the script uses can actually work pretty well — in a lot of the harder levels, running semi-blindly seems to work better than taking it slow and easy, which just puts you at greater risk of being attacked. I wonder — which video game will get solved next? Many of them seem trivially easy, but platformers (like Mario) seem relatively challenging, from an AI perspective.

Congrats, Robin! I am reminded of the Black Belt Bayesian post “Speedrunning Through Life”. Superintelligences will speedrun through real life problem-solving, analysis, and mediation in the same way that Robin’s AI speedruns through Mario. Though the real world is more complex than a game, the concept is fundamentally the same. I find it hilarious that there are humans who actually believe that their problem-solving acumen and speed is about as good as anyone could get, or that the real world is anything but a highly complex video game where you can feel pain and death is final.