We always say we're not going to talk about the robot apocalypse anymore, because it's sort of a silly trope that, surely, won't ever actually come to pass. And then, computer scientists go and do something like this...
According to John Timmer for Ars Technica, computer scientists not only taught a computer how to pretend to be a sniper hunting prey, they taught it how to teach itself to get better at its task. AND. They taught their new expert-hunting-sniper-bot how to teach other computers to develop this skill more quickly.
Combine this with the other ongoing project to teach robots how to mind meld with each other, and we're basically building the Geth, or those things from the Matrix, or whatever.
For StarCraft, the game was set up so that, using cover and firing from a distance, the human could take out the Zerg before it was killed. To keep the human from just staying in hiding, the player's score went down over time—the longer it took for the game to be completed, the worse the score. The algorithm evaluated a total of six factors, like distance and relative health, before choosing its next actions.
When self-teaching, the computers would start off avoiding the Zerg. They'd then end up doing a sort of inverse-Zerg-rush, quickly getting themselves killed before their entire score evaporated. Finally, after about 200 training runs, the computers started to get the idea of sniping and began killing the Zerg successfully, though performance was still pretty erratic.
When computers had to figure out how to play this modded version of StarCraft by themselves, they struggled through a series of failed strategies. When a computer that (who?) had already figured out the sniping strategies was paired up with a newbie, the learning curve was cut down drastically.
After talking to the researchers, Timmer says this isn't the beginning of the robot apocalypse. But, really, who's to say he isn't a cylon.