Wish They'd Let This Loose On Speedrunning Quake
The big news in the AI research community is the results that Alpha-Go program is getting against top Go players. It's currently leading 2-0 in a five-game match versus one of the world's top players. The previous results on Go have not been nearly this good.
Go, in case you are not familiar with it, is a Chinese board game that is also quite popular in Korea and Japan. It has long been held as the most difficult challenge remaining among traditional board games for creating AI players, because it has a huge state space that resists any sort of search-based approach that has worked quite well in Chess in the past. The new approach relies on neural networks, and its results are amazing so far.
https://deepmind.com/alpha-go.html
They have also researched the capabilities of the neural net algorithms by having them learn to play classic video games from scratch.
http://robohub.org/artificial-general-intelligence-that-plays-atari-video-games-how-did-deepmind-do-it/
Now we can only wait and hope they start to research actually important stuff like Quake speedrunning with this new technology. Because honestly we should be able to enjoy a few years of godlike AI performance in Quake speedruns before they inevitably create Skynet and we few, lucky enough to survive the nuclear strikes, get killed by rocket-jumping Terminators.