It wasn't enough for Google's DeepMind artificial intelligence researchers and developers to create a mental monster capable of decisively defeating a Go legend like Lee Sedol. It wasn't enough that they were already sending chills down the spines of scifi fans with nervous jokes about Skynet and robot overlords. Now they've got their sights set upon conquering the digital side of competitive gaming too, as they jointly announced with Blizzard over BlizzCon.
There's a few problems and potential challenges with that, though: Starcraft 2 is decidedly not Go. Though some basic concepts are agnostic from game to game, like controlling space and marshaling available resources, games like SC2 in particular have elements of ambiguity that makes it an even more of a difficult challenge than Go in some ways, and Go's inherent complexities were legendary as a benchmark.
But Go never had a fog of war.
As described by research scientist Oriol Vinyals, while self-learning AIs like DeepMind are really good at learning how to achieve an optimal end-goal with simple tasks like finding its way through a virtual maze, they can also act like a petulant child at times. If given a sufficiently complex game with a lot of failure states, they can be taught the wrong lesson entirely, and "learn" that the only way to not lose is to not act at all.
That's especially true for games that deliberately hide information from players, like with Starcraft 2, where most of the map and enemy actions are deliberately hidden from you unless you make the effort to unveil them. The ambiguities of a game with fog of war can easily befuddle even the most advanced self-learning AI.
On the other hand, most AI research doesn't get to have access to case studies of 1v1 interactions happening in parallel thousands of times a day at wide ranges of diverse skill levels. Which is where Starcraft 2 comes in. On one hand, we have countless case studies of what not to do, thanks to wood-tier players like you and me. And on the other hand, we have the entire Starcraft 2 pro circuit, staffed by the very best minds in the field of how to extrapolate information from indirect sources like timing, fog of war, when and how they spot a single unit, and turn that into a devastating all-encompassing strategy.
Thanks to the Blizzard/Google collaboration, AI researchers now get to feed their brainchild AI a veritable feast of data via replay analysis. It gets to chew through literal centuries' worth of player interaction data from both the best and worst players.
Effectively, we'll all have a hand in rearing what will hopefully turn out to be the second-most platonically ideal SC2 player ever.
Just behind MVP.