I'm not saying I'm not impressed, but doing a bit of machine learning research myself I can see how this is possible. However, StarCraft is intentionally hard for human players to control. You have to remember a ton of things, both long term and short term, you have to manage your attention and you have to train physically like a piano player to translate your actions into keyboard/mouse clicks.
The combination of these leads to a steadily growing learning curve the more time you put into the game, which is of course exactly what the game is designed for.
Now you want to test if your neural network (I refuse to call it AI, it's nowhere even close to that) can do strategy, so you give it a much simplified interface to work with. That's OK for this part of research, and it's also OK to pit it against human players to see how it performs, but don't interpret too much in the results. We will talk again when this neural net can move under it's own power in front of a screen and a keyboard and play the game under the same conditions as a human pro player.
Until then, we will have to learn how much processing power the bot needs after training (i.e. to just execute the neural net). This is usually far less then required for training the network, so if some better bots come out for us humans to train against, I wouldn't mind, especially if you can select the version of it. For example you have trouble defending against a certain strategy and can select a bot which implements it well.