I posted one of Greg Brockman's tweets right after the win under the "AI & Robotics News and Discussions" section.
I don't know enough about Starcraft 2 to make an informed bet -- if I knew more about it, I could probably give a good estimate.
The days of humans winning at games that can be fully simulated, are numbered. These training methods are highly scalable.
More impressive, by far, was OpenAI's recent dexterous robot hand work. The system was trained entirely in simulation, and then worked on real world robots. In fact, they first tried to mix real-world training data with simulated data, and it didn't help -- so they went with 100% simulated, and it worked beautifully on real-world robots! The key was doing the right kinds of randomization so that learning in simulation transfers to the real world. They've said that, in the future, they'd like to automate that part, too, so that the types of randomizations are chosen entirely by the machine.
That's the kind of thing that has a real chance of making home robots feasible -- especially given the rate of increase in compute being thrown at the largest AI projects (it grows by 10x every year!), and especially since there are now some pretty realistic home robot environments to train them in. Training probably won't just be "throw the robot in the environment, and see if it fails anywhere along the way"; it's probably going to have to be made up of a hierarchy of things -- but it's still probably doable near-term, in simulation -- and then also the real world, if those randomizations work as well on such a complicated task.
I would guess that mathematical theorem-proving and programming contest-level programming might also -- somehow -- succumb to these methods. This "world" can be simulated; what's missing is a way to generate a set of "natural" things to prove. Proving random things will not do; humans don't prove random things.
But none of this is going to make Turing Test-passing AIs. That's a different beast, and will require a much different set of approaches. BCIs + DL are our best chance; Big Data + DL + (statistical algorithms) might also work, but at a slower pace of development.