I recently read the paper, and there are a couple of things you need to keep in mind to understand the scope and how general the result is.
They were using a big cluster to do a brute-force tree search (not brute-force as in exhaustive, but still brute-force as in let's throw lots of hardware at this). According to the paper, this tree search was important in improving the play.
Basically they were using a combination of approaches, like the winners of the Netflix competition a couple of years ago, where each approach in its own was pretty good, but not on the level of Sedol.
The other thing is that this was bootstrapped using a gigantic database of human plays. It's not clear to me that they could have ever achieved what they did without this. Once they trained the neural networks up to the level of an expert player, they could make it play against itself and learn some extra things. But the question is how far this takes you? How much can an AI or a human learn by only playing with itself?
Clearly, it's not yet god-like, since Sedol managed to beat it by a move it wasn't really considering. It's not clear to me how you would improve what they have now, without adding yet another approach, like the Netflix competition where the mixed models got better by the sheer number of them.