Part of the reason people are so excited about recent Machine Learning milestones is that AlphaGo demonstrated a reproducible method of managing mathematical and computational intractability.
Go is interesting because it's impossible to solve. It cannot be brute-forced no matter how fast processors get. Go is so complex humans had failed to produce AI that could win against a skilled human player. The fact that a computer could teach itself to do something humans couldn't teach it, and something with a complexity analogous to nature to boot, is pretty extraordinary.
Combinatorial games in particular are useful because, unlike nature where it may be impossible to track or even be aware of every variable, intractability can be generated out of a simple set of elements and rules, and outcomes can be definitively evaluated.
As proof-of-concepts for methods go, AlphaGo seems like a pretty strong one. It allows us to definitively say "Machine Learning works", puts a lot of emphasis on the field, and raises confidence on extending the method to real world problems.
Beyond that, it suggests a feedback loop in which programs can improve at at improving, unrestricted by human limitations. Increase in processing power is bounded by physical limitations, but algorithms are not.