Can't find the reference now, but in recent interviews the AlphaGo team claimed that one of their next steps would involve training a system without that training database, from scratch (simply by playing lots of games against different versions of itself), and that they estimate that it would be just a bit weaker.