Oh man, it's like BOINC. (In fact, perhaps one could incorporate it into BOINC.) I wonder if, after a while, the most powerful ML platforms in the world will be volunteer-run distributed systems. That would be an interesting situation.

One interesting thing about the 'zero' approach is that it learns through self-play. So you can use a database to bootstrap the learning, but in principle you'd only need the game rules and enough training time.