“Based on a rule of thumb in Go, scaling AlphaGo's raw policy net from 3,000 Elo to AlphaGo Zero's 5,200 Elo would require a 100,000x increase in model size and training.”