From the perspective of AI and machine learning research, Leela Zero’s distributed training through a big population with generations and competitions of various networks, makes it more akin to “model-based evolution strategies” - agents evolved through variations and compete to survive over many generations. Many generations of LZ networks just spread out to seek a wider search space where the immediate training goal is not clearly set. A network can live as long as it doesn’t lose too badly, and a lot of the old networks got “achieved” and became self-playing sources for generating training games to smaller but more efficient networks. Leela Zero needs a big population to search wider and wider space, but the advance in play wouldn’t grow accordingly, and the distributed training can only go so far where hardware and people interested in helping will be limited. The search space will just keep growing without end.
KataGo on the other hand setup more immediate/intermediary goals, like how many stones to win, etc, not just win or lose, make it to seek a much narrower and smaller search space for training so it would grow much faster compared to LZ. It also makes the process more human-like with aggression to win (and why it play black better IMO comparably). I think KataGo is a big step for Go AI that is tailored for it, unlike Leela Zero (similar structure to Alpha Zero) which is meant to be more general and can be adapted to other games and problems. In the future, if KataGo like to adapt evolutionary strategies as part of the training, it should have more variety in them if we setup different goals and parameters to make them diversify which will be a big plus for more interesting play between different populations.