💭 I'm not really at all confident that muzero scales up to highly complex games the way alphazero clearly does given dominion. It might, but highly complex games are obviously a lot harder to simulate accurate, which muzero needs to learn to do itself. All the successes of muzero have been on relatively simple games, because that's all anyone's used board game ai for so far. I worry that without a really, really impressive trained simulator model, it'll make dumb mistakes with frequency since its entire search is based off of insufficient simulation. Needs more thought. I don't like relying on a highly performant and compatible game implementation, but the rules have to come from somewhere