Yeah, I just found out about this recently.
For those who don't know MuZero is like AlphaZero, but with even less human input.
In the game of chess, for example, AlphaZero was taught the rules of chess, then left to learn and master the game on its own.
In the newest iteration (MuZero), MuZero is shown the chess board and its pieces, but that's it. It doesn't know the purpose of the game, what a win/loss/draw is, or what the pieces are supposed to do.
It isn't told any of the rules of the game. It's only informed if it plays an illegal move. Or told if the game is over.
It's left to "discover" chess on its own.
This is pretty mind-blowing, if you think about it. Not just in terms of chess, but in the way this technology can be applied to other areas. (Disease research, perhaps?)
https://medium.com/applied-data-science/how-to-build-your-own-muzero-in-python-f77d5718061a
Obviously, the folks at DeepMind would not have a paper in Nature if this didn't work. MuZero took just under a million games against its predecessor AlphaZero in order to reach a similar level of performance in chess or shogi. For Go, it surpassed AlphaZero after only a half-million games. In all three of those cases, MuZero can be considered far superior to any human player.
https://arstechnica.com/science/2020/12/google-develops-an-ai-that-can-learn-both-chess-and-pac-man/