Someone wrote an open-source version of AlphaZero called Leela. You can donate CPU time to help it learn, or you can play against it at
lczero.org.
Judging by its game against the guy from ChessNetwork, it has a lot to learn:
White to play here. Leela (Black) evaluates its position at 56.9%, where 100% is a forced mate and 50% is a draw. But even to my amateur eye, this position is obviously a disaster for Black. White has a space advantage, the c-file on lockdown, his bishop is vastly better than Black's, Black's rooks are useless, especially the awful e8 one, and the pawn on c6 is a huge liability. I was like "how does it evaluate this positively, isn't that way off the mark?" so I put the position into Stockfish and it has +2.5.
Maybe Leela just needs more time and computing power, but her learning seems to be levelling off and not having learnt about basic things like file control, piece mobility, weaknesses etc suggests to me that there's a problem with the neural network architecture.