Stockfish wasn't being run under ideal conditions, and a video I saw from an IM who does a lot of deep analysis with Stockfish claimed that Stockfish would have drawn 7 or 8 of the 10 sample games under TCEC conditions. The full Bg5 game is
here, this IM claimed that Stockfish's losing move was 17. ... Qg6, which leaves the queen too vulnerable, and that instead Qd8 (preferred by Stockfish with deeper search) would hold the position. I have no idea if that's true or not.
It's not really the point whether AlphaZero is stronger than Stockfish or not though, the points are:
- Stockfish was designed by humans to play chess, AlphaZero is a general purpose neural network that can learn to play many different games
- Stockfish has been honed over tens of thousands of hours of self-play and tens of thousands more hours of programmer work, AlphaZero learnt in 4 hours (albeit with an enormous amount of computing power)
- When you look at AlphaZero's games, there's clear direction to its moves. It's evident that positional understanding is built into it. Games from other engines can often look directionless, with the engines shuffling pieces around and having no idea how to proceed. When AlphaZero drops one of its horrifying positional binds, you can look back in the game and see it patiently constructing the position well in advance. My favourite game is
this one, where Stockfish ends up so locked up that it has to give up its queen to get out of the bind. There, the fantastic move 37. Bd1! reroutes the bishop to the b2-g8 diagonal well in advance of any appearance of usefulness there. Fast forward 10 moves to move 47 and suddenly the bishop is the star performer in the position.