OP explicitly discusses the second big distinguishing in the opening paragraph of the section titled, 'The Alpha Zero Neural Ne':
“The Alpha Zero algorithm produces better and better expert policies and value functions over time by playing games against itself with accelerated Monte Carlo tree search. The expert policy π and the approximate value function Ŵ are both represented by deep neural networks. In fact, to increase efficiency, Alpha Zero uses one neural network f that takes in the game state and produces both the probabilities over the next move and the approximate state value. (Technically, it takes in the previous eight game states and an indicator telling it whose turn it is.)”
Regarding whether OP touches on temporal-difference learning I am unqualified to say but they do not explicitly mention it. Furthermore I am unqualified to judge how central this technique is to the level of play achieved. However in the DeepMind paper (pg. 20)† that start talking about temporal-difference learning thus:
“Self-play reinforcement learning has previously been applied to the game of Go. NeuroGo[40, 41] used a neural network to represent a value function, using a sophisticated architecture based on Go knowledge regarding connectivity, territory and eyes. This neural network was trained by temporal-difference learning[42] to predict territory in games of self-play, building on prior work[43]. A related approach, RLGO[44], represented the value function instead by a linear combination of features, exhaustively enumerating all 3 × 3 patterns of stones; it was trained by temporal-difference learning to predict the winner in games of self-play. Both NeuroGo and RLGO achieved a weak amateur level of play.”
I'm no expert but this implies to me that it was probably the sum total of all the subtle architectural decisions made by the DeepMind team plus their AI hardware and software platform that made AlphaGo Zero excel.
† https://deepmind.com/documents/119/agz_unformatted_nature.pd...