Yes, Alpha zero calculated 'best' opening moves during its training. That is equivalent to an opening book. Stockfish had to do all this on its own, one minute thinking time per move.
But again, it is impressive that an algorithm can reach a comparable strength to a handcrufted chess engine by just knowing how to move the pieces.
This is machine learning versus the imperative approach, two different paradigmas, and i agree with the people who say machine learning is the future.
As part of the parameters established for the match, SF opening and endgame tablebases were disabled. Also the time control was 1min./per move (max)
Is it bad that I find all of this hilarious? To be honest, though, SF is in trouble. A0 can only get better with time, SF too, but not by much...
Come on, what can a computer do in 4 hours of training calculating only a fraction of the number of moves Stockfish sees per second. It wasn't spending all of its time researching openings. I'm just waiting for the "equal" match to take place. Please don't exaggerate the disadvantages that Stockfish had (#6, for example).
SF took 9 years of programming and tweaking to reach it's current strength while Alpha took 4 hours of self learning. Yes there may be disadvantages to SF during the match but do you really think if played on equal standing, SF would then win easily? I really doubt so. Even if the results turn out to be less spectacular, it's still amazing how machine learning can speed things up. Give alpha 9 years of self-learning and I think SF will be outmatched completely.