A game in which A0 was required to play a defense it would never choose on its own, featured in Carlsen vs Li Chao, 2015. Excerpts from <Game Changer> are in brackets below.

<[When A0 is forced to play] openings in which it has no central presence as Black (as it likes to have in 1.e4 e5 openings) AlphaZero's evaluation drops significantly, and it feels it has to take immediate concrete action to rectify the situation as quickly as possible.

Whereas Stockfish is a master at absorbing the opponent's pressure, AlphaZero likes to complicate and confuse, and material investment is not an issue as long as it helps tip the play towards the type of play that AlphaZero excels in. In general, I would expect this to be a weakness in AlphaZero's play -- not every worse position should be defended in this way -- but AlphaZero usually gets away with it. It also leads to some great tactical episodes!>

The engines went out of book at move 10, and
in this position

click for larger view

AO immediately sacrificed a pawn with 10....f4

<A totally new idea, and one that I'm sure will be attempted in human games.

It feels like an overreaction to sacrifice a pawn so early in a situation where Black has many counterattacking chances. However -- as so often with AlphaZero's activity concepts -- it proves very difficult to handle.

I imagine the move appeals to AlphaZero as it frees the f5-square for the light-squared bishop and thus gives Black extra possibilities to attack White's king along the light squares. The fact that Black's rook is attacking White's dark-squared bishop after Bxf4 also introduces certain random tactical elements in the position.>

At 39 ply (77 minutes) my SF10 prefers the book move, 10….Nb4 (+0.44), followed by 10….f4 (+0.68) and 10….Qe8 (+0.73)

11.Bxf4 Nb4 12.a3(?) Nc6(!)(?)

<I find this followup even more astounding than the pawn sacrifice. Black often plays ....Nb4 in this line, but the followup is always either ...N4d5 or ...a7-a5 (trying to keep the knight on b4, and daring White to win a piece and open the a-file). AlphaZero sacrifices a pawn and then calmly spends two tempi to induce a small weakening of White's queenside light squares (the b3-square!). >

Incidentally, A0 thought the best continuation for both sides was the hair-raising 12.h4 Be6 13.h5 c5 14.hxg6 Qc7 15.gxh7+ Kh8 16.Bh6 cxd4

click for larger view

13.Nge2 Na5 14.Kb1 Nbc4 15.Qc1 b5

click for larger view

<By now we're not surprised to see this move from AlphaZero. Black offers a pawn to clear an open file pointing towards White's king.

AlphaZero is now up to a 48.7% expected score for Black and is clearly starting to enjoy itself.>

In the next few moves, SF decides to give the pawn back. Sadler doesn't address it, but my SF10 agrees with SF8's decision. 16.Ne4 Rb8 17.b4 (17.Bg5? b4 -0.87, 29 ply) 17....Nc6 18.N2c3 (18.Bg3 a5! -0.76, 25 ply) 18....Nxa3+ 19.Qxa3 Rxf4

click for larger view

<Making use of the 'coincidental' pressure on the bishop on f4 caused by 10....f4. Stockfish's evaluation in the coming phase was 0.00 with AlphaZero's expected score hovering around 49.0% expected score for Black, so both players felt the position was balanced.>

20.Bxb5

click for larger view

20....Nxb4

<This isn't the choice of any of my engines, but I'm not surprised that AlphaZero is looking to clear the material in front of the White king!>

At (white) move 20, my SF10 finds this move at about 30 ply. It prefers White for a long time, but at 38 ply (75 minutes) the evaluation is 0.00.

21.Qxb4 Kh8(!) 22.Kc2 a6 23.Ra1 axb5 24.Ne2 Rf7

click for larger view

<The position is an almighty mess: two knights and a solid centre against two bishops struggling to get free. In any case, AlphaZero can be happy with the result of its pawn sacrifice. The game stays roughly balanced until the final repetition at move 46.>

I like the f5-f4 pawn sac and am very surprised it has not been played before.

It reminded me right away of the c5-c4 pawn sac in the Benoni.

click for larger view

To free the c5 square for a Knight.

When I first went through the game I did not bat an eyelid at 10...f4.

click for larger view

(Looks playable, I probably thought it was theory, despite adopting a black KID set up v everything bar 1.e4 I've never faced 3.f3 before.)

Yup, Nb4, Bf5, Qd5 to get the Knight off c3 Nxd5 and Nxa7 mate. (pie in the sky line but these things seed my imagination.)

Add in a Queen defended piece on f4 opening the door on Rxf4 and Bh6 ideas to pull the Queen away from the Queen side when it may be critical.

After the sac the authors seem surprised at Alpha's follow up with them falling back on what played before. But Nc6 eyeing a5- b3 seems natural.

But I'm very interested in this.

"When A0 is forced to play openings in which it has no central presence as Black (as it likes to have in 1.e4 e5 openings) AlphaZero's evaluation drops significantly, and it feels it has to take immediate concrete action to rectify the situation as quickly as possible."

Lots of potential there. It does not like it's position so goes ape ASAP. Most humans leave it too late to go active.

It's swindling on a grander scale. We are closing the gap. I'm impressed. Looks like I may have to buy the book. (or next time I bump into Mathew I'll see if I can score a freebie)

I said 10 years ago when these things get around to understanding the concept of counter-play and not just sulking when they are losing then we will see some very interesting opening play (10...f4-f5).

I looked at a loss: Stockfish vs AlphaZero, 2018 to see how it wriggled when losing. (grabbed this because it was the longest. Is this game in the book?)

This time I read the posts first. Petrosonic makes an interesting comment Stockfish vs AlphaZero, 2018 (kibitz #6) as asking why here:

click for larger view

(White has just played 100.g2-g4 why does Black not play 100...hxg3.

Never mind that. First look at checks (check all checks) Why not 100...c5+ 101.Bxc5 Be6.

White has to play 102.Rc3 stop 102...Rc4+ winning a piece and after 102.Rc3 Rb2+ 103.Ka3 Ra2+ that is a stonewall perpetual unless WHite wants to get mated.

Why did Alpha not take that. If I saw it within 30 seconds it saw it possibly 20 moves ago.

In the match was it told to shun draws. Surely it cannot think it is winning. Before c5+ It's a pawn down though it's passed pawn has more potential. (WB wrong for the a-pawn.). But leaving the h3 pawn on the board on a dark square to me anyway gives White all the chances and eventaully it fell, the White h-pawn won the game.

Slightly baffled. Maybe Mathew's comment about these things not liking perpetuals kicked in. (are they told not too?)