Google Deepmind’s AlphaZero Chess Engine Makes ‘Inhuman’ Knight Sacrifice

Google Deepmind’s AlphaZero Chess Engine Makes ‘Inhuman’ Knight Sacrifice


You know those times
when you’ve gotten into trouble but you never realized
your idea was bad until it was too late? That’s sort of like what happened
to Stockfish in this game. On the white side of the King’s Indian,
Stockfish always liked its position according to the evaluation bar
you will see in this video, but AlphaZero had other ideas in mind. The game starts off in the King’s Indian. And Stockfish chooses to face it
with the Sämisch variation. And this is seen as very solid.
The pawn on f3 defends the pawn on e4. It also prevents any black piece
from going to the g4 square as, well, the pawn defends it. So, 0-0, Be3, c5. All of this is theory.
This is not a pawn blunder; it’s just a temporary sacrifice often when black gets
sufficient compensation and in practice does quite well. So, Nd2, Nc6, and now d5
to attack the knight. And all of these moves make perfect sense. The pawn on c4 was hanging,
so Ng3 protects it, and e6. When you have less space, you tend to want fewer pieces
to remain in the board because, otherwise, your pieces
are stepping on each other toes. So Be2, trades and h5. The knight on g3 is not a great piece. And black can expand on the king side,
kick that knight out of the g3 square, and potentially initiate
a king side attack. 0-0, h4, Nh1, only square, an ugly square, of course,
the knight is in the corner, but it’s gonna quickly get back
in the game via Nf2. Now, Nh7 makes a lot of sense here,
trying to go pawn to f5. You want to chisel away
white’s very good-looking center, and then get some counter-play. Now, AlphaZero played a move
that I would not play in a game, but maybe now I would
that I’ve seen this one: h3; because I would fear that
that pawn has gone too far. The only piece defending it
is this bishop on c8. And if this pawn moves later to g4
and the knight comes to f2, that pawn may be a goner. Also look at the evaluation
on the left side of the board here: 1.17 in white’s favor. So Stockfish, which is the engine that the chess.com server uses,
likes the position. Of course, it analyzes
the position in more depth, but it says that white is better. But AlphaZero doesn’t believe that. g3 was played, Bd7, Rc1. And now b5, a thematic move,
saying, “Please take this pawn.” If you take on b5 with the knight, I just go Rb8, and my rook is coming down
into the position on b2. So just to show a sample variation
of what can happen, if Nxd6, Rxb2. And let’s say you keep trying
to take my pawns with Bxc5. So white is temporally up two pawns, but black has great counter chances
with a move like Ba4. The bishop on e2 is going to be captured
if you move your queen and take on a4, which you have to do
because if you play Qe1, here comes a beautiful shot, Nd3! Now, if you take on d3 which… If you don’t do that,
you just lose all the material, here comes Rg2#, check and mate! So that would just be
absolutely devastating. So Qxa4, just to show a sample line. Rf2, trying to fend off
this rook on the second rank, then Nxf3+ comes. And you see this evaluation bar
sort of swinging here. Black is going to be the one
who is just completely crushing. And now you see, finally,
the evaluation bar continues to sink, and that shows black is the one in charge. So… After b5, no capture, Nf2. But here is sort of the depth that… It seems like black is in danger
of overextending because we’ll see Re8. Still don’t want to take that pawn on b5. And after Kh1, just getting the king
what looks like to safety, b4, Nb1, and black has
some structural issues here. This knight is trying
to quickly plug the c4 square. And from c4, it hits d6 and trades away
black’s best piece, the knight on e5. But check this move out. Bam! Neg4! I was not anticipating this at all
when I was first shown this game because what kind of move is that?
You’re giving up your knight for what looks like just one pawn. But the evaluation says 1.3
in white’s favor right now. Again, that’s silly Stockfish, but look how quickly the dynamic
of the game changes. fxg4, Ne4, all of the sudden,
I’m threatening Nxg3+ over here. And once that pawn is captured
then I can take you e3 bishop with my rook. Right? So… All of the sudden,
lines are opening up in black’s favor. Not to mention that this bishop on g7
has a new lease on the board because the b2 pawn is hanging. So, a trade on e4, Bf4,
coming after the d6 pawn. Qe7 protects, Bf3, and now Bxb2, offering a series of trades
because now, all of the sudden, black has two pawns for the piece,
and this g4 pawn might fall next, so that would be a third pawn. So, Bxd6, taking one pawn back. Qxd6, Bxe4. And we see that black currently has
just one pawn for the knight. But it’s not the quantity of pieces
that AlphaZero is sniffing out here, it’s the quality of them
because look at the play here: Bb5, hanging the rook on f1,
the rook moves. Re8, sort of pinning this bishop,
the bishop has nowhere good to go. And black is just trying to push
this c5 pawn down to the c3 square. When you have a passer,
you want to keep advancing it. Nd2, c4. What did Nd2 do to stop that move?
Absolutely nothing. If you take on c4,
you’re just going to lose very quickly because after Qc5,
my queen is coming to f2. Nd2, Qf2 and you can’t stop me
from sacrificing my rook on the e4 square, in which case you’re going to have
problems on the light squares. It’s going to be check and mating ideas. So Nxc4 doesn’t work, so Nf3 and… It looks like AlphaZero
plotted this out to equality because, again, white is still up material,
so if black isn’t doing something quick, white would just consolidate and win. Qf4. Ignoring the queen trade, Qc5. Actually avoiding it,
instead of ignoring it. Ne5 and now Re7,
calmly defending the f7 square. The pawn from c3 can be pushed
at c2 at a moment’s notice. And white’s king is still not very
happy there, trapped in the corner. So, Qf6, c2, threatening c1=Q and Nxg6, blasting open black’s king,
but this just fizzles out very quickly. AlphaZero can just take it and… d6 is one last-ditch effort,
just trying to… You know, it would confuse a human,
but these are not human. Right? These are engines here… working their magic. So, Bh7+. You could take it either way,
but now it’s just a series of checks, since the game ended
with a million checks. In this part of the game,
we’re just going back and forth until its natural conclusion of a draw. A really fascinating game this one was: sacrificing the knight on the g4 square,
catching me by surprise; perhaps catching the Stockfish
developers by surprise. But AlphaZero really plotted this ahead.
This was something fascinating. I’m getting to the game’s conclusion.
I might as well show you the rest, as it was just a bunch of checks. And the evaluation bar
is just sticking at zero because the king can’t
escape all these checks. So a draw was finally reached. But just to go back and emphasize
one last time for all the viewers here, is that, you know… You get this position,
you know the plan is to go b5, but it’s not clear that a move
like Neg4 will ever work. So, if you don’t capture my pawn on b5,
then, once I get b4 in, this bishop on g7 opens up,
the whole position opens up, and black had just enough play
there to hold the balance. A phenomenal move.
Just to show that on the board once more: Neg4! That’s for the highlight reel:
a full piece sacrifice, get a couple of pawns,
but the counter-play, the momentum gained,
was enough to hold the balance. A really nice showing
by AlphaZero taking advantage of that kind of critical calculation,
that really deep plan that no engine is seeing
several moves in advance.

Leave a Reply

Your email address will not be published. Required fields are marked *