If you like my videos do *subscribe* bit.ly/powerplaysubscription and do checkout the supporting options through *Patreon* : bit.ly/patreondanielking or through *PayPal* (links in the description)
One thing I find interesting about this series is it harkens back to the pre-computer age where we no one could be quite sure what the fatal mistake was in any given game or what best play would entail.
I like this game because it's so practical from a human perspective. It isn't some alien idea that sacs a ton of material but just a strong slow attack that squeezes black.
I thought there were at least some losses in all the games published with the full research paper (I think there were more than just those 100 games?). But I could be wrong about that.
@4:25 Here's the thing though, if a Bishop pair is more valuable than a bishop and a knight, then the h5 bishop has a pin on a comparatively more valuable piece on e2. If I'm white in this position, I now feel somewhat trapped into keeping my knight where it is to avoid the exchange and losing the bishop pair, or into playing something like g4 to unpin my knight from my bishop which would weaken my king's pawn structure. If I move my Bishop to something like c2, then black and exchange on f3 and damage my pawn structure. Either way, the white Bishop creates more value for black when it exchanges than white is compensated with by the recapture. How should a lowly 1200 rated scrub proceed as white if I want to keep my bishops and my kingside pawn structure intact? I always see positions like this in my games but rarely see the way forward. Thanks a lot!
I usually try to find a way to up the value of my knight through activity though in some positions g4 works and like in this game you can recapture with the g pawn and as such the bishop is not required to look after the knight
@@valdemarpaulusantreasyanhi7309 " recapture with the g pawn and as such the bishop is not required to look after the knight Right but I went over that, the problem is it breaks your pawn structure which is traditionally seen as a bad thing - though if you're A0 you seem to think it's a good idea because you're a machine that can calculate and remember exact positions you've been in over 1000's of games. The issue is that it is hard to keep the value of the bishops without losing pawn structure. What do you do about that if you want to keep your pawn structure AND maintain the bishop pair AND not have your light squared bishop tethered to the knights' defence?
A storm out of a clear blue sky. Alpha seems to have the antidote to the A17 Nimzo-English. There is more beauty on display in this series than a Miss World contest. I suspect the 6 games Alpha lost must have been from dubious set opening positions. This is like watching Hawk Eye Alpha v Mr Magoo Stockfish.
Thank you for once again a superb video. But who and why does Stockfish change its openings? Why didn't it play 6...a5 this time, when it was the strongest move (for stockfish) in the earlier game?
Stockfish (like all engines) deviates their games when having equally strong moves otherwise an engine vs engine match is going to have the same move sequence every game.
Those amazing Bishops again! Alpha Zero likes to keeps them on the back ranks. It would seem to maximize the range and keep them out of easy striking distance of Knights and Pawns. Combined with the Queen and Rooks, it's like some sort of rocket battery that dominates the field at long range. It seems to favor the idea quite a lot (as much as an old fart can tell). I find it quite incredible that it hasn't been more widely practiced. Or, maybe I'm just seeing the successes in hindsight? Whatever, it's still quite something to what I'm accustomed to. Anyway, thanks so much for your insight and enthusiasm Daniel. Watch out that Magnus and the rest don't become jealous!
In most of these games, seems like SF is playing with 1-2 pieces less (which are pretty much sitting on their home squares). See moves 20-30 in this game. The rook on A8 is as good as not there - when there is a battery forming on its king from the opponents heavy pieces.
What i don't understand about the Grand Master vs Deep Blue matches was if the computer was also allowed to think about moves during the player's turn. Surely this would have been unfair if it was the case. Does anybody know? Or is 1/2 the time just equal to 1 bit for a computer like Alpha Zero?
I only sez it would have been unfair because essentially for real players some element of the oponent's move must essentially be "waiting time" whereas for a computer this would not be the case.
Do you think these victories of AlphaZero demonstrate that these AI are superior in their “understanding” of chess, or is it only a result of their playing style somehow exploits engines' downfalls?
A0 has superior understanding than any other engine because it teaches itself by playing against itself. It's a learning machine that works in the same way the human brain does. It just does it in a more efficient and much faster way. Which human has played millions of games and was able to keep track of all of them to advance his/her understanding? None of course. i don't know how stockfish works exactly but it's basically a brute force machine that evaluates unclear positions based on set indicators. A0 can constantly evolve and change the way it evaluates a position. Stockfish cannot do this.
@@TheRubsi, What do you mean with "understanding". Does it have a concept like strategy or ideas ? Does it rely on blunders ? It is a Reinforcement Learning enginge, so a Markov state machine. But what is a state ? Just concrete position on the board or more general ideas as human player have ?
Alpha Zero doesn't play any official games where everything is even(TCEC). Leela who is very similar to Alpha does and is getting crushed by Stockfish on a regular basis in TCEC. Google is fooling you.
I am wondering what you think A0's ELO would be. StockFish is rated at 3300 but the score of the match is +155 -6 =839. Doesn't that justify a 4000 ELO for A0?
Not at all A0 scored 574.5/1000 which is the expected score of around 50 elo difference. A 400 point difference is close to scoring 100% (always winning) so you couldn't make a claim unless it was winning every game
That bishop was more powerful than the queen. What if Stockfish had played Re7, sacrificing the exchange to get rid of that dangerous dark-squared bishop? It would probably still be losing, but it looks simpler at least.
Are the machines running on equal hardware ? I keep on thinking about the remark of Nakamura that " AlphaZero is virtually running on a supercomputer and Stockfish on a computer that can be compared with my laptop."
i can't answer your question but daniel has said that stockfish actually calculates more moves than A0. So basically A0 calculates less moves but evaluates the positions in a much more precise way.
The paper says SF was running at 60Mn/s and A0 at 60kn/s. On my laptop, SF runs at 2.5Mn/s which is 24x times slower. I would expect similar slow-down for a neural engine LeelaZero, so 2.5kn/s. You know what? Leela gets at most 25n/s (25 nodes, not kilonodes!!!) which is 24000x slower. So how much would I need to spend for a GPU which would be 24x slower than the hardware given to A0? $5000? More? Are such fast GPUs available on laptops at all? Seems like neural engines are only for chess professionals who are willing to spend thousands of $ for hardware. You need a desktop computer with very expensive GPU to get their full strength. On my laptop A0 would lose to Fritz 8 unless they played a correspondence game 7 days per move :) On the contrary, SF on the same laptop would still beat Magnus Carlsen.
Nc5-a4-c5 looks incredibly stupid. The only reasoning I could come up with, is that around the horizon of the engine the evaluation of the position tends to get worse, so any intermezzo pushes the bad late evaluation across the horizon. But in general: why the hell do engines keep doing these stupid maneuvres? There were so many decent alternatives and Stockfish is just like: I will move back and forth, do your worst!
Steinitz rules are rubbish. Like the dogmatic theories on doubled and isolated pawns or king safety. Chess strategy must be rewrote in its fundamentals. Chess is a game whose scope is to checkmate the opponent king: piece dynamics, open lines, centralization, profilaxis, piece restriction are most important element. Not the integrity of pawn structure. Nimzowitsch well understood this one century ago.
This is a good move by AlphaZero. That’s a good move by AlphaZero. Blah, blah, blah blah all the pros of AZ play and never mentioning the downsides. Another miserable video. (An example of con trolling.)
Agreed, he needs to show the games where A0 lost too and focus on some of its weaknesses. Sometimes SF just out calculated it but in certain position SF significantly outplayed it, like in game 23 with the opening book.
If AlphaZero is capable of making a mistake, are humans able to detect this mistake? Thats why all its moves seem good to us, even to Daniel King, who is a grandmaster.
If you like my videos do *subscribe* bit.ly/powerplaysubscription and do checkout the supporting options through *Patreon* : bit.ly/patreondanielking or through *PayPal* (links in the description)
you are the AlphaZero of chess commentators!
thank you very much.
I agree !!!
It is very interesting to see how you talk about AlphaZero, it seems no longer to be a chess engine in your eyes, but rather some higher entity.
Chess teacher: A bishop and Knight are each worth about 3 pawn points.
ALphazero: Hold my beer...
Haha. Well put mate.
That bishop is worth 15 points
One thing I find interesting about this series is it harkens back to the pre-computer age where we no one could be quite sure what the fatal mistake was in any given game or what best play would entail.
Stockfish is already beyond human understanding, and AI is above even that, so to us, it just looks like two Gods playing chess.
AZ really likes the long bishops.
Do you have access to these 6 out of 1000 games that AlphaZero lost to Stockfish? If yes, it'd be awesome to see your analysis of at least one of them
Hey buddy you can download the games here
deepmind.com/research/alphago/alphazero-resources/
@@Phurngirathaana Thanks, but are you sure that there are game there that were lost by AlphaZero?
@@pkacprzak yes there are several
@@Phurngirathaana would you be so kind and point in which of the zip files?
@@pkacprzak in all.pgn 110 games see through the first 45 games
In tcec 100 games the first 30 or so games
Good work Daniel .
Thank you for doing these videos :)
I like this game because it's so practical from a human perspective. It isn't some alien idea that sacs a ton of material but just a strong slow attack that squeezes black.
I prefer the alien chess ideas :) Becos it was typical of mikhail tal
can u explain of the games A0 lost?
That's a nice idea. I wouldn't mind seeing a game to see why and how alphazero loses games.
I thought there were at least some losses in all the games published with the full research paper (I think there were more than just those 100 games?). But I could be wrong about that.
Those games were all played from set positions and alpha zero lost quite badly
@@Phurngirathaana he is from 2017
@Tiwaking Tiwaking the recent paper says "In chess, AlphaZero defeated Stockfish, winning 155 games and losing 6 games out of 1000"
@4:25 Here's the thing though, if a Bishop pair is more valuable than a bishop and a knight, then the h5 bishop has a pin on a comparatively more valuable piece on e2. If I'm white in this position, I now feel somewhat trapped into keeping my knight where it is to avoid the exchange and losing the bishop pair, or into playing something like g4 to unpin my knight from my bishop which would weaken my king's pawn structure. If I move my Bishop to something like c2, then black and exchange on f3 and damage my pawn structure. Either way, the white Bishop creates more value for black when it exchanges than white is compensated with by the recapture.
How should a lowly 1200 rated scrub proceed as white if I want to keep my bishops and my kingside pawn structure intact? I always see positions like this in my games but rarely see the way forward.
Thanks a lot!
I usually try to find a way to up the value of my knight through activity though in some positions g4 works and like in this game you can recapture with the g pawn and as such the bishop is not required to look after the knight
@@valdemarpaulusantreasyanhi7309 " recapture with the g pawn and as such the bishop is not required to look after the knight Right but I went over that, the problem is it breaks your pawn structure which is traditionally seen as a bad thing - though if you're A0 you seem to think it's a good idea because you're a machine that can calculate and remember exact positions you've been in over 1000's of games.
The issue is that it is hard to keep the value of the bishops without losing pawn structure. What do you do about that if you want to keep your pawn structure AND maintain the bishop pair AND not have your light squared bishop tethered to the knights' defence?
Queens and rooks all over the place? Let's bring in our king to help assault the enemy king!
A storm out of a clear blue sky.
Alpha seems to have the antidote to the A17 Nimzo-English.
There is more beauty on display in this series than a Miss World contest.
I suspect the 6 games Alpha lost must have been from dubious set opening positions.
This is like watching Hawk Eye Alpha v Mr Magoo Stockfish.
Can't wait for more alpazero!!
Chess has become more fun! Thank you for very entertaining analasys
Thank you for once again a superb video. But who and why does Stockfish change its openings? Why didn't it play 6...a5 this time, when it was the strongest move (for stockfish) in the earlier game?
Stockfish (like all engines) deviates their games when having equally strong moves otherwise an engine vs engine match is going to have the same move sequence every game.
13:39 Queen E2 threaten h5 gh and rook h6 check mate
Those amazing Bishops again! Alpha Zero likes to keeps them on the back ranks. It would seem to maximize the range and keep them out of easy striking distance of Knights and Pawns. Combined with the Queen and Rooks, it's like some sort of rocket battery that dominates the field at long range. It seems to favor the idea quite a lot (as much as an old fart can tell). I find it quite incredible that it hasn't been more widely practiced. Or, maybe I'm just seeing the successes in hindsight? Whatever, it's still quite something to what I'm accustomed to.
Anyway, thanks so much for your insight and enthusiasm Daniel. Watch out that Magnus and the rest don't become jealous!
In most of these games, seems like SF is playing with 1-2 pieces less (which are pretty much sitting on their home squares). See moves 20-30 in this game. The rook on A8 is as good as not there - when there is a battery forming on its king from the opponents heavy pieces.
How many of the wins were played in the English opening, does anybody know?
What i don't understand about the Grand Master vs Deep Blue matches was if the computer was also allowed to think about moves during the player's turn. Surely this would have been unfair if it was the case. Does anybody know?
Or is 1/2 the time just equal to 1 bit for a computer like Alpha Zero?
I only sez it would have been unfair because essentially for real players some element of the oponent's move must essentially be "waiting time" whereas for a computer this would not be the case.
Great analysis, but please add an analysis of one of the rare Stockfish wins. How to beat AZ?
Do you think these victories of AlphaZero demonstrate that these AI are superior in their “understanding” of chess, or is it only a result of their playing style somehow exploits engines' downfalls?
A0 has superior understanding than any other engine because it teaches itself by playing against itself. It's a learning machine that works in the same way the human brain does. It just does it in a more efficient and much faster way. Which human has played millions of games and was able to keep track of all of them to advance his/her understanding? None of course.
i don't know how stockfish works exactly but it's basically a brute force machine that evaluates unclear positions based on set indicators. A0 can constantly evolve and change the way it evaluates a position. Stockfish cannot do this.
@@TheRubsi, What do you mean with "understanding". Does it have a concept like strategy or ideas ? Does it rely on blunders ? It is a Reinforcement Learning enginge, so a Markov state machine. But what is a state ? Just concrete position on the board or more general ideas as human player have ?
Alpha Zero doesn't play any official games where everything is even(TCEC). Leela who is very similar to Alpha does and is getting crushed by Stockfish on a regular basis in TCEC. Google is fooling you.
alpha zero plays anti computer style against stockfish !
It's really hard to see the moves as there are no animations of them moving
You must have noticed that a yellow arrow indicates where the piece has come from
@@PowerPlayChess yes I can see that but I'm watching on my phone and it's not as easy to see
Have you tried to flip the phone so you watch it on the whole screen?
I am wondering what you think A0's ELO would be. StockFish is rated at 3300 but the score of the match is +155 -6 =839. Doesn't that justify a 4000 ELO for A0?
Not at all A0 scored 574.5/1000 which is the expected score of around 50 elo difference. A 400 point difference is close to scoring 100% (always winning) so you couldn't make a claim unless it was winning every game
at this level 100 elo equals like 500 elo. SF 10 is like 3400 A0 is prob 3500-3600 rated
I would give AZ 10,000 points. Whatta heck, it doesn't lose anyway.
@@YevgeniyShcherbakov SF crushed it bro
@@chrisiver8506 I just want to see a match where SF is fully maxed out.
I find AlphaZero's techniques in exploiting its unopposed bishop really educational
What is nice is that AlphaZero's plan wasn't super complicated.
That bishop was more powerful than the queen. What if Stockfish had played Re7, sacrificing the exchange to get rid of that dangerous dark-squared bishop? It would probably still be losing, but it looks simpler at least.
ty
Did AZ lose 6 games? Where are they?
Are the machines running on equal hardware ? I keep on thinking about the remark of Nakamura that " AlphaZero is virtually running on a supercomputer and Stockfish on a computer that can be compared with my laptop."
i can't answer your question but daniel has said that stockfish actually calculates more moves than A0. So basically A0 calculates less moves but evaluates the positions in a much more precise way.
From the original paper, Stockfish looked at a thousand times more positions than A0, so A0's positional evaluation must be way better.
The paper says SF was running at 60Mn/s and A0 at 60kn/s.
On my laptop, SF runs at 2.5Mn/s which is 24x times slower. I would expect similar slow-down for a neural engine LeelaZero, so 2.5kn/s. You know what? Leela gets at most 25n/s (25 nodes, not kilonodes!!!) which is 24000x slower.
So how much would I need to spend for a GPU which would be 24x slower than the hardware given to A0? $5000? More?
Are such fast GPUs available on laptops at all?
Seems like neural engines are only for chess professionals who are willing to spend thousands of $ for hardware. You need a desktop computer with very expensive GPU to get their full strength.
On my laptop A0 would lose to Fritz 8 unless they played a correspondence game 7 days per move :)
On the contrary, SF on the same laptop would still beat Magnus Carlsen.
awesome!
Nc5-a4-c5 looks incredibly stupid. The only reasoning I could come up with, is that around the horizon of the engine the evaluation of the position tends to get worse, so any intermezzo pushes the bad late evaluation across the horizon. But in general: why the hell do engines keep doing these stupid maneuvres? There were so many decent alternatives and Stockfish is just like: I will move back and forth, do your worst!
mas mas mas
Beautiful.
Sadly more enjoyable and instructive than the FIDE World Chess Championship.
Is human chess at an end?
Steinitz rules are rubbish. Like the dogmatic theories on doubled and isolated pawns or king safety. Chess strategy must be rewrote in its fundamentals. Chess is a game whose scope is to checkmate the opponent king: piece dynamics, open lines, centralization, profilaxis, piece restriction are most important element. Not the integrity of pawn structure.
Nimzowitsch well understood this one century ago.
use bitcoin wallet donation! why paypal or pathreon? :S
stop anthropomorphising daniel it's a computer program and as such it's as interesting as lewis hamilton's car.
This is a good move by AlphaZero.
That’s a good move by AlphaZero.
Blah, blah, blah blah all the pros of AZ play and never mentioning the downsides. Another miserable video.
(An example of con trolling.)
Agreed, he needs to show the games where A0 lost too and focus on some of its weaknesses. Sometimes SF just out calculated it but in certain position SF significantly outplayed it, like in game 23 with the opening book.
If AlphaZero is capable of making a mistake, are humans able to detect this mistake? Thats why all its moves seem good to us, even to Daniel King, who is a grandmaster.
@@chrisiver8506 AlphaZero only lost 6 games out of 1000 so thats literally looking for needles in a haystack (it won more than a hundred vs stockfish)
@@whisperwalkful I'm just glad it actually lost a few games, shows that its not a god lol
@@chrisiver8506 Agree but those games were from set opening positions like the French etc which A0 dislikes strongly