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)
@Byron Lovelace Excuse me...we were both given 20 games in advance by Deep Mind - along with a handful of other broadcasters. We were not told who else was going to receive the games. There was an embargo on showing the games before 19.00 GMT tonight, so neither Anna, nor I, shared information with any other party beforehand, and we had no idea that we would both be reporting on the same story. By chance we both picked the same game to annotate - clearly we have a similar taste and it was after all very beautiful! Here is Anna's excellent recording and I would encourage you to watch it as inevitably we highlight different aspects of the game: ruclips.net/video/nPexHaFL1uo/видео.html
Byron Lovelace you think the game highlights are unique to the commentator?? Haha, and you think a well known chess commentator would have the audacity of commenting on another well known commentator’s video if she had stolen intellectual property? Get real.
To my eyes, there's a sense of unstoppable force and inevitability behind Alpha Zero's moves in this game. So fascinating. While watching this game unfold, I was somehow reminded of Fischer at the height of his career.
Yep, it reminds me of my own "immortal". Classic bishop sacrifice on h7, followed by more sacrifices to lure the black King into my camp until I suddenly realized that I didn't have enough pieces to checkmate the black King, who leisurely went on to butcher my remaining pieces before majestically waddling off back to his side of the board...
@@sam-lz6pi Years ago I made the double bishop sacrifice, had a won game but later I blundered and lost my imortal...A fragment of the game appeared in a book about the Colle...
In addition to being truly extraordinary for non-chess reasons, these games are so beautiful to leave me moved. It's like Shirov and Kasparov had a super human child!
It's obvious Alpha Zero is incredibly strong. However, I really I want to know about its 6 losses, and the manner of those losses. At what point did it start to fail in those games? Also, it would have been interesting to see StockFish's evaluation of the position, when AZ was sacrificing those pawns for what it sees as long term compensation.
Those losses probably occurred early in the 1000 game match. We may find that the majority of its wins occurred much later on. Due to its evolution over so many games we should be able to see how it changed from David Banner into the Incredible Hulk.
as Daniel mentioned, some of the games were played not from the opening chess position, but from a selected middlegame position that is equal or slightly advantageous for one side (but not winning). Since A0 has not trained itself starting from random middlegame positions (and would never get itself in such a position if allowed to play from the beginning), throwing it in the deep end and expecting it to draw or win a random middlegame position would definitely cause some A0 loses. Stockfish on the other hand should be able to better handle being given a random middlegame position as it has all those heuristics built into its eval function that is using a mini-max alpha beta search
No human logic in the form of theory or search optimization, also means no chess dogma! That allows AlfaZero to squeeze through seemingly “crazy” lines and come out better on the other side. It not only expands its horizon with a fraction of the computing power of other engines, but gives great insight on what is perceived to be “correct” by humans and other machines we have “trained” to play like humans. Will having access to such beast change human chess in stylistic terms? Will AlphaZero type of technology steer elite chess towards studying, as Daniel says, “uncompromising” chess? This game is an ode to the power of the bishop and AlphaZero does it by giving up material to prove it: one bishop stays on b2 guarding AND attacking the rival King, while the other keeps black on its toes. 2 hard working bishops with clearly different strategic functions!
Still, a lot of what was perceived to be correct by humans turned out to be correct. Good piece placement is important, open files for rooks is important, etc.
I think what is most interesting about Alphazeros style is that it plays more like Morphy than Carlsen. Morphy also basically had only the chess rules to work with while now in 2018 we have worked out openings up to 25 moves deep. It is fascinating to see how humans developed a certain human way to handle the game and that we (sometimes as a reflex) follow along with well researched principles while Alphazeros "understanding" of positions seems to be much deeper and more free. When any human would give away 4 pawns, how would you be able to hold your nerves like that and casually shove your king to safety, amazing stuff. Thank you again for this coverage!
Nothing is further from truth by the way. Morphy had the best book knowledge of any chess player in his era. Steinitz correctly pointed out, that while Morphy always played book moves, he and some other players(like Zukertort and later Chigorin) had to be caught up by books because they experimented with some cool ideas. Steinitz even said that Morphy never made any opening novelties and mostly followed book to gain advantage over his opponents.
Great stuff. Great game by alpha zero & great analysis by Dani. Can't wait to see the rest 19 games of course analysed by the King himself. Absolutely amazing maneuvering of pieces and use of space by AZ.
I share DK's sentiment about this game. Truly astonishing the cavalier disregard for material in favor of activity. Must be nice being able to run heaps of Monte Carlo simulations of move possibilities. A0 is showing us the weaknesses of Stockfish, and what it takes to actually beat it.
I watched numerous chess analyses videos created by you and I thank you for your incisive analysis, extraordinary enunciation and for exuding a charming elegance as befits the game.
@@PowerPlayChess By the way, GM King, are you by any chance a twin of Nikolaj Waldau? Also thanks to you and another RUclipsr Mordimer my game has improved. I am able to formate plans and strategies that I never delineated earlier leading to wins by middle game.
What an amazing game. I also want to say that the way you present these games is really elegant amd beatiful. Thank you Mr. King (and your team) for your awesome work on this channel! 😘
And now we have stockfish 10!! The difference is that it analyzes much faster than the previous engines. I compared it with an old Houdini 1.5 and the moves were the same, but the dephts is reached in a fraction of a second.
Great video. Congratulations to you and Matthew Sadler on your Alpha Zero windfalls. Very worthy recipients. Intro was inordinately long though. Max 4 minutes I reckon and then bang, straight into it.
Also, I notice a recurring theme in that Alpha Zero will drop some pawns, and use its pawn structure in order to immobilize black's queen-side pieces. Notice that we are on move 25 and the queen rook and bishop haven't been moved and the knight is barely doing anything.
Good commentary Mr. King. What an impressive game! The multiple positional pawn sacrifices performed by AlphaZero, followed by a number apparently harmless moves, are the apotheosis of chess strategy. Chess strategy and piece dynamics principles need to be radically reviewed. I will buy the book of Sadler as soon as it will be published.
dont think human players will play like this, because one suboptimal move and you lose, it is too risky, calculating every tactical variation in an open game is impossible for humans
I don't know about that. No human can defend like Stockfish either. "All" that's needed is a Tal like talent and confidence that this way of playing is correct. To turn our risk aversion on its head, one suboptimal move by the defender and it loses too. Here defender made no suboptimal move I can see and the attacker made a whole lot of "suboptimal moves' that I could see, yet the attacker won. What does it matter what you understand or don't understand chess when checkmate ends the game?
When you say 'neural networks' it reminds me of the famous scene in Wargames (1983) where the supercomputer plays out every possible scenario in a real time game of thermonuclear war between two superpowers, only to conclude that the game ultimately ends in a draw. ....'Greetings Professor Falken. Would you like to play a game of chess?'
i wonder if Alpha Zeros play is based on like 95% tactics also, like chess seems to be these days, or more on principles that it developed itself..for us humans it´s fascinating to see piece activity, development, value of pawns and pieces, king safety and long term compensation ruling over material advantage against such an advanced competition for the whole game and seemingly freeing up realms of sound thinking about chess. It sidesteps somewhat unseen counterplay tactics and rules over the board majestically in complete freedom to manoeuver its pieces and point out blacks weaknesses in all lines..I´m sure it does the same with black too, reaching a win, draw or loss, just playing based on its own general knowledge and creating a chess world in which a chess game is not a draw by force but if it were to be one by them basics Alpha Zero uses but also in which it´s a game that applauds to rich ideas in your play and does not destroy them with crazy-as-20-move-tactical-lines but rebuilds those basics within the same game in different positions and stages. Really wonder too if it does the same with Shogi and Go..
Beautiful. This is the final realization of AI, and it's fascinating that it is not just a further extension of the search tree per se. For one, greedy pawn-grabbers really do get punished by a superior chess intelligence, only, there was not a chess intelligence strong enough to prove that until now. A0 demonstrates that optimal piece placement should trump a pawn deficit. At the same time, pawn-grabbing itself is exposed for what it is -- not a profound middlegame strategy, but just what a tactical monster does when it can't think of anything better to do. SF met its match in an opponent that knows just when to lose material and how. (Or to put it another way, pawns really are to be sacrificed, but only if you know when to do it!).
Thanks for this, Daniel. Very interesting. Watching your analysis, Anna's and Agadmator's is a real education. Would love to hear Judit Polgar's take on this.
Totally agree with Matthew if you're in the top 50 you have to study these games they will spark new ideas at the subconscious level and elevate your own games.
Absolutely amazing! Such an elegance, it feels like if alpha had and organic sense of the board and feels every possibility of the pieces as we feel the possbility of our movements in our surroundings. The white b6 move is just surreal, it would be great to know how back into the game did the AI actually see this.
Really enjoyed the game and presentation. Nimzowitch would be proud. Perfect examples of his teachings of sacrificial blockades and alternation, attacking on both sides of the board. His book examples might be subject to tactical scrutiny, but he had the right ideas.
3:56 I wonder if having 4 TPUs doesn't give Alpha Zero a huge processing advantage. Or maybe the TPUs are only to evaluate the position but they don't help in the searching process
@@klieu90210 I know, but wont the extra hardware power give Alpha Zero unfair advantage? Lets say Stockfish is calculating 50 moves deep and Alpha Zero is calculating 55 moves deep, that would be a huge advantage. Or lets say both are calculating 50 moves deep but Stockfish is analyzing 5 branches and Alpha Zero is analyzing 10 branches that would be also a huge advantage. Even a slightly inferior engine running in a really powerful hardware can outplay a superior one but in a slower hardware. My question is: is there some hardware advantage on Alpha Zero's side? If that were the case, it can be compensated by giving Stockfish more computational power
Neural networks don't calculate nearly as deep as alpha-beta engines in any case. Rather, they try to swing from one known position to a later known position with what it thinks is the best chance of winning. So it'll calculate what position/patterns it can force, at roughly a third the depth but at higher complexity. I'm not sure about A0, but Leela currently can't even properly use 4 TPU, doing best on just 2 TPU. Given that NNs can't even reach A-B depth, I find it fair enough that they're allowed some equipment that the A-B engines can't use properly.
@@klieu90210 Sorry, now I have another question. Where does Alpha Zero stores the positions he likes? Has he a hard drive with a massive database of favorite positions or something like that?
Mr. King, thank you for your brilliant analysis of this fascinating game! I wonder if there's something about the position that tells us a slow-moving attack is in order after a sacrifice--that it isn't necessary to hurry. Surely there are other situations (it's probably even the norm) where it would be a mistake to take too much time and allow the opponent to consolidate with a material advantage.
This is a brilliant game in terms of king position.. AlphaZero's pieces basically said to the king "Just go stand back in the corner and watch what happens"
Imagine a live world championship match between these two engines. Both teams (creators) gathered in an arena with their computers (running on the same hardware) Also of course live analysis from GM`s. It would be great.
For any other chess players who find this interesting, Garry Kasparov wrote a book called “Deep Thinking” that is related to the rise of AI. Thanks for the video! Super interesting game! :)
Have any of the games been replayed at a later date to see how long A0 still plays the same moves after the corresponding SF moves (which I assume would remain the same unless A0 made a change)? Does A0 try different moves in games it has previously lost? I'm guessing it would never lose a game it has previously won and wouldn't have any reason to risk deviating.
Incredible stuff! Those monstrous bishops sitting back & dictating everything! Could you also enlighten us with the match conditions - time control, hardware etc.
Noob question: Isn't the idea of sacrificing a pawn to attack with the rook on the open file pretty basic? I do it sometimes in blitz games. A rook looking at the opponents pawn does more than having the pawn yourself, doesnt it? Especially if your own king is save, far away.
Those days are long over. Super GM + engine vs another engine has been tried. But it ends up being the attempt of the human player to set up an anti-computer (mostly closed) position, provoking strategic mistakes that are beyond the calculation horizon of the engine. Thinking of it, you may have a point in thats exactly the position A0 is different to other engines who seem to need materialistic evaluations.
The current dictum of chess is there are no rules of strategy, everything is decided by concrete considerations. You wonder after this how much of that is because the engines didn't operate on rules just concrete calculations. But if A0 is playing by rules it discovered and doing a better job than Stockfish, perhaps that signals also a change in Grandmaster approaches to chess.
Its a bit more nuanced imo: Traditional engines and ofc also humans do abide by rules of strategy - only they nowadays are less general and as you said, applied or determined by specific situations. Maybe we could say that A0 applies them even more freely - and/or simply has developed different rules.
It always amazes me how well Daniel understands both brute-force engines and the new neural-network driven chess-engines! Are you a computer programmer yourself Daniel ?
Perhaps they could demonstrate a more convincing challenge (for those who doubt these existing 2017 games) by undertaking a correspondence-like match, on the order of something such as performing a few dozen simultaneously-played games of Alpha Zero vs. Stockfish, maybe played out at the rate of say, one half-move per day, and also broadcast the whole event as it occurs, much like the infamous Kasparov vs. The World Game of 1999. - j q t -
AlphaZero played this game like Kasparov of the late 80s and early 90s sacrificing pawns for the initiative. BTW, SF 8 defeated AZ in 6 games, any chance we can see a analysis of any one of those games?
It is as if AlphaZero has developed a sense for future piece activity and coordination. It all falls into place so incredibly well without any kind of rush. It is as if it has no horizon and move numbers do not matter. It is as if AlphaZero sees the connections and ideas instead of the pieces and squares. At least that is my impression. It is as if AlphaZeros pieces are one larger unit that blindly finds its path, while the enemy pieces get stuck all over the place. Also a much simpler thought: somehow Stockfish in its losing games tends to have the combination of Ra8, b7, Bc8, Nd7, e6 which every human knows is terrible in most cases. Somehow this structure seems to be incredibly bad, even if white cannot exploit it straightaway, or black seems to have compensation.
So Alpha Zero scored only 57.45% against a Stockfish 8 (1Gb of hash). Alpha Zero performance rating is +54 Elo over SF8 1Gb, that would probably not be enough to be in the top 3 engines today.
Did you read the full paper (not the preprint that's been out for a year, the full paper that was just released)? The hash size used for this match was 32 GB, not 1 GB.
I think that too many people are focusing on the game, which I also follow, as if this were an ordinary player. Since I have significant knowledge, and since I believe that Hawking and Musk were right, I am really anxious by the self-taught nature of this AI. This particular AI (and its more generalized, even more recent variant MuZero) is not the worrisome thing, albeit it has obvious, potential applications in military logistics, military strategy, etc. The really scary part is how fast these were developed after AlphaGO debuted. We are not creeping up on the goal of human-level intelligence. We are likely to shoot past that goal amazingly soon without even realizing it, if things continue progressing as they have. The early, true AIs will also be narrow and not very competent or threatening, even if they become "superhuman" in intelligence. They will also be harmless, idiot savants at first. Upcoming Threat to Humanity. The scary thing is the fact that computer speed (and thereby, probably eventually AI intelligence) doubles about every year, and will likely double faster when super-intelligent AIs start designing chips, working with quantum computers as co-processors, etc. How fast will our AIs progress to such levels that they become indispensable -- while their utility makes hopeless any attempts to regulate them or retroactively impose restrictions on beings that are smarter than their designers? At first, they may have only base functions, like the reptilian portion of our brain. However, when will they act like Nile crocodiles and react to any threat with aggression? Ever gone skinny dipping with Nile crocodiles? I fear that very soon, before we realize it, we will all be doing the equivalent of skinny dipping with Nile crocodiles, because of how fast AIs will develop by the time that the children born today reach their teens or middle age. Like crocodiles that are raised by humans, AIs may like us for a while. I sure hope that lasts. As the announcer in Jeopardy said about a program that was probably not really an advanced AI long ago, I, for one, welcome our future, AI overlords.
The first GM to learn the Alpha Zero wisdom will be the next World Champion. It seems that piece activity counts above all else and that pawn munching just allows you to die on a full stomach.
Please also note, as I mentioned in the introduction, that AlphaZero was tested in several different matches against Stockfish: with an opening book; without; playing from various standard opening positions AND against a more advanced version of Stockfish.
@@PowerPlayChess Please also note that the more advanced version of Stockfish was not even Stockfish 9. Please also note that A0 had a huge hardware advantage. Please note that A0 only had a +52 elo advantage over Stockfish 8 with a hardware advantage after 1000 games. And please also note that Stockfish 10 is rated +112 elo over Stockfish 8. All hype aside the evidence shows that it is more likely that Stockfish 10 is stronger then A0. I expected more from this channel and not to swallow the Google hype without any critical thinking.
@@MCvanVeen The thing is that adjusting or improving a traditional engine means including the evaluation of human heuristics. Thats exactly the point A0 differs - it makes up its own rules depending on the outcomes of games, not on how humans evaluate positions. Judging from the relatively less ressources A0 has needed vs the decade-long development of traditional chess engines and a resulting comparable strength - its the superior approach.
@@ChessQuizToday Come on! The hole point is not the hole crush(155-6) but the style of playing. Is very possisional strategy and yes i believe that Google can easy win Stockfish 25 (not only 10) if he really want. I don't care about elo, but about style. Is not Rock=5 Queen=9(or10) no! is playing from very activity style and i love this! I am happy to test now A0 vs Stockfish 10 from stoping posting this idiot messege abut elo
![BLACK BAG - Official Trailer [HD] - Only in Theaters March 14](http://i.ytimg.com/vi/Du0Xp8WX_7I/mqdefault.jpg)
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)
Really enjoyed watching your analysis, Mr King! The 36...Re8 line with b6 and Ba6 is so beautiful and instructive! Excellent work.
Love you Anna!!
You are one of the nicest people in the world of chess, Anna.
Just not on the board, because you crushed me once online! It was great. :)
@Byron Lovelace Excuse me...we were both given 20 games in advance by Deep Mind - along with a handful of other broadcasters. We were not told who else was going to receive the games. There was an embargo on showing the games before 19.00 GMT tonight, so neither Anna, nor I, shared information with any other party beforehand, and we had no idea that we would both be reporting on the same story. By chance we both picked the same game to annotate - clearly we have a similar taste and it was after all very beautiful! Here is Anna's excellent recording and I would encourage you to watch it as inevitably we highlight different aspects of the game: ruclips.net/video/nPexHaFL1uo/видео.html
Byron Lovelace you think the game highlights are unique to the commentator?? Haha, and you think a well known chess commentator would have the audacity of commenting on another well known commentator’s video if she had stolen intellectual property? Get real.
To my eyes, there's a sense of unstoppable force and inevitability behind Alpha Zero's moves in this game. So fascinating. While watching this game unfold, I was somehow reminded of Fischer at the height of his career.
Stockfish grabbed 4 pawns and still lost. When Yasser Seirawan will watch this he will have an heart attack.
I play in the same dynamic way as alpha zero. I give up everything but sadly I don't follow up with anything..
Don't we all
Yep, it reminds me of my own "immortal". Classic bishop sacrifice on h7, followed by more sacrifices to lure the black King into my camp until I suddenly realized that I didn't have enough pieces to checkmate the black King, who leisurely went on to butcher my remaining pieces before majestically waddling off back to his side of the board...
At least you've got half the right idea 😁👍
@@sam-lz6pi Years ago I made the double bishop sacrifice, had a won game but later I blundered and lost my imortal...A fragment of the game appeared in a book about the Colle...
I use to "sacrifice" a piece in the last 2 years. What I mean by sacrifice is just giving away a piece for no reason, LOL.
In addition to being truly extraordinary for non-chess reasons, these games are so beautiful to leave me moved.
It's like Shirov and Kasparov had a super human child!
So happy you are covering this Daniel
It's obvious Alpha Zero is incredibly strong. However, I really I want to know about its 6 losses, and the manner of those losses. At what point did it start to fail in those games? Also, it would have been interesting to see StockFish's evaluation of the position, when AZ was sacrificing those pawns for what it sees as long term compensation.
The question is: Did Alpha Zero lose Tal style?
Exactly, after the first 100 games, I thought that AlphaZero was simply invincible.
im pretty sure those were inferior openings where both engine won with their favorable color
Those losses probably occurred early in the 1000 game match. We may find that the majority of its wins occurred much later on. Due to its evolution over so many games we should be able to see how it changed from David Banner into the Incredible Hulk.
as Daniel mentioned, some of the games were played not from the opening chess position, but from a selected middlegame position that is equal or slightly advantageous for one side (but not winning).
Since A0 has not trained itself starting from random middlegame positions (and would never get itself in such a position if allowed to play from the beginning), throwing it in the deep end and expecting it to draw or win a random middlegame position would definitely cause some A0 loses.
Stockfish on the other hand should be able to better handle being given a random middlegame position as it has all those heuristics built into its eval function that is using a mini-max alpha beta search
No human logic in the form of theory or search optimization, also means no chess dogma! That allows AlfaZero to squeeze through seemingly “crazy” lines and come out better on the other side. It not only expands its horizon with a fraction of the computing power of other engines, but gives great insight on what is perceived to be “correct” by humans and other machines we have “trained” to play like humans.
Will having access to such beast change human chess in stylistic terms? Will AlphaZero type of technology steer elite chess towards studying, as Daniel says, “uncompromising” chess?
This game is an ode to the power of the bishop and AlphaZero does it by giving up material to prove it: one bishop stays on b2 guarding AND attacking the rival King, while the other keeps black on its toes. 2 hard working bishops with clearly different strategic functions!
Still, a lot of what was perceived to be correct by humans turned out to be correct. Good piece placement is important, open files for rooks is important, etc.
John Kom for sure! I think the use of the bishops in this game is a perfect examples of good use of fundamentals...on crack! ;)
Fgh
I double dare you to sac 2 pawns for activity and win against stockfish.
Well if you would ever win against stockfish, then it would be this way, I guess.
you have to double quadruple dare at the minimum
I think there is a game where A0 sacked three pawns for the win
I think what is most interesting about Alphazeros style is that it plays more like Morphy than Carlsen. Morphy also basically had only the chess rules to work with while now in 2018 we have worked out openings up to 25 moves deep. It is fascinating to see how humans developed a certain human way to handle the game and that we (sometimes as a reflex) follow along with well researched principles while Alphazeros "understanding" of positions seems to be much deeper and more free. When any human would give away 4 pawns, how would you be able to hold your nerves like that and casually shove your king to safety, amazing stuff. Thank you again for this coverage!
Nothing is further from truth by the way. Morphy had the best book knowledge of any chess player in his era. Steinitz correctly pointed out, that while Morphy always played book moves, he and some other players(like Zukertort and later Chigorin) had to be caught up by books because they experimented with some cool ideas. Steinitz even said that Morphy never made any opening novelties and mostly followed book to gain advantage over his opponents.
"I told you ka1 was the right move.. Makes room for White's pieces" and a smug face... Wow Danny you are the best!
It's called irony!
The famous bishop pair from hell.
BRILLIANT. Both your analysis, Mr. King, and the game.
Great stuff. Great game by alpha zero & great analysis by Dani. Can't wait to see the rest 19 games of course analysed by the King himself. Absolutely amazing maneuvering of pieces and use of space by AZ.
Absolutely stunning stuff! Thanx, Daniel :-D
I share DK's sentiment about this game. Truly astonishing the cavalier disregard for material in favor of activity. Must be nice being able to run heaps of Monte Carlo simulations of move possibilities. A0 is showing us the weaknesses of Stockfish, and what it takes to actually beat it.
Thank you very much for putting together this video.
I watched numerous chess analyses videos created by you and I thank you for your incisive analysis, extraordinary enunciation and for exuding a charming elegance as befits the game.
Thank you!
@@PowerPlayChess By the way, GM King, are you by any chance a twin of Nikolaj Waldau?
Also thanks to you and another RUclipsr Mordimer my game has improved. I am able to formate plans and strategies that I never delineated earlier leading to wins by middle game.
Wow happy to hear you got these games in advance. Nice to see you gwt some appreciation.
What an amazing game. I also want to say that the way you present these games is really elegant amd beatiful. Thank you Mr. King (and your team) for your awesome work on this channel! 😘
Amazing game! Fascinating video. Many thanks
What an incredible display of complete mastery of the game. Such confidence. Frightening.
Stockfish disliked this video...
And now we have stockfish 10!! The difference is that it analyzes much faster than the previous engines. I compared it with an old Houdini 1.5 and the moves were the same, but the dephts is reached in a fraction of a second.
I mean unreal... What a game - really strange
Love this! Excellent way to follow up the World Championship
Great video. Congratulations to you and Matthew Sadler on your Alpha Zero windfalls. Very worthy recipients. Intro was inordinately long though. Max 4 minutes I reckon and then bang, straight into it.
Fantastic subject material! Thanks!
Also, I notice a recurring theme in that Alpha Zero will drop some pawns, and use its pawn structure in order to immobilize black's queen-side pieces. Notice that we are on move 25 and the queen rook and bishop haven't been moved and the knight is barely doing anything.
AlphaZero destroying Stockfish seems like Morphy vs NN. Develop, develop, sac, sac, mate.
Good commentary Mr. King.
What an impressive game! The multiple positional pawn sacrifices performed by AlphaZero, followed by a number apparently harmless moves, are the apotheosis of chess strategy. Chess strategy and piece dynamics principles need to be radically reviewed.
I will buy the book of Sadler as soon as it will be published.
dont think human players will play like this, because one suboptimal move and you lose, it is too risky, calculating every tactical variation in an open game is impossible for humans
I don't know about that. No human can defend like Stockfish either. "All" that's needed is a Tal like talent and confidence that this way of playing is correct. To turn our risk aversion on its head, one suboptimal move by the defender and it loses too. Here defender made no suboptimal move I can see and the attacker made a whole lot of "suboptimal moves' that I could see, yet the attacker won. What does it matter what you understand or don't understand chess when checkmate ends the game?
When you say 'neural networks' it reminds me of the famous scene in Wargames (1983) where the supercomputer plays out every possible scenario in a real time game of thermonuclear war between two superpowers, only to conclude that the game ultimately ends in a draw.
....'Greetings Professor Falken. Would you like to play a game of chess?'
i wonder if Alpha Zeros play is based on like 95% tactics also, like chess seems to be these days, or more on principles that it developed itself..for us humans it´s fascinating to see piece activity, development, value of pawns and pieces, king safety and long term compensation ruling over material advantage against such an advanced competition for the whole game and seemingly freeing up realms of sound thinking about chess. It sidesteps somewhat unseen counterplay tactics and rules over the board majestically in complete freedom to manoeuver its pieces and point out blacks weaknesses in all lines..I´m sure it does the same with black too, reaching a win, draw or loss, just playing based on its own general knowledge and creating a chess world in which a chess game is not a draw by force but if it were to be one by them basics Alpha Zero uses but also in which it´s a game that applauds to rich ideas in your play and does not destroy them with crazy-as-20-move-tactical-lines but rebuilds those basics within the same game in different positions and stages. Really wonder too if it does the same with Shogi and Go..
Beautiful. This is the final realization of AI, and it's fascinating that it is not just a further extension of the search tree per se.
For one, greedy pawn-grabbers really do get punished by a superior chess intelligence, only, there was not a chess intelligence strong enough to prove that until now.
A0 demonstrates that optimal piece placement should trump a pawn deficit.
At the same time, pawn-grabbing itself is exposed for what it is -- not a profound middlegame strategy, but just what a tactical monster does when it can't think of anything better to do. SF met its match in an opponent that knows just when to lose material and how. (Or to put it another way, pawns really are to be sacrificed, but only if you know when to do it!).
Thanks for this, Daniel. Very interesting. Watching your analysis, Anna's and Agadmator's is a real education. Would love to hear Judit Polgar's take on this.
Really enjoy seeing AlphaZero’s games analyzed at such a high level.
Such a beauty, waouhh! Thanks a lot!👍👍👍👍
Glad you are analyzing all of them. They chose the right person!
Those Sniper Bishops were attackers and also the defenders. Amazing.
Totally agree with Matthew if you're in the top 50 you have to study these games they will spark new ideas at the subconscious level and elevate your own games.
I never thought I ll hear the word reinforcement learning from Mr. King. Thanks.
Fascinating stuff thanks Dan. I’ll for sure check out that book.
The human who analyzed the games of the Carlsen and the Caruana at the board!
I love this game too, as the two Bishops are like snipers hidden in the trees picking off Blacks position.
I love this game. Thanks Daniel.
Absolutely amazing! Such an elegance, it feels like if alpha had and organic sense of the board and feels every possibility of the pieces as we feel the possbility of our movements in our surroundings. The white b6 move is just surreal, it would be great to know how back into the game did the AI actually see this.
Really enjoyed the game and presentation. Nimzowitch would be proud. Perfect examples of his teachings of sacrificial blockades and alternation, attacking on both sides of the board. His book examples might be subject to tactical scrutiny, but he had the right ideas.
3:56 I wonder if having 4 TPUs doesn't give Alpha Zero a huge processing advantage. Or maybe the TPUs are only to evaluate the position but they don't help in the searching process
Neural networks absolutely need GPU/TPU to effectively run; CPU alone would cripple them.
@@klieu90210 I know, but wont the extra hardware power give Alpha Zero unfair advantage? Lets say Stockfish is calculating 50 moves deep and Alpha Zero is calculating 55 moves deep, that would be a huge advantage. Or lets say both are calculating 50 moves deep but Stockfish is analyzing 5 branches and Alpha Zero is analyzing 10 branches that would be also a huge advantage. Even a slightly inferior engine running in a really powerful hardware can outplay a superior one but in a slower hardware.
My question is: is there some hardware advantage on Alpha Zero's side? If that were the case, it can be compensated by giving Stockfish more computational power
Neural networks don't calculate nearly as deep as alpha-beta engines in any case. Rather, they try to swing from one known position to a later known position with what it thinks is the best chance of winning. So it'll calculate what position/patterns it can force, at roughly a third the depth but at higher complexity. I'm not sure about A0, but Leela currently can't even properly use 4 TPU, doing best on just 2 TPU.
Given that NNs can't even reach A-B depth, I find it fair enough that they're allowed some equipment that the A-B engines can't use properly.
@@klieu90210 That's very interesting. Thanks!
@@klieu90210 Sorry, now I have another question. Where does Alpha Zero stores the positions he likes? Has he a hard drive with a massive database of favorite positions or something like that?
Wow! I am speechless... what a game.
Just mind-boggling . . .
Mr. King, thank you for your brilliant analysis of this fascinating game! I wonder if there's something about the position that tells us a slow-moving attack is in order after a sacrifice--that it isn't necessary to hurry. Surely there are other situations (it's probably even the norm) where it would be a mistake to take too much time and allow the opponent to consolidate with a material advantage.
i think alphazero takes "controlled aggression" to a new level. thanks for this amazing analysis!
This is a brilliant game in terms of king position.. AlphaZero's pieces basically said to the king "Just go stand back in the corner and watch what happens"
One of the most remarkable games I've ever seen
Imagine a live world championship match between these two engines. Both teams (creators) gathered in an arena with their computers (running on the same hardware) Also of course live analysis from GM`s. It would be great.
For any other chess players who find this interesting, Garry Kasparov wrote a book called “Deep Thinking” that is related to the rise of AI.
Thanks for the video! Super interesting game! :)
Have any of the games been replayed at a later date to see how long A0 still plays the same moves after the corresponding SF moves (which I assume would remain the same unless A0 made a change)? Does A0 try different moves in games it has previously lost? I'm guessing it would never lose a game it has previously won and wouldn't have any reason to risk deviating.
Incredible stuff! Those monstrous bishops sitting back & dictating everything! Could you also enlighten us with the match conditions - time control, hardware etc.
Please see the introduction and the insert with the bullet points!
simply fascinating !
If i played this way the chess.com engine would tell me i blundered >:(
Fascinating. What about 15...d5 blocking the Pd4? Simply 16.a4?
Its the Ron Henley sniper bishops that do all the damage
I swear alpha plays on the natural greed of chess engines
Finally a good level for me after 12 videos
9:42 Carlsen's Queenside?
Noob question: Isn't the idea of sacrificing a pawn to attack with the rook on the open file pretty basic? I do it sometimes in blitz games. A rook looking at the opponents pawn does more than having the pawn yourself, doesnt it? Especially if your own king is save, far away.
I wonder if a super GM has the courage to fight this beast. I would recommend a team of GMs to try...
AZ vs Carlsen + Caruana, would be nice...
Those days are long over. Super GM + engine vs another engine has been tried. But it ends up being the attempt of the human player to set up an anti-computer (mostly closed) position, provoking strategic mistakes that are beyond the calculation horizon of the engine. Thinking of it, you may have a point in thats exactly the position A0 is different to other engines who seem to need materialistic evaluations.
absolutely amazing...
Fantastic positional games by Alfazero
The current dictum of chess is there are no rules of strategy, everything is decided by concrete considerations. You wonder after this how much of that is because the engines didn't operate on rules just concrete calculations. But if A0 is playing by rules it discovered and doing a better job than Stockfish, perhaps that signals also a change in Grandmaster approaches to chess.
Its a bit more nuanced imo: Traditional engines and ofc also humans do abide by rules of strategy - only they nowadays are less general and as you said, applied or determined by specific situations. Maybe we could say that A0 applies them even more freely - and/or simply has developed different rules.
Simply amazing
It always amazes me how well Daniel understands both brute-force engines and the new neural-network driven chess-engines!
Are you a computer programmer yourself Daniel ?
Perhaps they could demonstrate a more convincing challenge (for those who doubt these existing 2017 games) by undertaking a correspondence-like match, on the order of something such as performing a few dozen simultaneously-played games of Alpha Zero vs. Stockfish, maybe played out at the rate of say, one half-move per day, and also broadcast the whole event as it occurs, much like the infamous Kasparov vs. The World Game of 1999. - j q t -
AlphaZero played this game like Kasparov of the late 80s and early 90s sacrificing pawns for the initiative. BTW, SF 8 defeated AZ in 6 games, any chance we can see a analysis of any one of those games?
Superb game !
It is as if AlphaZero has developed a sense for future piece activity and coordination. It all falls into place so incredibly well without any kind of rush. It is as if it has no horizon and move numbers do not matter. It is as if AlphaZero sees the connections and ideas instead of the pieces and squares. At least that is my impression. It is as if AlphaZeros pieces are one larger unit that blindly finds its path, while the enemy pieces get stuck all over the place.
Also a much simpler thought: somehow Stockfish in its losing games tends to have the combination of Ra8, b7, Bc8, Nd7, e6 which every human knows is terrible in most cases. Somehow this structure seems to be incredibly bad, even if white cannot exploit it straightaway, or black seems to have compensation.
Why only 20 games have been published
So Alpha Zero scored only 57.45% against a Stockfish 8 (1Gb of hash). Alpha Zero performance rating is +54 Elo over SF8 1Gb, that would probably not be enough to be in the top 3 engines today.
Did you read the full paper (not the preprint that's been out for a year, the full paper that was just released)? The hash size used for this match was 32 GB, not 1 GB.
Alpha Zero is really an artist, I wonder if it has some sense of beauty and grace rather than just heuristics.
I always wondered how God would play chess... Now I know...
@Daniel : alpha zero games are so eye opening. Please do cover more of these...
More on their way!
I love this game, too.
Game starts at 6:00
Alternatively you can also click through the game yourself, if you want to skip the great commentary ;)
Anyone else play the intro at 2x speed? That was an amazing game.
😍 I love this game
Calsen's queen side? So alpha zero is Carlsen, no wonder its good.
I think that too many people are focusing on the game, which I also follow, as if this were an ordinary player. Since I have significant knowledge, and since I believe that Hawking and Musk were right, I am really anxious by the self-taught nature of this AI.
This particular AI (and its more generalized, even more recent variant MuZero) is not the worrisome thing, albeit it has obvious, potential applications in military logistics, military strategy, etc. The really scary part is how fast these were developed after AlphaGO debuted.
We are not creeping up on the goal of human-level intelligence. We are likely to shoot past that goal amazingly soon without even realizing it, if things continue progressing as they have.
The early, true AIs will also be narrow and not very competent or threatening, even if they become "superhuman" in intelligence. They will also be harmless, idiot savants at first.
Upcoming Threat to Humanity.
The scary thing is the fact that computer speed (and thereby, probably eventually AI intelligence) doubles about every year, and will likely double faster when super-intelligent AIs start designing chips, working with quantum computers as co-processors, etc. How fast will our AIs progress to such levels that they become indispensable -- while their utility makes hopeless any attempts to regulate them or retroactively impose restrictions on beings that are smarter than their designers?
At first, they may have only base functions, like the reptilian portion of our brain. However, when will they act like Nile crocodiles and react to any threat with aggression? Ever gone skinny dipping with Nile crocodiles?
I fear that very soon, before we realize it, we will all be doing the equivalent of skinny dipping with Nile crocodiles, because of how fast AIs will develop by the time that the children born today reach their teens or middle age. Like crocodiles that are raised by humans, AIs may like us for a while. I sure hope that lasts. As the announcer in Jeopardy said about a program that was probably not really an advanced AI long ago, I, for one, welcome our future, AI overlords.
2 pawns up - moves king
gg AlphaZero
The game - 5:52
Could someone post the pgn? I would like to send this game to a friend who doesn't speak english.
The first GM to learn the Alpha Zero wisdom will be the next World Champion.
It seems that piece activity counts above all else and that pawn munching just allows you to die on a full stomach.
Amazing!!!! almost human!!!
If a0 is the best, wait for a1...
Such interesting games more so than human vs human.
Chess at the level: god.
Out of 1000 games, only 161 were decisive? So around 16% What does that say about the future of chess? Is every game a draw with best play?
Funny how the machine is a romantic at heart.
The king is dead ! Long live the king !
With 155 wins, and 6 losses out of 1000 games, AlphaZero scored 57.5%. This is about an advantage of 50-60 ELO.
Please also note, as I mentioned in the introduction, that AlphaZero was tested in several different matches against Stockfish: with an opening book; without; playing from various standard opening positions AND against a more advanced version of Stockfish.
@@PowerPlayChess Please also note that the more advanced version of Stockfish was not even Stockfish 9. Please also note that A0 had a huge hardware advantage. Please note that A0 only had a +52 elo advantage over Stockfish 8 with a hardware advantage after 1000 games. And please also note that Stockfish 10 is rated +112 elo over Stockfish 8. All hype aside the evidence shows that it is more likely that Stockfish 10 is stronger then A0. I expected more from this channel and not to swallow the Google hype without any critical thinking.
@@MCvanVeen The thing is that adjusting or improving a traditional engine means including the evaluation of human heuristics. Thats exactly the point A0 differs - it makes up its own rules depending on the outcomes of games, not on how humans evaluate positions. Judging from the relatively less ressources A0 has needed vs the decade-long development of traditional chess engines and a resulting comparable strength - its the superior approach.
@@ChessQuizToday Come on!
The hole point is not the hole crush(155-6) but the style of playing. Is very possisional strategy and yes i believe that Google can easy win Stockfish 25 (not only 10) if he really want.
I don't care about elo, but about style. Is not Rock=5 Queen=9(or10) no! is playing from very activity style and i love this! I am happy to test now A0 vs Stockfish 10 from stoping posting this idiot messege abut elo
amazing
Alpha Zero should play itself. Those games would be interesting too.
I think it did that when it was learning how to win at Go.
It does play itself lol it's played itself thousands of times, learning from each game
Not the learning ones but an actual game as a real opponent (like against Stockfish). Were any of those played and released?