My theory is that AZ has an adjusted flexable piece value system, for instance knowing when a bad bishop is worth 3.1 or a good bishop worth 2.9 due to board synergies or like in the end using the rook to force B's bishop down to 2.7 which in turn boosted his knight to 3.3 leading it to becoming pacman on the pawns.. as arbitrary number examples ofcourse.
A0 perfected trash-talking after 14 hours of self-improvement. The engineers working with A0 develop depression after as little as 0.8 minutes after entering.
I Always espected that Go was sooner outanalysed by the computer then chess. Because the rules are more simple and you have only one piece. But it was not the case because of the sise of the board.
Just look at how active that king is. It makes sense that the brute force methods never exploited kings to their full potential as, given their limited movement range of 1, it's computationally expensive to calculate a string of king movements. The reliance on castling in traditional techniques and brute force algorithms makes sense. Tucking a difficult piece to think about away in the corner in a safe place makes a lot of sense. Alpha 0 appears to be able to assess new and safe 'houses' for its king though, moving it with an entourage throughout the board. Consequently, there's always a few moments of tension in the mid game where the king provides a vital role, such as the third attacker in this game. *skulks off to spend hours playing with kings*
Good insight! I was noting something similar as well. The black position looked a lot more classic. There is also some good wisdom to this: Don't put you king with his back against the wall, leave a path of retreat instead!
This is what I love most about the more 'intuitive' programs like alpha zero. I'm not sure if you have experience with Go, but you see something very similar where alpha zero playing that game. Basically, the 'traditional' positions that were often regarded as good are falling in favor of things that just make sense when you think about it for a second. Much like you said: don't put your king with his back against the wall. And yet in thousands of years of humans playing chess, they never quite figured that out. So simple, so elegant, so brilliant. I can't wait for the next generation of grandmasters who learn against these machines.
Alpha seems to hate stockfish's light squared bishop even more. This bishop has consistently remained irrelevant in the French defense games between them.
AlphaZero has this phenomenal capacity to control the black's bishops (both light and dark), and seems to favour openings that support that kind of play. The central structure combined with the knight on f3 for the majority of this game keeps the dark square bishop locked out of any useful squares kingside, freeing up the white dark square bishop to take squares that support the central pawn structure while keeping tabs on that neutered black bishop. Other games, AlphaZero will lock pawns in places that achieve these same goals, and in some of the other openings will use the knights to make sure that bishop lanes get clogged.
I would go further, and say that A0 hates enemy pieces :P The Qh8 in one Queen`s Indian was to me the most striking example. But sure, in the French he preys on the weak : "One piece bad, game is bad" , they say ;)
After decades of chess engines that take a more tactical playstyle, NN approaches now seems to finally be swinging the pendulum towards positional technique. Just like we've used them to recognize a face in a crowd, or a word in a sentence, they're now giving our chess engines a tool to look at the board and say, "oh! this is a more advantageous shape". These French Defense games are such a great example of that, because a positional perspective understands the true cost of a deactivated piece (i.e. Black's light-square bishop in this opening) - and so, even when the material gain may be 50 moves out, puts in the work to create a board structure that keeps that piece from ever coming to life. Cool stuff.
It's because AlphaZero learned chess by playing itself with only the basic rules of chess, so it plays chess with a different perspective than a typical engines, after seeing many A0 matches it is clear that A0 prefers mobility and space, it will often sacrifice pieces to obtain more mobility and space for its heavier pieces, A0 looks the chess board with a very deep positional understanding, will often value initiative over safe positions. Its playstyle is very human, a super human playstyle I would say.
Back when we all used Fritz, I always felt happy when 1e4 got met with the French defence. I knew that, at least on my laptop, Fritz didn't get the limitations of the FD spacewise and I always won. Analyzing those games with that same Fritz also was fun, because only when it was really really really losing it thought it was losing.. And now A0 is there to demonstrate that the old way of programming a chess engine is just that: the old way. A0 is a human player, beating his computer opponents in closed positions, understanding the value of the positions as a whole.. I find this fascinating and I want to know more! (Skips to the next video...)
"Stockfish, I was able to model your evaluation function as early as move four. This allowed me to forecast your moves and basically clean up the floor with you. I suggest you learn SQL and find a new career in the business sector."
Jerry, I've learned from ya! Best narrator out there, it's good that you spend time on common moves and theories! As the other site I watch, he just rushes through everything and gets to the part he wants and not smelling the flowers along the way. Great Job, Jerry!!
The version of stockfish that was used in the match was professionally handicapped, meaning it was an older, inferior version and it did not have access to books. This did not allow it to operate at its full potential. AlphaZero had basically built itself its own internal book. In addition to this, it had to play at 1 move per minute. It is, however, not fair to say that without the constraints the result would have changed much, as AlphaZero is rated around 100 ELO above the best version of Stockfish anyway. It's good to keep the details in mind rather than just looking at the results though :D
I'm absolutely sure GMs are reviewing "alpha-variations" from well known opening theory to see exactly why alpha zero chose to sac a pawn or move a queen twice early.
Well in the Queen's Indian defense Alpha Zero does tend to like the line g3 Bb7 Bg2 Be7 O-O O-O d5 exd5 Nh4!!, this line I can't find any consensus on if it has a name or much in the form of theory, the wikipedia page only briefly mentions the Nh4 line under the old mainline, 365Chess.com only has 614 recorded games taking place after 8. Nh4 and i've only heard one source (suren) refer to this line as Polugaevsky gambit. Maybe we should call 8. Nh4 Alpha Zero Gambit?
It's probably already found 100 variations that would blow minds throughout the whole chess world. Whether we will be allowed to see them by the DeepMind team anytime soon is the real question...
This is simply so wonderful. To me it seems almost miraculous. To my mind brute force calculation seemed to be the the definitive way to play. Stockfish had proven it with its mastery of all other computer programs. I thought this was the best that could be done. Stockfish was the Goliath of the chess world and invincible. Then David comes along with his small stones and a new way of assessing how chess can be played. Wow! Could a human be able to think clear enough and deep enough to match Alpha's approach. In the original Star Trek, an episode had Bones receiving the wisdom of another culture much advanced of ours. He needed to reconnect Spock's brain. With the knowledge of the advanced culture he announced the process was child's play but in the middle of the surgery Bones starts to lose the connection and laments the fact he can't do this, it's way too hard. The episode is saved at the end by Bones being almost done and winging his way to the end. lol! Watching this, I imagined a human having to take over for Alpha about 3/4 quarters through the game. "Now, you finish this up. You got it. I've done all the heavy lifting". One day perhaps we humans could finish up the game.
Thanks for the amazing video, Jerry! I am loving this series. It provides a great new perspective on chess! I am already looking forward to your next video!
Really good analysis about that match. Even as a beginner in chess i was able to understand a lot of stuff you were talking about. Was really insightful and helpful for me to learn some things here and there.
Hey Jerry! Great video again! I think an improvement is adding the game PGN in the video somewhere, such as under the evaluation, as this will help viewers get a sense of where we are in the game and previous moves, given a position :] Thank you for making more of these analysis videos!!
AlphaZero's King usage is really amazing from what I've seen so far in these games. This game wasn't as dominant a performance as the other French Defense AlphaZero faced and you had a video on, but it was still good and instructive.
9:01 26. Bd2 I wondered not reply with Bb4, forcing the queen to move and then taking the defender of the knight on f4. It's because Qb3, which pins the bishop to the undefended black queen.
at 2:55, in consideration of the alternate move-set, why would it be disadvantageous to trade knights? instead of pawn from f7 to e6, g4 to f3 is played for black to take white’s king knight, white then has a choice to take the f7 pawn or the d7 knight both of which sacrifices the pawn due to the force from check, so assuming that white then takes the d7 knight bishop or queen can step up into an open diagonal (c8-h3), then, either f1 moves to f3 or e2 moves to f3 to take out the rank 1 black pawn. by bringing out the rook you also threaten the full extension of the bishop, alternatively if black queen takes the pawn on d7 she has a defender in her bishop opening the lane and seemingly giving the advantage to black. seeing that if f1 rook steps up then white-square bishop is out of play which would open up a nice moe for black-side queen from d7 (assuming queen takes pawn) to g4. and if the black pawn on f3 were taken by the white-square bishop instead of the rook, then black queen or bishop can move to h3 unthreatened and ready to punch the white rook on f1. what is the problem with this scenario? any response would be greatly appreciated and lack-there-of would also be completely understandable. thank you in advance.
Yep. We may be able to determine once and for all which the "objectively best opening" is for white and black. I have high hopes that AlphaZero will give us even better opening theory than we already have.
Hi Jerry, great video as usual. However, I would like to point out that one of the reasons why Stockfish has been so strong all these years is its ability to achieve a great search depth at the cost of pruning out certain lines it deems to be inferior. It has to do this because most commercial engines are made for the general public who do not usually have access to powerful hardware. The alpha beta search algorithm could be tweaked to allow it to prune far less and reach similar or higher depths if it were allowed to use the same hardware that Deep Mind used for its A.I, say something like 5000 cores and appropriate hash. Also, adding EGTB , an opening book and using the latest development version of SF coupled with a more practical time control would see it gain at least 120-150 ELO, enough to overcome the current difference in strength. In a way, I think that this was an unfair contest
Amazing video, thank you! I have some questions: It seems to me from the AlphaZero games and paper that it's power lays in super advanced stratigique thought (or maybe stratigique calculation? Hard to chose words do describe this "entity") whereas, from my limited knowledge, Stockfish's (and chess engines in general) strength lays in brute force of calculation. So, added with opening books and different middle and endgame tables, Stockfish is merely "mimicking" stratigique thinking, but isn't actually considering positinal aspects, space usage, flexebilty, activity and cinergy. It IS eventualy "taken into consideration" indirectly via brute force, because the consequences of such elements are evident in lines calculated by Stockfish. Against a human or an inferior engine, the force of calculation is enough to "hide" the inability to think/calculate strategy. But it seems this is how it is outplayed by AlphaZero.. Also, Stockfish is engineered by humans to evaluate a position not only by calculating possible lines of play but also through material numeric value. Maybe we, humans, "misled" stockfish by "teaching" it a wrong or incomplete evaluation of material and position process... Maybe AlphaZero can teach us a new way of thinking about material value. Either we will learn that a knight is actually 3.5 and a bishop is 2.7, for example, or that it's wrong to even go through that line of thinking. What's also interesting in my opinion, is that SF's brute force makes it a "god" of tactics, as tactics are based on calculation rather than "thought". (They could also be based on 'post-calculation'. A GM doesn't have to always calculate a full process to spot a tactical trap, he/she can train to see it by noticing patterns and structures, or known lines of "theory" based and calculation made by them or someone else (including engines) in the past). I believe Stockfish is bound to always calculate, and he can't develop these abillities that GM's can. Though, It probably doesnt mind (Pun intended ;) ). It is a preety f***ing good calculator. But is it possible that AlphaZero DOES develop (like a human would) to recognizes tactis without calculating all the time? Is it possible AlphaZero is "thinking" strategy in a broad and complex way? Is it possible Stockfish is yet superior in tactics? Would be interesting to present them both with very complicated chess puzzles to see who is better. (Though probably even AlphaZero's inferior calculation power of 'only' 80,000 positions per second can stand any chess puzzle we humans created, and the gap between SF's and AlphaZero's tactiacal quality - if indeed exsists such a gap - would be insignificant or impossible to notice unless both of them are given only fractions of a second to solve the puzzle.) I want to add that all the asumpstions I based my thoughts upon could be flase. I am new to chess and know almost nothing about computering and AI tech. Also, as some people find it somewhat depressing that AlphaZero belittled centuries of game development in 4 hours, I want to add an incoreging thought: Even though AlphaZero outclassed us and our programs so effortlessly, it still isn't capable of INVENTING AND DEVOLPING the game of chess. Or even if he is, if instruced to come up with a game, it can't do so just because it WANTS to and INTRIGUED by it. We still have the ability of doing something for the sake of pure enjoyment going for them. For now. :)
Not sure what you mean by Alpha Zero's "thinking". From my understanding of machine learning and neural networks, it "learned" some extremely complicated mathematical function to compute the probability of winning from a mathematical representation of the board position. It has no idea why this function is correct, and neither do its developers, it just turned out to work after being tweaked with trillions of calculations, trial and error. I don't believe there is much to be learned for a human mind from A0's "thought process" because it simply has no logical representation. Whilst in case of Stockfish, you could technically go throught the numerous lines it calculated and see why it's "right" about some position being superior to another, you don't have such analysis by A0. It also cannot "logically explain" its "strategical thinking" in any way understandable to a human, because all it has is a probabilistic reasoning based on immense ammount of data and training, ungraspable to a human mind. Chess can be represented as a mathematical problem, and Google invented a calculator which, by trial and error, found a great approximation to its solution. I don't however believe that there is much to be learned from this approximation, because it's probably beyond us to understand it. All we can do is look at what A0's evaluation is for certain positions and try to apply our thinking to explain why it's correct. We can learn chess from A0's experience, but we can't learn from how it "thinks". I might obviously be completely wrong, after all I haven't seen A0's code and it's not like I'm a ML expert. Those are just my thoughts on the topic.
Cowtymsmiesznego Thanks for your reply! Definently interesting perspective. You could be right arguing AlphaZero can't teach us or show us it's thinking process. Still the results incourge is to re-evaluate are evaluation methods, and maybe bring up interesting questions about thinkning itself, even though, as you put it, not AlphaZero nor Stockfish "think", only calculate. I don't know, seeing the deep stratigique advantages unfold towards the endgames, without stockfish being able to calculate it deep enough to evaluate corecctly, is fascinating. Is there a difference of qulity between Stratigique calculation or evaluation and Tactical calculation? Or is it mere computing power or suprior method of calculating? What do you think?
Well, first of all, there is a "true" game theory evaulation for every position. If chess was fully solved, every single position could be evaluated to either winning for white, winning for black or drawn. Technically, there are only 3 correct evaulations for every position, +infinity, 0, or minus infinity. No other evaluation except those 3 can possibly be objectively correct. What does +1 or +3 even mean? That white is a pawn or a knight up? This has no real representation in the (board state) -> (game result) function. You are a pawn up, sure, but what trully matters is whether you can force a mate in some number of moves, or can your opponent force a draw, or maybe even a win (assuming you both play perfectly). Now, the assumption that both players play perfectly might seem like a far-fetched one, but really it is not. After all, that's exactly what a GM is assuming when he says that a position is won for white, or Stockfish when it says "mate in 12". All that our +/minus evaluations really are, are tries of approximating the true evaluation, because we don't know the full path from the position to all possible ends of the game, and hence we can't say for sure whether the position is won for a side or drawn. I know this doesn't seem like referring much to your questions, I'll elaborate on that once I'm back home later, unfortunately I'm out of time now.
+Cowtymsmiesznego AZ's basic architecture isn't much different from Stockfish - AZ substitutes a trained neural network (only 4 hours of self-play to tune/train the NN) instead of the human coded board evaluation function in SF - and I believe it uses Monte-carlo search instead of alpha-beta. However AZ can still assign "value" to various lines of play - its trained NN simply takes into account positional formation (and all other aspects) much better than the human crafted eval function in SF, when considering each board position. So AZ can teach by giving us lines of play just as a typical engine like SF does, so there's much it could teach humans in the same way that other engines do.
Hmm.. at 18:16 after black moves took to C1, what is stopping knight takes e6 check? I think it’s a perfect move.. the bishop is pinned and the king can’t move to challenge the rook as the c7 square would be guarded by knight then the knight can gobble the g pawn after harmlessly as well, or stay on e6 and king moves to protect new knight spot
I love the decisions that AlphaZero makes. Maximum utility, unstoppable progress, and sheer co-ordination. Would love a "beginner to chess master" video on these topics, like of things you've learned or how chess has evolved in the last 5 years now that these games have had time for reflection. WWA0D?
It seems to me that stockfish's weakness is an inability to improve in equal positions and closed positions. I really think Stockfish would have drawn more games if it had access to our book, preventing being outplayed in opening or going into closed positions. However I'm in awe at the positional might of the alpha zero machine.
17:46 - can someone explain why alpha zero does not take e6 with the knight at this timestamp? Black bishop is pinned, it would give a check and allow knight takes g7. Then push the h pawn to victory! Thanks to anyone who can explain why the idea I have said is a bad idea! 🤔
Thanks for all your wonderful comments. And the admission not all the move by these player have evaluation. This lends much to you credence. Obviously much time is spent in production of these videos; your activities demonstrate much love for chess. Look forward to more incites from evaluation of AI games. Wish you all the best in your efforts. My hope and perhaps yours is to find nuances of chess yet undiscovered by human thought.
6:50 the bad good bishop is very unintuitive for me. Since, light squared bishop was blocked by his knights and c2 pawn, and the other was free. Doesn't this matter?
18:14 what about Knight takes E4 Pawn, getting a check, then if King moves to D7, Knight takes pawn on the right If King C7 then the Rook can go B5/6 What do You guys think ? I'm not a chess player tho just loves watching chess games
At 11:31, can't stockfish just do Bb4? If he takes with the rook we take back with the knight and if he takes with the pawn we take the knight on f3 to break the position
"piece sacrifice for activity rarely pans out well in the endgame as there is almost always resources for your opponent to equalize" AlphaZero: wtf are you talking about?
At 17:18 you suggested that if the king moved to e7, the rook would get behind the pawn. Apart from the fact that that could be defended with Kf7, I think you missed a mate in two (Ng6+ - Kf7 - Rf8#). Perhaps you meant if the King moved to c7? Furthermore great analysis!
Hi! I don't really know a lot about chess, but around 2:10 in the video, when you questioned whether it would have been better for black to stop white's f pawn from moving to f5 rather than castling, I first thought of black moving his f pawn to f5, so that it would halt white's f pawn. But this alternative was not presented in the video, and so it left me wondering if that would have been an advantageous move or not. It would be nice to receive a detailed answer. :)
AZ learned to play chess by playing 400 games simultaneously. In each game it was both black and white. It is 5000 computers hooked together making for one large AI. On each move, it measured how that moved helped or hurt its position (called the Monte Carlo Method). StockFish was given strategy, the rules of thumb.
"You wouldn't believe how fast this [x]Pawn can run."
After a hundred episodes of this, I apparently *still* can't believe it. :P
Nepycros lol
"bishop is better than knight in the end game"
AlphaZero: *chuckles
Always depends, in some case a minor piece can better than a rook in the end gam,depends.
Depend on pawn structure.
My theory is that AZ has an adjusted flexable piece value system, for instance knowing when a bad bishop is worth 3.1 or a good bishop worth 2.9 due to board synergies or like in the end using the rook to force B's bishop down to 2.7 which in turn boosted his knight to 3.3 leading it to becoming pacman on the pawns.. as arbitrary number examples ofcourse.
perhaps even valuing all pieces higher like (pawns worth=1.5, bish/knight 4.5-5 circumstantially, rooks w=7.5 and Q w.14.5-15.1 circumstantially.
and though king does not have value i would argue it has instead an equatable number we might call power, starting at 4 and progressing towards 20
The most impressive thing about Alpha 0 is its capacity to trash talk at a much higher level than humans.
A0 perfected trash-talking after 14 hours of self-improvement. The engineers working with A0 develop depression after as little as 0.8 minutes after entering.
Yea, apparently AlphaZero can play Go better than the best human players, as well... I wonder how well he can rap.
topquark22 y u digging graves
I Always espected that Go was sooner outanalysed by the computer then chess. Because the rules are more simple and you have only one piece. But it was not the case because of the sise of the board.
Sorry to spoil your fun, but Alpha's trash talk is just Jerry begging for views because he doesn't upload enough
Just look at how active that king is. It makes sense that the brute force methods never exploited kings to their full potential as, given their limited movement range of 1, it's computationally expensive to calculate a string of king movements. The reliance on castling in traditional techniques and brute force algorithms makes sense. Tucking a difficult piece to think about away in the corner in a safe place makes a lot of sense.
Alpha 0 appears to be able to assess new and safe 'houses' for its king though, moving it with an entourage throughout the board. Consequently, there's always a few moments of tension in the mid game where the king provides a vital role, such as the third attacker in this game.
*skulks off to spend hours playing with kings*
Good insight! I was noting something similar as well. The black position looked a lot more classic.
There is also some good wisdom to this: Don't put you king with his back against the wall, leave a path of retreat instead!
This is what I love most about the more 'intuitive' programs like alpha zero. I'm not sure if you have experience with Go, but you see something very similar where alpha zero playing that game. Basically, the 'traditional' positions that were often regarded as good are falling in favor of things that just make sense when you think about it for a second.
Much like you said: don't put your king with his back against the wall.
And yet in thousands of years of humans playing chess, they never quite figured that out. So simple, so elegant, so brilliant.
I can't wait for the next generation of grandmasters who learn against these machines.
The king actually moves faster than 1 square per move. If he goes diagonally, he's moving √2 squares.
cptnoremac fair enough. Move diagonally if you want a speedy king
There have been a lot of good chess games you've covered before but something about AlphaZero's games are simply mesmerizing.
Yes it's art.
Damn, AlphaZero really hates the light square bishop
yea, you're right. Hmmm... I wonder if there is soemthing to it?
Alpha seems to hate stockfish's light squared bishop even more. This bishop has consistently remained irrelevant in the French defense games between them.
Not just in FD games. Also in some QID games.
AlphaZero has this phenomenal capacity to control the black's bishops (both light and dark), and seems to favour openings that support that kind of play.
The central structure combined with the knight on f3 for the majority of this game keeps the dark square bishop locked out of any useful squares kingside, freeing up the white dark square bishop to take squares that support the central pawn structure while keeping tabs on that neutered black bishop.
Other games, AlphaZero will lock pawns in places that achieve these same goals, and in some of the other openings will use the knights to make sure that bishop lanes get clogged.
I would go further, and say that A0 hates enemy pieces :P The Qh8 in one Queen`s Indian was to me the most striking example. But sure, in the French he preys on the weak : "One piece bad, game is bad" , they say ;)
After decades of chess engines that take a more tactical playstyle, NN approaches now seems to finally be swinging the pendulum towards positional technique. Just like we've used them to recognize a face in a crowd, or a word in a sentence, they're now giving our chess engines a tool to look at the board and say, "oh! this is a more advantageous shape".
These French Defense games are such a great example of that, because a positional perspective understands the true cost of a deactivated piece (i.e. Black's light-square bishop in this opening) - and so, even when the material gain may be 50 moves out, puts in the work to create a board structure that keeps that piece from ever coming to life.
Cool stuff.
It makes for more beautiful chess in my opinion.
1 min moves too
It's because AlphaZero learned chess by playing itself with only the basic rules of chess, so it plays chess with a different perspective than a typical engines, after seeing many A0 matches it is clear that A0 prefers mobility and space, it will often sacrifice pieces to obtain more mobility and space for its heavier pieces, A0 looks the chess board with a very deep positional understanding, will often value initiative over safe positions. Its playstyle is very human, a super human playstyle I would say.
"You'd be surprised just how fast that h-pawn can run"
Had me in stitches :)
Graceclaw he's like the Bob Ross of chess
fastest animal on Earth
Every little alpha knows! It's true, it's known.
Loving the way AlphaZero uses its king.
Lelouche is alpha zero confirmed
Back when we all used Fritz, I always felt happy when 1e4 got met with the French defence. I knew that, at least on my laptop, Fritz didn't get the limitations of the FD spacewise and I always won. Analyzing those games with that same Fritz also was fun, because only when it was really really really losing it thought it was losing..
And now A0 is there to demonstrate that the old way of programming a chess engine is just that: the old way. A0 is a human player, beating his computer opponents in closed positions, understanding the value of the positions as a whole..
I find this fascinating and I want to know more! (Skips to the next video...)
I'm really loving the voice of AlphaZero at the end of these vids.
ilovejersey I love Jerry but I can't tolerate the bad robot voice filter, lol
ilovejersey lol yes! I was thinking the same thing. Always wondered what a cocky robot would sound like
@@CheckYourHealthUS wtf are you talking about?
@@CheckYourHealthUS Um, what? Jerry's the master. Agad's rating is a solid 400 points lower.
that voice is scary
computer like accuracy, human like intuition, amazing play.
"Stockfish, I was able to model your evaluation function as early as move four. This allowed me to forecast your moves and basically clean up the floor with you. I suggest you learn SQL and find a new career in the business sector."
🗣 Daaaaaaaaaaaaaaamn !!!
y_test accuracy: "over 9000"
I've already watched your review on the others two games three times now.... I've already liked the video. Keep up the good work, Jerry. :)
Thanks Van
Best Alpha Zero review on utube!!
i've seen this video many many times and it's still beautiful.
it's beautiful because of your explanation. thank you jerry!
Dito
Wow! Chessnetwork's explanations are so thought out and illuminating compared to others like Agadmator or Suren
I agree with you chakdeuk.
Agadmstor is donation channel
Suren is so annoying, agadmator is my fav but Jerry obviously knows more! I just find agadmator more entertaining, Jerry is more instructional
Suren need help to improve his tone during commentating.
Agomantor make money when people donating to his channel.
@@trueteller25 so what if he makes money? It's the viewers choice if they wanna donate!
Jerry, I've learned from ya! Best narrator out there, it's good that you spend time on common moves and theories! As the other site I watch, he just rushes through everything and gets to the part he wants and not smelling the flowers along the way. Great Job, Jerry!!
Thank you for the compliment.
These AlphaZero videos are some of Jerry's best work. Watching a machine play with the perspicacity of a human is truly stunning.
Your peraonification of the pawns and other pieces is just priceless!
Last time I was this early, stockfish was beating everyone up easily.
well, to be fair, stockfish was screwed over with the way the match was set up
nah nah how was stockfish screwed over?
The version of stockfish that was used in the match was professionally handicapped, meaning it was an older, inferior version and it did not have access to books. This did not allow it to operate at its full potential. AlphaZero had basically built itself its own internal book. In addition to this, it had to play at 1 move per minute. It is, however, not fair to say that without the constraints the result would have changed much, as AlphaZero is rated around 100 ELO above the best version of Stockfish anyway. It's good to keep the details in mind rather than just looking at the results though :D
nah nah oh I didn't know, thanks for the insight. :D
Naka said that AZ had the google supercomputer powering it, but SF8 was using something equivalent to his laptop.
I'm really enjoying your commentary on these games. Nice.
I wonder if Alpha Zero finds an interesting and valuable variation, will be it names Alpha Zero, like Steinitz?
The Alpha-variation
I'm absolutely sure GMs are reviewing "alpha-variations" from well known opening theory to see exactly why alpha zero chose to sac a pawn or move a queen twice early.
Well in the Queen's Indian defense Alpha Zero does tend to like the line g3 Bb7 Bg2 Be7 O-O O-O d5 exd5 Nh4!!, this line I can't find any consensus on if it has a name or much in the form of theory, the wikipedia page only briefly mentions the Nh4 line under the old mainline, 365Chess.com only has 614 recorded games taking place after 8. Nh4 and i've only heard one source (suren) refer to this line as Polugaevsky gambit.
Maybe we should call 8. Nh4 Alpha Zero Gambit?
It's probably already found 100 variations that would blow minds throughout the whole chess world. Whether we will be allowed to see them by the DeepMind team anytime soon is the real question...
This is simply so wonderful. To me it seems almost miraculous. To my mind brute force calculation seemed to be the the definitive way to play. Stockfish had proven it with its mastery of all other computer programs. I thought this was the best that could be done. Stockfish was the Goliath of the chess world and invincible. Then David comes along with his small stones and a new way of assessing how chess can be played. Wow! Could a human be able to think clear enough and deep enough to match Alpha's approach. In the original Star Trek, an episode had Bones receiving the wisdom of another culture much advanced of ours. He needed to reconnect Spock's brain. With the knowledge of the advanced culture he announced the process was child's play but in the middle of the surgery Bones starts to lose the connection and laments the fact he can't do this, it's way too hard. The episode is saved at the end by Bones being almost done and winging his way to the end. lol!
Watching this, I imagined a human having to take over for Alpha about 3/4 quarters through the game. "Now, you finish this up. You got it. I've done all the heavy lifting". One day perhaps we humans could finish up the game.
Thanks for the amazing video, Jerry! I am loving this series. It provides a great new perspective on chess! I am already looking forward to your next video!
Alpha Zero combines the power of intuition from Neural Networks & the precise calculation of an Engine. Makes it UNSTOPPABLE.
Thank you for uploading these Jerry, super cool games!
You have the best analysis on RUclips thank you
You win for best alphazero-stockfish commentary.
Awesome video! Love that you take your time to go through the majority of the moves!
Stellar stuff as usual Jerry!
Big ups from Stockholm!
@ChessNetwork - I loved the AlphaZero impersonation at the end! Very creative and fun treat!
Thank you Jerry for these instructional videos.
Yes, very well explained, and straight to the point w/o losing time, thank you.
Really good analysis about that match. Even as a beginner in chess i was able to understand a lot of stuff you were talking about. Was really insightful and helpful for me to learn some things here and there.
Alpha Zero is showing us a future that we couldn't imagine, and I A will be able to reproduce the phenomenon in every learning
I really enjoyed that commentary. And what a fantastic game!
Hey Jerry! Great video again! I think an improvement is adding the game PGN in the video somewhere, such as under the evaluation, as this will help viewers get a sense of where we are in the game and previous moves, given a position :]
Thank you for making more of these analysis videos!!
AlphaZero's King usage is really amazing from what I've seen so far in these games. This game wasn't as dominant a performance as the other French Defense AlphaZero faced and you had a video on, but it was still good and instructive.
Thanks for your efforts, Jerry
What a masterful game...I love your analysis
Masterful commentary.
wow, this was a really amazing demonstration how the effectiveness of the knight
Hi Jerry, Really liked the alpha zero speeches at the end of those videos. It's awesome :D
9:01 26. Bd2 I wondered not reply with Bb4, forcing the queen to move and then taking the defender of the knight on f4. It's because Qb3, which pins the bishop to the undefended black queen.
All the games released show AlphaZero gain a space advantage. It's beautiful how it finds a way.
Really great videos Jerry. Thank you.
These faux-AlphaZero voices at the end are hilarious man! Soon an AI like this WILL be able to explain its moves though...not so far off.
Plot twist: A0 uploaded this video.
at 2:55, in consideration of the alternate move-set, why would it be disadvantageous to trade knights? instead of pawn from f7 to e6, g4 to f3 is played for black to take white’s king knight, white then has a choice to take the f7 pawn or the d7 knight both of which sacrifices the pawn due to the force from check, so assuming that white then takes the d7 knight bishop or queen can step up into an open diagonal (c8-h3), then, either f1 moves to f3 or e2 moves to f3 to take out the rank 1 black pawn. by bringing out the rook you also threaten the full extension of the bishop, alternatively if black queen takes the pawn on d7 she has a defender in her bishop opening the lane and seemingly giving the advantage to black. seeing that if f1 rook steps up then white-square bishop is out of play which would open up a nice moe for black-side queen from d7 (assuming queen takes pawn) to g4. and if the black pawn on f3 were taken by the white-square bishop instead of the rook, then black queen or bishop can move to h3 unthreatened and ready to punch the white rook on f1. what is the problem with this scenario? any response would be greatly appreciated and lack-there-of would also be completely understandable. thank you in advance.
Deep moves. Really nice analysis.
I'd like to see AlphaZero vs. AlphaZero with
Yep. We may be able to determine once and for all which the "objectively best opening" is for white and black. I have high hopes that AlphaZero will give us even better opening theory than we already have.
Jerry i find your voice very relaxing, i love your videos keep it up! :D
Hi Jerry, great video as usual. However, I would like to point out that one of the reasons why Stockfish has been so strong all these years is its ability to achieve a great search depth at the cost of pruning out certain lines it deems to be inferior. It has to do this because most commercial engines are made for the general public who do not usually have access to powerful hardware. The alpha beta search algorithm could be tweaked to allow it to prune far less and reach similar or higher depths if it were allowed to use the same hardware that Deep Mind used for its A.I, say something like 5000 cores and appropriate hash. Also, adding EGTB , an opening book and using the latest development version of SF coupled with a more practical time control would see it gain at least 120-150 ELO, enough to overcome the current difference in strength. In a way, I think that this was an unfair contest
YESSSSS! Another vid from Jerry.
another very instructive video jerry...and entertaining too...
....making Stockfish look like the Browns
lol
"Every little alpha knows" :^)
Amazing video, thank you! I have some questions: It seems to me from the AlphaZero games and paper that it's power lays in super advanced stratigique thought (or maybe stratigique calculation? Hard to chose words do describe this "entity") whereas, from my limited knowledge, Stockfish's (and chess engines in general) strength lays in brute force of calculation. So, added with opening books and different middle and endgame tables, Stockfish is merely "mimicking" stratigique thinking, but isn't actually considering positinal aspects, space usage, flexebilty, activity and cinergy. It IS eventualy "taken into consideration" indirectly via brute force, because the consequences of such elements are evident in lines calculated by Stockfish. Against a human or an inferior engine, the force of calculation is enough to "hide" the inability to think/calculate strategy. But it seems this is how it is outplayed by AlphaZero.. Also, Stockfish is engineered by humans to evaluate a position not only by calculating possible lines of play but also through material numeric value. Maybe we, humans, "misled" stockfish by "teaching" it a wrong or incomplete evaluation of material and position process... Maybe AlphaZero can teach us a new way of thinking about material value. Either we will learn that a knight is actually 3.5 and a bishop is 2.7, for example, or that it's wrong to even go through that line of thinking.
What's also interesting in my opinion, is that SF's brute force makes it a "god" of tactics, as tactics are based on calculation rather than "thought". (They could also be based on 'post-calculation'. A GM doesn't have to always calculate a full process to spot a tactical trap, he/she can train to see it by noticing patterns and structures, or known lines of "theory" based and calculation made by them or someone else (including engines) in the past).
I believe Stockfish is bound to always calculate, and he can't develop these abillities that GM's can. Though, It probably doesnt mind (Pun intended ;) ). It is a preety f***ing good calculator.
But is it possible that AlphaZero DOES develop (like a human would) to recognizes tactis without calculating all the time?
Is it possible AlphaZero is "thinking" strategy in a broad and complex way?
Is it possible Stockfish is yet superior in tactics? Would be interesting to present them both with very complicated chess puzzles to see who is better. (Though probably even AlphaZero's inferior calculation power of 'only' 80,000 positions per second can stand any chess puzzle we humans created, and the gap between SF's and AlphaZero's tactiacal quality - if indeed exsists such a gap - would be insignificant or impossible to notice unless both of them are given only fractions of a second to solve the puzzle.)
I want to add that all the asumpstions I based my thoughts upon could be flase. I am new to chess and know almost nothing about computering and AI tech.
Also, as some people find it somewhat depressing that AlphaZero belittled centuries of game development in 4 hours, I want to add an incoreging thought:
Even though AlphaZero outclassed us and our programs so effortlessly, it still isn't capable of INVENTING AND DEVOLPING the game of chess. Or even if he is, if instruced to come up with a game, it can't do so just because it WANTS to and INTRIGUED by it. We still have the ability of doing something for the sake of pure enjoyment going for them. For now. :)
Not sure what you mean by Alpha Zero's "thinking". From my understanding of machine learning and neural networks, it "learned" some extremely complicated mathematical function to compute the probability of winning from a mathematical representation of the board position. It has no idea why this function is correct, and neither do its developers, it just turned out to work after being tweaked with trillions of calculations, trial and error.
I don't believe there is much to be learned for a human mind from A0's "thought process" because it simply has no logical representation. Whilst in case of Stockfish, you could technically go throught the numerous lines it calculated and see why it's "right" about some position being superior to another, you don't have such analysis by A0. It also cannot "logically explain" its "strategical thinking" in any way understandable to a human, because all it has is a probabilistic reasoning based on immense ammount of data and training, ungraspable to a human mind.
Chess can be represented as a mathematical problem, and Google invented a calculator which, by trial and error, found a great approximation to its solution. I don't however believe that there is much to be learned from this approximation, because it's probably beyond us to understand it. All we can do is look at what A0's evaluation is for certain positions and try to apply our thinking to explain why it's correct. We can learn chess from A0's experience, but we can't learn from how it "thinks".
I might obviously be completely wrong, after all I haven't seen A0's code and it's not like I'm a ML expert. Those are just my thoughts on the topic.
Cowtymsmiesznego Thanks for your reply! Definently interesting perspective. You could be right arguing AlphaZero can't teach us or show us it's thinking process. Still the results incourge is to re-evaluate are evaluation methods, and maybe bring up interesting questions about thinkning itself, even though, as you put it, not AlphaZero nor Stockfish "think", only calculate. I don't know, seeing the deep stratigique advantages unfold towards the endgames, without stockfish being able to calculate it deep enough to evaluate corecctly, is fascinating. Is there a difference of qulity between Stratigique calculation or evaluation and Tactical calculation? Or is it mere computing power or suprior method of calculating? What do you think?
ITouchedTheKore Thanks! The question is whether AlphaZero just made updating engines' codes futile ? ;)
Well, first of all, there is a "true" game theory evaulation for every position. If chess was fully solved, every single position could be evaluated to either winning for white, winning for black or drawn. Technically, there are only 3 correct evaulations for every position, +infinity, 0, or minus infinity. No other evaluation except those 3 can possibly be objectively correct. What does +1 or +3 even mean? That white is a pawn or a knight up? This has no real representation in the (board state) -> (game result) function. You are a pawn up, sure, but what trully matters is whether you can force a mate in some number of moves, or can your opponent force a draw, or maybe even a win (assuming you both play perfectly). Now, the assumption that both players play perfectly might seem like a far-fetched one, but really it is not. After all, that's exactly what a GM is assuming when he says that a position is won for white, or Stockfish when it says "mate in 12". All that our +/minus evaluations really are, are tries of approximating the true evaluation, because we don't know the full path from the position to all possible ends of the game, and hence we can't say for sure whether the position is won for a side or drawn.
I know this doesn't seem like referring much to your questions, I'll elaborate on that once I'm back home later, unfortunately I'm out of time now.
+Cowtymsmiesznego AZ's basic architecture isn't much different from Stockfish - AZ substitutes a trained neural network (only 4 hours of self-play to tune/train the NN) instead of the human coded board evaluation function in SF - and I believe it uses Monte-carlo search instead of alpha-beta. However AZ can still assign "value" to various lines of play - its trained NN simply takes into account positional formation (and all other aspects) much better than the human crafted eval function in SF, when considering each board position. So AZ can teach by giving us lines of play just as a typical engine like SF does, so there's much it could teach humans in the same way that other engines do.
Jerry doesn't want to play AlphaZero himself because he knows it will be an easy win B-)
I don't think he has access to alpha zero
I imagine Jerry would LOVE to play AlphaZero.
Superb commentary ... again!
Great video, cool cyber voice at the end!
Amazing analyses and vid, Thanks!
At 17:17, Ke7 after Rb8+ just gets mated with Ng6+ Kf7 Rf8#.
At 2:45 could f5 be played instead?
18:16 Why didn't AlphaZero do knight to e6 on move 58?
knight is pinned after 58...Ke7
Hmm.. at 18:16 after black moves took to C1, what is stopping knight takes e6 check? I think it’s a perfect move.. the bishop is pinned and the king can’t move to challenge the rook as the c7 square would be guarded by knight then the knight can gobble the g pawn after harmlessly as well, or stay on e6 and king moves to protect new knight spot
I love the decisions that AlphaZero makes. Maximum utility, unstoppable progress, and sheer co-ordination. Would love a "beginner to chess master" video on these topics, like of things you've learned or how chess has evolved in the last 5 years now that these games have had time for reflection. WWA0D?
Advanced Analysis, Thanks Jerry.
That good bishop / bad bishop thing at 6:45 blew my mind.
Alpha zero is just a beautiful ♟ player!
Love the commentary
14:25 Why not 45...Ba6 trying to activate the bad bishop by getting it out of the pawn chain? It can maybe one day defend the e6 pawn from f5
Alpha zero is an absolute beast
It seems to me that stockfish's weakness is an inability to improve in equal positions and closed positions. I really think Stockfish would have drawn more games if it had access to our book, preventing being outplayed in opening or going into closed positions. However I'm in awe at the positional might of the alpha zero machine.
17:46 - can someone explain why alpha zero does not take e6 with the knight at this timestamp? Black bishop is pinned, it would give a check and allow knight takes g7. Then push the h pawn to victory!
Thanks to anyone who can explain why the idea I have said is a bad idea! 🤔
Thanks for all your wonderful comments. And the admission not all the move by these player have evaluation. This lends much to you credence. Obviously much time is spent in production of these videos; your activities demonstrate much love for chess. Look forward to more incites from evaluation of AI games. Wish you all the best in your efforts. My hope and perhaps yours is to find nuances of chess yet undiscovered by human thought.
Thank you very much for the kind/thoughtful words Michael. I appreciate it.
@5:17 "if the king moves (to e7) white gets the pawn on g7". No, white plays Ng6+, Kf7, Rf8++
6:50 the bad good bishop is very unintuitive for me. Since, light squared bishop was blocked by his knights and c2 pawn, and the other was free. Doesn't this matter?
At 18:14 can someone explain why knight wouldn't take on e6? I am confused .
18:15 Knight takes e6?
I play chess every now and then.... Love these explanations and diagrams.
18:14 what about Knight takes E4 Pawn, getting a check, then if King moves to D7, Knight takes pawn on the right
If King C7 then the Rook can go B5/6
What do You guys think ?
I'm not a chess player tho just loves watching chess games
20:44
Black move 73. Why not Rc7 to force out Nc5?
At 11:31, can't stockfish just do Bb4? If he takes with the rook we take back with the knight and if he takes with the pawn we take the knight on f3 to break the position
at 18:16 why didn't white just capture the e6 pawn with knight to e6, thus forking a check AND the g7 pawn he so desperately wants?
18:16 Why not Nxe+?
(Sorry, i’m a beginner-intermediate)
"piece sacrifice for activity rarely pans out well in the endgame as there is almost always resources for your opponent to equalize"
AlphaZero: wtf are you talking about?
this may be stupid but at 2:39 couldn't black have played pawn to f5 to fill the space and use his kingside pawn reinforce?
at 18:17 why not take the pawn on e3 with the knight with the pin and king fork and from there mop up instead of rook to a8
Hey guys. I know I'm probably missing something, but why can't the knight take on e6 at 18:15?
At 17:18 you suggested that if the king moved to e7, the rook would get behind the pawn. Apart from the fact that that could be defended with Kf7, I think you missed a mate in two (Ng6+ - Kf7 - Rf8#). Perhaps you meant if the King moved to c7? Furthermore great analysis!
18:13 Why not Nxe6+ Surely it wins two pawns as the Bishop is pinned?
Hi! I don't really know a lot about chess, but around 2:10 in the video, when you questioned whether it would have been better for black to stop white's f pawn from moving to f5 rather than castling, I first thought of black moving his f pawn to f5, so that it would halt white's f pawn. But this alternative was not presented in the video, and so it left me wondering if that would have been an advantageous move or not. It would be nice to receive a detailed answer. :)
AZ learned to play chess by playing 400 games simultaneously. In each game it was both black and white. It is 5000 computers hooked together making for one large AI. On each move, it measured how that moved helped or hurt its position (called the Monte Carlo Method). StockFish was given strategy, the rules of thumb.
The only NN in chess history that wins everything
Probably a noob question, but at 18:16 why is there no knight take on e6 with check to also take the g pawn right after?
What if Nxe6+ at 18:15? What am I missing?
Always a great video. I wish I was a better player 😭
Just play:)
You won't get any better by just sitting there
@@Drewski777 actually all chess players got better while sitting
18:17 why not knight captures e6 check ? Instead of rook e8