I work in big data and did a detailed reading of the paper the team released. I think the amazing thing for me is the difference in the way the two algorithms' approach the game. SF can be described more as maximizing or minimizing material gain or loss with ruthless efficiency. A0 played 44 million games of chess against itself and trained to optimize the final outcome of the game--not the material. It used neural nets to recognize patterns and updated Bayesian priors based on the move effectiveness as judged by the result of the game. You watch these games and these algorithm differences become so obvious. A0 plays positional and sacrifices material. SF hordes material at the cost of trapping itself in the corner of the board. Playing Chess myself I've always had my best games against computer opponents when the position restricted mobility and there weren't a lot of deep tactics to calculate. Clearly there's something there! This is the start of a chess renaissance!
In a way, we can have this strange realization that SF is actually a strong marker for the current human wisdom in chess up until AlphaZero. Codifying arbitrary value systems that have nothing to do with the rules of chess. In the documentary AlphaGo, the key realization was that the NN did not care about how much it will win. It only mattered that it does. In contrast, the prevailing human play has long used "score lead" as a proxy to the "probability" of winning when in fact, AlphaZero shows, it really is not. You can see that this is the same now-refuted wisdom that SF operates in.
Sacrificing the knight at the beginning to get a clear advantage in the endgame was so obvious to me. I think Alpha zero has been programmed to examine all my 1350 level games.
DatWill its crazy u said that. About halfway through the 3rd AlphaZero game i saw analized i was wondering if it was really a super A.I. or if they were just mimicking your ideas. Another conspiracy theory proved correct. Ty.
All of these games seem to have a common theme, and that is restricting the opponents piece mobility and activity. This long-term piece immobility is likely something where the time horizon is too long to calculate the benefits of through brute force, but something AlphaZero has "learned intuitively." This seems to be the major weakness of the Stockfish engine which AZ repeatedly exploited to win.
Yeah, I think you nailed it. There is obviously a lot of stuff to learn from the games. Most of it though is out of reach for low rated player like myself. You need a lot of chess knowledge if you are going to get something out of trying to analyze the games. However one thing which no matter your skill you can make use of is, like you say, the lesson of activity and restriction. Feels pretty revolutionary. These are well known basic core concepts ofc but seems like the alpha zero games are showing that we might not have the right approach to the game overall. Feels like a new way of chess thinking could result from it.
Also with a Queen's Indian opening. To have a pawn push early to have that advantage. I've always opened with my Queen's pawn for that reason unless i'm just trying something different...
Yes, this is not a new concept, but I don't think it's ever been so comprehensively and transparently demonstrated before in (super) high level chess. The idea that you could sacrifice multiple pawns and entire minor pieces for long-term positional advantage, combined with the continuous application of prophylaxis... it's just stunning. A0 doesn't hesitate (not that a computer would) to repeatedly do this and it does it ruthlessly. This has already made a contribution to chess knowledge IMO, and I look forward to more A0 games.
Tal and his predecessors would be smiling ("See! We are right! That is how chess is played.") It is so good to see that Chess, like Go, has unique qualities and such complexity that you cannot consider theories and strategy settled. Makes for an exciting future for the game.
It has been shown through these games that Stockfish simply prioritizes amount of pieces on the board, and that many can move. As long as Alpha Zero can manipulate positions with sacrifices, he can and will exploit Stockfish's weakness.
Wanna know my favorite part about these games? "Usually you don't want to..." "Usually you don't see..." "Usually it's not a good idea to..." But AlphaZero fucking does it anyway lol #thuglife These games renew my hope that chess can be exciting, and it makes me want to play some games.
@@sliver170 Such a regressive remark. Progress is small incremental steps. If we don't try to think like A0 on chess, we won't ever be. Applicable on anything else.
Most traditional engines seem to do poorly against fortress or blockade positions. A0 seems to be able to avoid the pitfalls other engines suffer from as it actually sees over the horizon rather than relying strictly on number crunching. All that is ultimately meaningless if you can't see the limitations and dangers that the edge of the board present. In many games I've looked at it's almost as if the engine believes it has a phantom rank behind the back rank or an extra file beside the a and/or h-file. For example, in the games where SF lost repeatedly due to a cramped QID it almost seems like it would like to either move it's knight back or it's rook around, in which case true equality might have been attained. However in all these games notice how SF always seem to have either a bad bishop or at least one other inactive piece such as a knight or rook and even sometimes the queen. A0 plays in a way that almost seems to check all it's opponents pieces at once on every move, never losing tempo until literally the whole board in is zugwang.
The talk of engines "learning" in-between games is non-sense. They play different moves because the pruning process is somewhat random. Stockfish is not an A.I, it´s just a sofisticated algorithm. AlphaZero wasn´t learning anything while playing against Stockfisk either. Training is a seperate process from preformance. After AlphaZero's 4 hour training section it will remain unchanged, no matter how many games it plays.
It's basically a case of our human narrator anthropomorphizing the chess engines (as that is what a human player in Stockfish's position would probably be thinking and doing in a real match) as a way of creating a more engaging narrative for us humans in the audience.
Human created monster who can dig out our brain as dirt ! We can not beat AI because our brain is made of jelly neuron ! Jelly neron protein can be denatured ! AI only costs electricity and precision oil !
Another brilliant summary Danny, thanks again for taking the time to analyse these games and for sharing your insights and enthusiasm, much appreciated.
Qc4!! is pretty sick from a computer's perspective, and a typical brilliant move of alphazero in these wins. It's a very human thing- provoking ...b5 in response and improving upon the position you would get after hxg5 fxg5 and Qh4+. It turns out to be crucial in the build up of white's attack in the end, but there is no human or (current) computer capable of calculating all of that. It's a very human move- you can imagine a lot of strong grandmasters (Carlsen, Karpov, Kasparov) to be able to see it (as it's a strict improvement over taking on g5 straight away), but before alphazero something had to be hardcoded in for a computer to spot it. Alphazero shows a new age for chess AI and ofc beyond... Simply amazing.
To all the people emotionally attached to Stockfish: SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.
Alpha Zero is definitely a very impressive chess engine and probably stronger overall than Stockfish (especially since it is so young and had really good growth potential compared to Stockfish), however with more common/logical timecontrols, more comparable hardware as well as opening tables (especially since AlphaZero from my understanding basically build it's own opening table it could take into the games)we definitely wouldn't have seen a 25-25-0 score
AZ is not a chess engine. It plays as well in Go or Shogi. AZ in chess in only experiment how generic is theirs algorithm. Creating chess engine never was a real goal. They want to have algorithm which will be able to learn every task with only rules given.
Abbec : Well somebody had to give AZ more than the basic chess rules, I'm pretty sure. People often fantasize what AI can do, and it's always pure BS. Even in machine learning, you have to give your algorithm ways of evaluating how "something" is "good" related to "something else". Or at least how to learn how it can evaluate it by itself, be it by bruteforce (the learning phase of any machine learning algorithm), or by feeding it examples.
Whole idea of AI is to give method of learning instead of solution. I'm not fantasize, I just mention what is goal of deep mind team. It's not chess engine. It's bigger idea behind it.
mjgayle52 I'm pretty sure that Stockfish won more games than Alpha0 out of the hundred. But Danny chose the interesting ones (the ones where Stockfish lost) which is brilliant, because people would't be impressed, if he showed a random game where worlds strongest chess engine beats a random AI, but seeing stockfish 8 lose is something that attracts a lot of viewers and is interesting.
That’s the most amazing game I’ve ever seen. Human game does not even have the term like “Long term piece sacrifice” The knight sacrifice in the opening for a rook in the endgame. Wow!!! That’s increadible
I think it is evident that traditional engines which used human understanding of the game, use an " avoid loss at all costs" type of style, its tied to the human fear of loss ( which is why many games end with a draw). This Deepmind AI on the other hand, plays to win, instead of having a " Make no mistakes and avoid losses at all costs" mentality, its disregards even the potential of losing, and just plays to win.
This just shows again that chess can be played on much higher level than we could ever imagine. In early 2000s there was this fear that human chess would get killed by computers, that GMs of future are gonna be all about memorization of openings and variations. However we can see that chess has to offer so much more, positional, strategic, long term piece sacrifices and other futuristic things. I think that this could only help GMs of future. We may enter, or we are already in this new era of chess where games are attack minded and extremely exiting. Beware of human development also, i am confident that we are gonna see a lot more immortal games between humans. Just imagine if Tal and Capablanca were alive to see this, and what they could get from this. One thing is certain, chess survived this machine fear and it's back, stronger than ever.
AlphaZero is terrific. Computer Chess world is changed after just 4 hours of training. I hope they use AlphaZero to solve real world problems, it seems smart as god. I guess this opens new doors in Robot AI as well. New Era.
I play just barely enough chess to do a competent job of losing to most chess programs on a medium-low setting, and I certainly never sit around watching videos about chess... but you got me on the AI angle and I'm back for more the next day because you guys make this fairly easy to understand (provided I occasionally back up and dwell on it when something unexpected happens). Thanks!
So basically alpha zero starts of every game by trading material in order to gain a strategical position giving him control over the key squares on the board. Then when the right moment is there he trades everything with some fine tactics giving him the favorable end game, resulting in the win. AI is truly op, simply wow
It's simply incredible how the site's evaluation always initially reads something like -3 but then quickly adjusts to "only" -0.5 or so in a couple of seconds, and as time goes on slowly gets less and less negative, after every silly-looking Alpha move
Danny, SF says after Qc4 that hesIT is winning by 1.75. I really love this. For some reason I thought that this kinda of play was possible. What i mean is that Pure calculations from SF is not 100% accurate. As you can clearly see SF says heIT is winning but is not do to the fact of Skilled play by AZ. What do you think?
One of the (many) things to take away from this, is that the concept of "correct move" goes by the wayside, unless you have AlphaInfinity or something, which can solve chess to the bottom. Projects such as chess.com's CAPS needs re-thinking. Magnus was right all along in not always placing too much confidence in the (pre-AlphaZero) engines. What wonders will all these new A0-lines do for him and other top-GMs...
"We also report the win/draw/loss results of 100 game AlphaZero vs. Stockfish matches starting from each opening, as either white (w) or black (b), from AlphaZero’s perspective." From The Paper: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm by David Silver, Thomas Hubert, Julian Schrittwieser et all, page 6 / arxiv.org/pdf/1712.01815.pdf So there was total 1200 games played between AlphaZero and Stockfish with stats: Total games: w 242/353/5, b 48/533/1 Overall percentage: w 40.3/58.8/0.8, b 8.0/88.8/3.2 (copy pasted directly from the paper) Why people talk about only 100 games played and that Stockfish didn't win any games?
Taking a material loss, in order to get a positional advantage for mating pressure that is leveraged back into a material advantage. I want to introduce strong ai to other turn based games
Great analysis of the games. Alphazero plays more like a pure tactician and recognizing and aiming for patterns on the board and using the computer's access to book moves and programmed lines(memorized lines if it was a human player) against it. Alphazero plays more like a human by using long term strategy and calculation, humans just have the flaw of habit moves (as you mention saying "normally this happens") because of time and can only calculate so far in the time given during games. In these games AZ calculated long term strategy within a minute per move...That would be roughly a 45-60 minute game.
There's a big question whether AlphaZero was trained playing only against itself. Maybe they trained it against Stockfish as well. To find its weaknesses. Making it a Stockfish-killer. But maybe it would lose to Comodo.
cbr -- no, I think that 10000 number is wildly wrong. Every 100 points of ELO represents about a 2 to 1 win ratio for the stronger player. Go take a look at the Deepmind paper: arxiv.org/pdf/1712.01815.pdf -- you'll see the ELO gain based on more training just levels off at about one quarter of the way into the four hour training period. (Remember, this 4 hours of training *is on thousands of TPU cores* so that's a lot of training steps!) More time would have _maybe_ given a little more ELO, I think. To get any significant gains, the AlphaZero algorithm would probably need to be enhanced to allow some amount of incorporation of alpha-beta search, or other traditional engine techniques into the play style, rather than just the MCTS (Monte Carlo Tree Search). Again, from the paper: "None of the techniques described in this section are used by AlphaZero. It is likely that some of these techniques could further improve the performance of AlphaZero; however, we have focused on a pure self-play reinforcement learning approach and leave these extensions for future research."
I think past a certain point, chess will essentially be solved, and the AI will be able to either draw every game, or the win/loss will only depend on who is white. In either case, ELO won't work anymore. So I don't believe it is possible to reach ELO 10000.
5:25 lol'd so hard. "... world will never know....like the center of a tootsie roll pop." AKA " Like the number of licks it takes to get to the center of a tootsie roll pop..." idk why it's just how his comments seem to be, like intentional (or unintentional) but just a few degrees of freedom between his reality and the reality he remembers, the real one. It's genius!
Although stockfish was restricted in some ways, I still find it amazing that it only took 4 hours for this neural network to win against computer, which is basically fed with information that human gathered 1500 years...
it's important to point out that AlphaZero was on a far more powerful machine than Stockfish, meaning Stockfish couldn't anaylise as far ahead and that they were given 1 min/move, rather than a classical time control. Something Stockfish is good at is knowing when to spend the time in critical positions. Not trying to take anything away from AZ, it's clearly becoming a force to be reckoned with.
at 6:44 instead bishop c8, why not move rook to h8 and then king back to f8? At least, the king is somewhat secured and protected by pawn, bishop and other pieces.
I appreciate these videos very much! Love your sense of humor too, Mr. Rensch. But I'm a patzer and sometimes you leave me hanging and I'm frustrated. You analyzed Kg7 (which was not played) after White went h4. Your analysis ended with White playing Qe5 and "forking" Black's bishops but then Black's obvious reply (which you did not mention, much less analyze) is Be6, protecting both bishops. What might White do then? What ideas are in this position for White? I mean, I'm sure there must be something because according to the bar at the left of the screen, Stockfish feel inferior as Black after Qe5. But poor patzer me, I'm at a loss for how White might proceed. Can you give a clue please?
Such a good video, can’t say it enough. I’m sitting here at 14:32 trying to think what was the next amazing move A0 played, as we the viewers are challenged to do ... and I’ve already seen this game! Ok, think, think ...
I think Stockfish have a problem of database in this opening, from game many games won alpha zero that way, that's is why the rest of the won game have not bin made public. Alpha zero discovered a hole in Stockfish
![โอ้เธอช่าง... - บี้เดอะสกา ft. ต๋อง เทวัญ (Prod. by ป๋าเพชร) [Official MV]](http://i.ytimg.com/vi/zbWhiLFmVL0/mqdefault.jpg)
Catch Danny's match highlight wrap-up here! th-cam.com/video/6z1o48Sgrck/w-d-xo.html
Chess.com thanks
H
I work in big data and did a detailed reading of the paper the team released. I think the amazing thing for me is the difference in the way the two algorithms' approach the game. SF can be described more as maximizing or minimizing material gain or loss with ruthless efficiency. A0 played 44 million games of chess against itself and trained to optimize the final outcome of the game--not the material. It used neural nets to recognize patterns and updated Bayesian priors based on the move effectiveness as judged by the result of the game.
You watch these games and these algorithm differences become so obvious. A0 plays positional and sacrifices material. SF hordes material at the cost of trapping itself in the corner of the board. Playing Chess myself I've always had my best games against computer opponents when the position restricted mobility and there weren't a lot of deep tactics to calculate. Clearly there's something there! This is the start of a chess renaissance!
In a way, we can have this strange realization that SF is actually a strong marker for the current human wisdom in chess up until AlphaZero. Codifying arbitrary value systems that have nothing to do with the rules of chess. In the documentary AlphaGo, the key realization was that the NN did not care about how much it will win. It only mattered that it does. In contrast, the prevailing human play has long used "score lead" as a proxy to the "probability" of winning when in fact, AlphaZero shows, it really is not. You can see that this is the same now-refuted wisdom that SF operates in.
Sacrificing the knight at the beginning to get a clear advantage in the endgame was so obvious to me. I think Alpha zero has been programmed to examine all my 1350 level games.
rekt
DatWill i
DatWill its crazy u said that. About halfway through the 3rd AlphaZero game i saw analized i was wondering if it was really a super A.I. or if they were just mimicking your ideas. Another conspiracy theory proved correct. Ty.
orochimarujes lol
Lol
All of these games seem to have a common theme, and that is restricting the opponents piece mobility and activity. This long-term piece immobility is likely something where the time horizon is too long to calculate the benefits of through brute force, but something AlphaZero has "learned intuitively." This seems to be the major weakness of the Stockfish engine which AZ repeatedly exploited to win.
Yeah, I think you nailed it. There is obviously a lot of stuff to learn from the games. Most of it though is out of reach for low rated player like myself. You need a lot of chess knowledge if you are going to get something out of trying to analyze the games.
However one thing which no matter your skill you can make use of is, like you say, the lesson of activity and restriction. Feels pretty revolutionary. These are well known basic core concepts ofc but seems like the alpha zero games are showing that we might not have the right approach to the game overall. Feels like a new way of chess thinking could result from it.
Also with a Queen's Indian opening. To have a pawn push early to have that advantage. I've always opened with my Queen's pawn for that reason unless i'm just trying something different...
These are only a few of the 100 games, but yeah it does look like that.
Yes, this is not a new concept, but I don't think it's ever been so comprehensively and transparently demonstrated before in (super) high level chess. The idea that you could sacrifice multiple pawns and entire minor pieces for long-term positional advantage, combined with the continuous application of prophylaxis... it's just stunning. A0 doesn't hesitate (not that a computer would) to repeatedly do this and it does it ruthlessly. This has already made a contribution to chess knowledge IMO, and I look forward to more A0 games.
Most of the games were draws though. I'd like to see those games, which probably also means both engines aren't too far off.
Tal and his predecessors would be smiling ("See! We are right! That is how chess is played.") It is so good to see that Chess, like Go, has unique qualities and such complexity that you cannot consider theories and strategy settled. Makes for an exciting future for the game.
Tal is my favourite. AlphaZero is just a more computationally intensive Tal.
That is crazy. SF likes it's position right up to the point where it is lost. I for one welcome our new computer overlords.
It has been shown through these games that Stockfish simply prioritizes amount of pieces on the board, and that many can move. As long as Alpha Zero can manipulate positions with sacrifices, he can and will exploit Stockfish's weakness.
E
@@gredangeo Not anymore...
I am from the future and Stockfish 15 crushes AlphaZero
@@phase0400 Stockfish went over to the 'Dark Side' starting at SF12. 😏
This is mind blowing and I am shaken.
Danny , make them release all the games - MAKE THEM!
Most of them were draws.......
@@razinoid6863 27 wins
Absolutely love your comments, analysis and zest. In this sea of youtube chess commentators catering to the absolute beginners... I love it!
You're doing a helluva job with these. Keep it up.
David Woosley i think we dont have more than 10 matches :(
Scroogie TW, that's clearly not my fault. Gimme more analysis, dammit.
This series was incredibly high quality stuff for free
Wanna know my favorite part about these games?
"Usually you don't want to..."
"Usually you don't see..."
"Usually it's not a good idea to..."
But AlphaZero fucking does it anyway lol #thuglife
These games renew my hope that chess can be exciting, and it makes me want to play some games.
Yep, A0 is causing people to rethink a lot. And that's a good thing.
Since humans cannot think this far ahead, what use is following AlphaZero's example?
@@sliver170 Such a regressive remark. Progress is small incremental steps. If we don't try to think like A0 on chess, we won't ever be. Applicable on anything else.
That was a masterpiece of a review.
Haven't even watched the video and I'm already psyched. Loving the coverage.
Most traditional engines seem to do poorly against fortress or blockade positions. A0 seems to be able to avoid the pitfalls other engines suffer from as it actually sees over the horizon rather than relying strictly on number crunching. All that is ultimately meaningless if you can't see the limitations and dangers that the edge of the board present. In many games I've looked at it's almost as if the engine believes it has a phantom rank behind the back rank or an extra file beside the a and/or h-file. For example, in the games where SF lost repeatedly due to a cramped QID it almost seems like it would like to either move it's knight back or it's rook around, in which case true equality might have been attained. However in all these games notice how SF always seem to have either a bad bishop or at least one other inactive piece such as a knight or rook and even sometimes the queen. A0 plays in a way that almost seems to check all it's opponents pieces at once on every move, never losing tempo until literally the whole board in is zugwang.
Very good analysis Danny! . You are doing a really good job..I'm learning more from your analysis and brawls n stuff from ur you- tube channel..
mind blown
9.42 Queen h1 -- that woke me up
John Bicycle 9:42
our youtube names have issue.
Amazing knight sacrifice by Alpha Zero. It's positional understanding is mind boggling!!!
The talk of engines "learning" in-between games is non-sense. They play different moves because the pruning process is somewhat random.
Stockfish is not an A.I, it´s just a sofisticated algorithm. AlphaZero wasn´t learning anything while playing against Stockfisk either.
Training is a seperate process from preformance. After AlphaZero's 4 hour training section it will remain unchanged, no matter how many games it plays.
It's basically a case of our human narrator anthropomorphizing the chess engines (as that is what a human player in Stockfish's position would probably be thinking and doing in a real match) as a way of creating a more engaging narrative for us humans in the audience.
I wonder if it even plans any moves in front. Is it just playing the best move at that time, with no future moves in mind?
One of the best games i seen lately TY!
I consider this game, The AlphaZero Immortal
miximup every game we've seen so far is an immortal :)
Human created monster who can dig out our brain as dirt !
We can not beat AI because our brain is made of jelly neuron !
Jelly neron protein can be denatured !
AI only costs electricity and precision oil !
AlphaZero has taken it to the next level! I mean that Knight sac and c4!!! Thanks Danny for the amazing analysis!
9:20 - Qh1 ... !!! When have we ever seen a move like this?
modolief Carlsen played it against Anand in the World Championship!!
www.chessgames.com/perl/chessgame?gid=1736710
Ah, well there you go ... by no less than Carlsen!
Now this is certainly something. I didn't think Machine Learning would win against brute-force alos in games like chess.
kasparov9 LOL Thank you!
Another brilliant summary Danny, thanks again for taking the time to analyse these games and for sharing your insights and enthusiasm, much appreciated.
Qc4!! is pretty sick from a computer's perspective, and a typical brilliant move of alphazero in these wins. It's a very human thing- provoking ...b5 in response and improving upon the position you would get after hxg5 fxg5 and Qh4+. It turns out to be crucial in the build up of white's attack in the end, but there is no human or (current) computer capable of calculating all of that. It's a very human move- you can imagine a lot of strong grandmasters (Carlsen, Karpov, Kasparov) to be able to see it (as it's a strict improvement over taking on g5 straight away), but before alphazero something had to be hardcoded in for a computer to spot it.
Alphazero shows a new age for chess AI and ofc beyond... Simply amazing.
Beautiful lesson. Just beautiful
Absolutely impressive play from AZ. Great analysis Danny! Thanks!
To all the people emotionally attached to Stockfish:
SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls.
The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening.
EGTB-Stockfish had lost positions before the EGTB would be of any use.
Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound.
Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.
Alpha Zero is definitely a very impressive chess engine and probably stronger overall than Stockfish (especially since it is so young and had really good growth potential compared to Stockfish), however with more common/logical timecontrols, more comparable hardware as well as opening tables (especially since AlphaZero from my understanding basically build it's own opening table it could take into the games)we definitely wouldn't have seen a 25-25-0 score
AZ is not a chess engine. It plays as well in Go or Shogi. AZ in chess in only experiment how generic is theirs algorithm. Creating chess engine never was a real goal. They want to have algorithm which will be able to learn every task with only rules given.
Abbec : Well somebody had to give AZ more than the basic chess rules, I'm pretty sure. People often fantasize what AI can do, and it's always pure BS. Even in machine learning, you have to give your algorithm ways of evaluating how "something" is "good" related to "something else". Or at least how to learn how it can evaluate it by itself, be it by bruteforce (the learning phase of any machine learning algorithm), or by feeding it examples.
Whole idea of AI is to give method of learning instead of solution. I'm not fantasize, I just mention what is goal of deep mind team. It's not chess engine. It's bigger idea behind it.
arizona beat san francisco, gotcha
chess has a new king
mjgayle52 I'm pretty sure that Stockfish won more games than Alpha0 out of the hundred. But Danny chose the interesting ones (the ones where Stockfish lost) which is brilliant, because people would't be impressed, if he showed a random game where worlds strongest chess engine beats a random AI, but seeing stockfish 8 lose is something that attracts a lot of viewers and is interesting.
what i have read is that in a 100 game match AlphaZero won 28 games - lost 0 games - with 72 draws
Can't Catch Me
The fuck you are saying, my dude?
+ Can't Catch Me
Nah dude, Stockfish lost hard. I think they actually did 1000 matches and stockfish only one like 2%. Alpha Zero is way better.
@@cantcatchme6749 alpha zero won way more games than stockfish are you high?
Excelent analysis! Thank you for this series!!
All of this is mind-blowing. Great coverage too.
That’s the most amazing game I’ve ever seen. Human game does not even have the term like “Long term piece sacrifice” The knight sacrifice in the opening for a rook in the endgame. Wow!!! That’s increadible
I think it is evident that traditional engines which used human understanding of the game, use an " avoid loss at all costs" type of style, its tied to the human fear of loss ( which is why many games end with a draw).
This Deepmind AI on the other hand, plays to win, instead of having a " Make no mistakes and avoid losses at all costs" mentality, its disregards even the potential of losing, and just plays to win.
This just shows again that chess can be played on much higher level than we could ever imagine. In early 2000s there was this fear that human chess would get killed by computers, that GMs of future are gonna be all about memorization of openings and variations. However we can see that chess has to offer so much more, positional, strategic, long term piece sacrifices and other futuristic things. I think that this could only help GMs of future. We may enter, or we are already in this new era of chess where games are attack minded and extremely exiting. Beware of human development also, i am confident that we are gonna see a lot more immortal games between humans. Just imagine if Tal and Capablanca were alive to see this, and what they could get from this. One thing is certain, chess survived this machine fear and it's back, stronger than ever.
insane game, mind blown. alphazero is amazing
Alpha also blocked the check forcing the queen trade with the pawn move to c4. O.O
AlphaZero is terrific.
Computer Chess world is changed after just 4 hours of training.
I hope they use AlphaZero to solve real world problems, it seems smart as god.
I guess this opens new doors in Robot AI as well. New Era.
Alpha Zero 2.0: Absorbs all human knowledge in 0.0042 nanoseconds, reaches singularity.
4 hours of tons of computing time.
Other than that, yes.
S L
AZ didn't need to absorb human filth. It made its own art.
AlphaZero will never solve human problems !
because his master is our human brain !
sabbas eleftheriadis AlphaZero will probably solve real world problems by killing all humans
This is crazy lol almost every move was like wtf to me but it all made sense
Fascinating
just speechless alpha zero is god
Devvan Butler Jehovah would argue against Clapton or AZ! ! !
AI learned that the key to winning is sacrificing material to gain long term positional advantages. We are so screwed.
I play just barely enough chess to do a competent job of losing to most chess programs on a medium-low setting, and I certainly never sit around watching videos about chess... but you got me on the AI angle and I'm back for more the next day because you guys make this fairly easy to understand (provided I occasionally back up and dwell on it when something unexpected happens). Thanks!
Game ten watched ❤
Loved this serie Danny!
Qh1 .. move of the year!! Incredible
So basically alpha zero starts of every game by trading material in order to gain a strategical position giving him control over the key squares on the board. Then when the right moment is there he trades everything with some fine tactics giving him the favorable end game, resulting in the win.
AI is truly op, simply wow
1989Brumo Yeah thats pretty much sums it up.
It's simply incredible how the site's evaluation always initially reads something like -3 but then quickly adjusts to "only" -0.5 or so in a couple of seconds, and as time goes on slowly gets less and less negative, after every silly-looking Alpha move
Danny,
SF says after Qc4 that hesIT is winning by 1.75. I really love this. For some reason I thought that this kinda of play was possible. What i mean is that Pure calculations from SF is not 100% accurate. As you can clearly see SF says heIT is winning but is not do to the fact of Skilled play by AZ. What do you think?
One of the (many) things to take away from this, is that the concept of "correct move" goes by the wayside, unless you have AlphaInfinity or something, which can solve chess to the bottom. Projects such as chess.com's CAPS needs re-thinking. Magnus was right all along in not always placing too much confidence in the (pre-AlphaZero) engines. What wonders will all these new A0-lines do for him and other top-GMs...
Another great game with excellent commentary. I always learn something by listening.
c4 also stops Qd5+ later after B and R exchange. Such a visionary move.
Great analysis
This is so good.. i watch it many times 😄
Awesome review, really cool to see how advanced techniques find value in minor positional advantages
This is Danny at his best
Mindboggling. Wow
Way to go Danny; excellent and professional commentary. Until Google teaches Alpha Zero how to commentate on chess games this is as good as it gets.
Very nice analysis by Danny 👍👍
amazing anaalysis Daniel
"We also report the win/draw/loss results of 100 game AlphaZero vs.
Stockfish matches starting from each opening, as either white (w) or black (b), from AlphaZero’s
perspective."
From The Paper: Mastering Chess and Shogi by Self-Play with a
General Reinforcement Learning Algorithm by David Silver, Thomas Hubert,
Julian Schrittwieser et all, page 6 / arxiv.org/pdf/1712.01815.pdf
So there was total 1200 games played between AlphaZero and Stockfish with stats:
Total games: w 242/353/5, b 48/533/1
Overall percentage: w 40.3/58.8/0.8, b 8.0/88.8/3.2
(copy pasted directly from the paper)
Why people talk about only 100 games played and that Stockfish didn't win any games?
SF only won games that had pre-determined opening lines, (in the first match).
Awesome extended analysis! Totally, quells the idea that Stockfish engine was somehow neutered with limited parameters.
Wow!!!!!! Thanks Danny, great analysis, much appreciated!
Excellent danny
the weakness of stockfish looks like its evaluation of material
Taking a material loss, in order to get a positional advantage for mating pressure that is leveraged back into a material advantage. I want to introduce strong ai to other turn based games
we need more of these videos.
Thank you Daniel.
Great analysis of the games. Alphazero plays more like a pure tactician and recognizing and aiming for patterns on the board and using the computer's access to book moves and programmed lines(memorized lines if it was a human player) against it. Alphazero plays more like a human by using long term strategy and calculation, humans just have the flaw of habit moves (as you mention saying "normally this happens") because of time and can only calculate so far in the time given during games. In these games AZ calculated long term strategy within a minute per move...That would be roughly a 45-60 minute game.
Make a TH-cam channel of your own. you teach and analyse games very well.
he does have one
Valdemar Paulus Antreasyan Hirshals
What is the channel name
th-cam.com/users/ACEChess
srikanth t 18:36
16:18 my old Houdini 3 suggests that Rd8 was a blunder from Stockfish and Re5 would have been a better move.
Mikhail Tal would be proud
There's a big question whether AlphaZero was trained playing only against itself. Maybe they trained it against Stockfish as well. To find its weaknesses. Making it a Stockfish-killer. But maybe it would lose to Comodo.
Give Alphazero a full month of self learning. it's elo will be like 10000
Sanjay Goyal probably not; the neural net reaches saturation and learning levels off.
Still would beat every human and engine, so I guess he's right with his elo guess
cbr -- no, I think that 10000 number is wildly wrong. Every 100 points of ELO represents about a 2 to 1 win ratio for the stronger player. Go take a look at the Deepmind paper: arxiv.org/pdf/1712.01815.pdf -- you'll see the ELO gain based on more training just levels off at about one quarter of the way into the four hour training period. (Remember, this 4 hours of training *is on thousands of TPU cores* so that's a lot of training steps!) More time would have _maybe_ given a little more ELO, I think. To get any significant gains, the AlphaZero algorithm would probably need to be enhanced to allow some amount of incorporation of alpha-beta search, or other traditional engine techniques into the play style, rather than just the MCTS (Monte Carlo Tree Search). Again, from the paper: "None of the techniques described in this section are used by AlphaZero. It is likely that some of these techniques could further improve the performance of AlphaZero; however, we have focused on a pure self-play reinforcement learning approach and leave these extensions for future research."
Max elo is 3700.I think
I think past a certain point, chess will essentially be solved, and the AI will be able to either draw every game, or the win/loss will only depend on who is white. In either case, ELO won't work anymore. So I don't believe it is possible to reach ELO 10000.
5:25 lol'd so hard. "... world will never know....like the center of a tootsie roll pop." AKA " Like the number of licks it takes to get to the center of a tootsie roll pop..." idk why it's just how his comments seem to be, like intentional (or unintentional) but just a few degrees of freedom between his reality and the reality he remembers, the real one. It's genius!
Danny, 34. Rd8 was a blunder by SF. Better was bc,a5,a6,Kf7 stays 0.0. Why did SF not choose that path?
after i watch those videos i always forget how did the opening went to the middlegame
especially after the "e4, e5, Nf3, Nc6, Bc4, Bc5..." openings
Although stockfish was restricted in some ways, I still find it amazing that it only took 4 hours for this neural network to win against computer, which is basically fed with information that human gathered 1500 years...
Ok.now im subbed to both chess.com and your TH-cam channel. Like a good Russian schoolboy.
really love these analysis!!!
Wonderful analysis! Congratulations for the great work. So far we've just seen AlphaZero winning with White... but how does AlphaZero play as Black?
Loved the series!
So Tal's approach to playing Chess is the right approach after all ...
Last moves were déjà vu
Really great analysis, thanks a lot.
it's important to point out that AlphaZero was on a far more powerful machine than Stockfish, meaning Stockfish couldn't anaylise as far ahead and that they were given 1 min/move, rather than a classical time control. Something Stockfish is good at is knowing when to spend the time in critical positions. Not trying to take anything away from AZ, it's clearly becoming a force to be reckoned with.
A0 80 Knps, SF 70,000 Knps. Any questions?
what does that mean?
at 6:44 instead bishop c8, why not move rook to h8 and then king back to f8? At least, the king is somewhat secured and protected by pawn, bishop and other pieces.
I appreciate these videos very much! Love your sense of humor too, Mr. Rensch. But I'm a patzer and sometimes you leave me hanging and I'm frustrated. You analyzed Kg7 (which was not played) after White went h4. Your analysis ended with White playing Qe5 and "forking" Black's bishops but then Black's obvious reply (which you did not mention, much less analyze) is Be6, protecting both bishops. What might White do then? What ideas are in this position for White? I mean, I'm sure there must be something because according to the bar at the left of the screen, Stockfish feel inferior as Black after Qe5. But poor patzer me, I'm at a loss for how White might proceed. Can you give a clue please?
5:40 why not Bf6? Am I missing something(Nh6+ Kg-h8....or after 1.Bf6 2.Nxg7 Bxg7 3.Bh7 for white and then just Qd8-f6)?
So nasty, would love to see a fully powered stockfish with an opening book etc
Man good job with the analysis, thank you
I think the winning idea after 26 Qh1 h5 27 Be4 Q:a2 28 Bd4 Bd6 29 h4 hg4 is the incredible 30 c4!!, when if 30...Q:c4 then 31 B:f5+ mates
Such a good video, can’t say it enough. I’m sitting here at 14:32 trying to think what was the next amazing move A0 played, as we the viewers are challenged to do ... and I’ve already seen this game! Ok, think, think ...
love this! Anyone know why the other 90 games weren't released? What would be the result of a 100 game match if AlphaZero plays itself?
Excellent analysis
This is what gingergm played, Williams vs Wells at the Bumrattey (sic) tournament.
Fantastic entertaining commentary!
I think Stockfish have a problem of database in this opening, from game many games won alpha zero that way, that's is why the rest of the won game have not bin made public. Alpha zero discovered a hole in Stockfish
what are you smoking, kid?
In your line after 26 Qh1 h5!? 27 Be4 Q:a2 28 Bd4 Bd6 29 g4 h:g4 30 Bf6, what if 30...Bh2+! which seems to draw
that C4!!!!
exactly , just as amazing as Qh1
AlphaZero was like "I guess I'll play C4 coz it's explosive lol"