Botvinnik was once quoted : If Tal offers a sacrificial, take then think. If I offer a sacrifice, think then take. If Petrosian offers a sacrifice, resign.
Yes! Botvinnik was more human than he appeared beneath that stern facade. The Petrosian quote is indeed brilliant. Petrosian wasn't known for his sacrificial play. He was probably rather underestimated. He was a very strong player. His win in the second game of the candidates match against Fischer in 1971, (or was it early 72?) was hugely impressive, considering what Fischer had done to Taimanov and Larsen.
This would require that Stockfish has attained extraordinary insight into Chess, so that it can see its position as black is SO hopeless after the initial White move that it can't even expect to draw. Without this insight, it would logically try for the draw. It is very possible neither humans nor AI will ever achieve this.
I think alphazero also thinks that winning chess simply involves paralysing the opponents queens bishop - mostly when playing white, rendering the light-squared bishop useless, but also as black treating the dark-squared bishop the same way. This might be distorted by so many queens indian games, but it seems universal in the games we have seen. Though it doesn't seem to have the same problems liberating and using its own queens bishop.
thisnicklldo you're kinda right kinda wrong. It is more about the difference between how stockfish and alphazero values the activity of their pieces. The thing is it isn't all too simple to determine the impact of a piece's lack of activity, how bad it is depends heavily on the position. which I believe alphazero understands better than stockfish because a machine learning approach lends itself to discovering and "understanding" nuances better
Well, I'm only partly serious. I don't think the evidence of what we have seen is enough to imagine that alphazero has a particular view about a particular piece, in general. Though with a neural net, it's not impossible - its learning may well have ended up with neurons related to the activity of individual pieces, but probably all of them or none of them is likeliest. But maybe the queens bishop really is the easiest piece to limit, and alphazero knows it. I'm not sure you are saying anything different to me - of course the relevance of the activity of a particular piece depends on the position, but alphazero seems to be generating positions in which the low activity of the queens bishop is relevant. It would have to be like that wouldn't it? Else it wouldn't bother trying, and would never have learned that it helps it to win. In the much less complex neural networks I have played with, this could prove to be a problem, as they tend to be self-reinforcing (well, mine did). So there must be some risk that alphazero progressively gets more and more concentrated on generating positions where the activity of the queens bishop can be limited, and where it matters (this is a feature of 'the local minimum problem', which alphazero might be too sophisticated to suffer from, for all I know). This could, just possibly, prove a weakness in the future, since perhaps it does not learn the value of a similar strategy against the kings rook, or whatever, or at least, not as well. I absolutely would expect it to have specific views on specific pieces and structures, unlike a brute force engine like stockfish which knows nothing except the outcome of all the deep sequences it tests.
I don't think alpha zerois evaluating piece values individually, rather the entire position for patterns. It understands the value of pinning, forking, blockading activity, sacrificing, etc... of all pieces relative to all other pieces to understand from a fundamentally positional foundation. I don't think it is targeting the queen's bishop, but rather this occurrence is a consequence of the pitfalls of stockfish's specific style of play.
Maybe it's as general as you say, but don't understimate the ability of a neural network to get quite specific. Your vision may generally interpret lines, areas, intersections etc and learn many objects in a generic object-neutral way, such as cars, pencils etc etc, but it also respond quite specifically and non-generically to faces, and a baby specifically learns its mothers face very early indeed. I'm not suggesting that alphazero has the general power or structure of the human brain, but it is a very, very big network, with many layers and will operate with much more than just forks, pins and other high-level concepts. One of the points of neural networks is that they are simultaneously working on both high level categories and low-level detail.
I agree it must be operating its value system on many deeper more dynamic layers of concepts than the surface level. Forks and pins may be a second layer concept relating 2 or more pieces at once.. General activity might actually be represented as a much deeper layer, and how many open squares there are available for your pieces in the future. Everything is simultaneous and we can never know the exact purpose of any single node or function. 1 bishop is worth 1 bishop. Sacrificing it is up to the remaining layers to determine the entire value structure of the whole position once every piece on the board has been evaluated in reference, through all of alphazeros layers. My point was, the bishop was never assigned a value. The next move was assigned a value based on the whole board.
@9:40 why not capture pawn on e6? Forces check. Looks like if it tries to take pawn on g7 afterwards it will lose the knight due to a pin, but it could always just retreat after taking that pawn for free and be right back to the same position instead, no?
7:51 "for a human this might be a bit troublesome to convert into a victory, but not for Alpha, and Stockfish knows this...." Very interesting statement....I'm sure that there is no way for Stockfish to know that it's playing another computer, but are you saying that Stockfish might accept this endgame if it "thought" (computed) that the opponent was a weak player..? Does it have such alternatives built in..? Or were you anthropomorphizing...? :-D Great channel, keep it up, I'm addicted... B)
Stockfish does have a "contempt" setting for how it estimates its opponent. When playing against a human or a weaker engine, a higher contempt setting will let it win faster, while the standard contempt setting has it assume its opponent is "another one of itself" and will make it play conservatively and take less risks. Stockfish does not have a system to estimate its opponent as the game progresses, this is a setting you need to choose manually. Google "stockfish contempt"
Hmm interesting, thanks......this whole endeavour is so fascinating, now I wonder if a different contempt setting would make for a better chance for Stockfish, or just even more of a beating... :-D I hope they release the rest of the games (and play some more...) soon, Agadmator will be analysing them for months... :-)
Question: at 9:40 after ...Rc1, why can’t white play Nxe6+ then after ...Ke7, Nxg7. Isn’t this completely winning for white? The black bishop is pinned and can’t defend e6. Why does white play Ra8? Thanks!
after the Ke7 by black, the black bishop is no longer pinned so it can defend e6. Also even if the bishop was pinned to the king (it isn't), you still would not be able to do Nxg7 because the knight is pinned as well.
Last December/January I watched several videos about the Alpha Zero-Stockfish match, some better than others. Yet somehow this one evaded my attention until just now. It's an excellent video, up to your usual high standard. It happens to have been worth watching purely for the Ian McDonald quote which, I agree, is very powerful. I shall read the book, now (I just downloaded it to my Kindle). Of course the game (which I had seen before) is also very powerful, and your analysis succeeds, as always, in being both insightful and succinct. Thank you.
So... what is the refutation? and idk if alpha zero does probability like when it plays go but if so whats the probability of winning from such position?
As a note, we can see the latest versions of Leela chess 0 (the other self learning ai) are now also "refuting" E4. She also will always play the Berlin against e4. This leads me to believe that BOTH self learning engines believe that the Berlin is indeed the refutation of E4. Now this is unlikely to be significant in terms of human over the board chess, but it is indeed, very interesting. Leela now seems to be enjoying playing the Catalan, (Leela ID42916 as of May the first, 2019) as well as other d4 openings. However it should be said that it will take much longer to get a conclusive outcome as Leela is far from being "completed"
it's probably more about the sequence of the opening moves. Depending on how your oponent responds to your move, then you move to counter that makes the better response
opening sequences leads to distinct middle game advantages and a slight variation might transpose into other lines if you play 1 e4 as white, black could play the scandinavian d5 or other lines that A0 might consider equal..
There are likely many ways to draw, but in my research, black conspicuously can mess up the Sicilian. Many engine vs engine matches I played resulted in a loss for black when the Sicilian was played by my opponent. Yet, when I played it, the game was always a draw. This could be because the opponent had a second rate engine, but more likely, the opponent chose a bad opening sequence. The most important factor when choosing the best opening sequence is to see the average rating of the player who chose each move. Secondarily, the quantity of players that chose a move should be taken into account. Although, a move overwhelmingly preferred by masters, is better than one with a slightly higher average rating. In my experience, most of the time a player lost after playing the Sicilian, it was because they chose the most popular move instead of the one with the highest average rating in a masters database.
i know that but you can use any opening u want but if you dont make the best move in every variation it dontmean nothing cuz you cant keep forcing that line
I think AlphaZero's move order also tells us 2 more things: 1. Best way to meet the french is Nc3 2. Winawer might be better for black than Nf6 (otherwise AlphaZero would go 2.e4 which is more logical than 2.Nc3 as it occupies the center immediately) By using this very specific move order Alpha got the Steinitz variation which it may consider the best way for white to play against the french.
Hello agadmator, is there an app of Alpha zero in play store? I only see stockfish, where did you get the Alpha zero chess engine? I'm curious, please reply..;)
I think it can be bcause that will pin the knight to the king.. making the knight unable to move. So next turn stockfish king just moves to c7 and you have to move the rook away thus enabling the bishop to take the knight with check
That's actually quite an interesting position. I know it would become an illegal move for the knight to move after capturing the pawn on e6, but should it? The bishop is pinned. It makes more sense that a pinned piece shouldn't be able to put the opponent's king in check, until it's unpinned. It would also be more confusing.
@@hishamamir8186 king cant go c7 he has to go e7...then the rock cant take the bishop ...and the king takes knight cuz its pinned ...so alpha loses a piece
THANK.YOU! I noticed too and just realized I asked the same question once before but no1 noticed. Had to go through all the comments to look for any1 else seeing it. I cant really see how black would get out of the position, though I'm not stockfish or AG..
AlphaZero has liberated the sacrificial players like me by letting us know that Stockfish is not quite the all-knowing Oracle we had presumed. Many lines which should be theoretically sound (particularly positional sacrifices) are dismissed as a forced loss a pawn down when it's now clear that up to fifteen points can be sacrified for a king hunt. This opens up everything theoretically, creating a true golden age of chess, with the "spellchecking" of the computers now tapered by the uncertainty of the pre-computer era, where we never really knew if our moves were best. Get ready for some of the wildest chess in history from humans, who now know how to beat Stockfish.
Alpha is not telling us that e4 is bad. Alpha is telling us that based on positioning and how you started the game, if you take what looks like "free" material, even just a pawn you will lose the game
No it doesnt, because alpha zero ended with 2 pawns, the one alpha zero took, and the one he offered gave him 1 pawn. The point Im making is that if you play like stockfish, then you will not understand how Alpha Zero plays. Im making the point that even if it seems like there are lots of variables in chess based on positioning, there are actually very limited variables. Notice how halfway through alpha repeats the same moves over and over until stockfish does something different. Its because alpha understands that he will not give up the positioning for anything. You need to think like Alpha zero to understand, you are thinking like stockfish. The fact that alpha both took AND offered a pawn, and is playing against stockfish, an engine that doesnt think like Alpha zero, actually backs up my assertion. Get on alpha zeros level
It's amazing how good Alpha Zero managed to get itself at chess. Whoever thought that playing thousands of games against your own powerful processing power and placing the results into a ram, creating a library of the good and bad could do that? I think it's fascinating how it completely annihilates the likes of stockfish and has complete grasp of complicated endgame situations requiring scores of moves. It plays them like a person MIGHT play 3 or 4 move scenarios. Not your usual boring computer that's for sure. It feels so absolute that you are often on the edge of your seat and stopping the game to look at what's happening
Best chess channel on RUclips by far!! I have abandoned most of the others because they get into the game and talk about 100 possible options instead of getting on with the game. Agadmator mentions maybe a couple of lines when appropriate and otherwise presents the game as it was. MUCH better than watching a half hour of "coulda, woulda, shoulda" rhetoric. Hope my broke ass gets some money so I can contribute at some point!
Why in minute 9:40 the white knight didnt grabbed the pawn on e6 if the black bishop is pin and after grabbing the pawn it comes with a check??? Help pls
+Allofthegoodnameswheretaken Yesterday I prepared lunch for my girlfriend, went to pick her up at work, and when we got back, we saw Medo had devoured lunch all by himself... somehow he grabbed it from the table... But yes, usually, he's a good boy :)
I cut my teach programing chess engines while in highschool in the 90's by 2000 when I had finish my education I had started work on a theoretical AI project using hidden Markov models and Bayesian tables. I made a life change to focus more on family life, but reviewing these games makes me believe that the learning mechanic's behind Alpha zero is based on a combination of those models. Deep Mind I support you. -David_Dennis
I haven't seen anything concrete about Alpha Zero the chess engine, but there is a lot of material about AlphaGo Zero the Go engine. I imagine they are very similar, if not in fact the exact same architecture.
it's neural networks. These 80-x machine learning stuff isn't suitable for AI. Throw away and forget about it. Every time u'll end up with hard-coded evaluating function with only pieces cost undefined. It's not alphazero and any modern AI approach.
There's a difference in calculating all possibilities and predicting outcomes. If I was going to speculate I would guess the engine to work similar to this. Human players will often visualize a winning position based on how the current position is. More than likely what Alpha zeros is doing is looking ahead for winning positions then comparing them to the current position and calculating the probability of reaching that position. An AI that uses probability is an AI that has the ability to produce a plan. Something as complex a deepmind will have several layers that are checked against each but the real Holy Grail is in the layer that manages probability. The learning aspect of it or the "Deep neural network" or "modelled after the human brain" is just another way a saying the computer learns like we do. So how do we learn? Here's a thought experiment, imagine the lights off in your house. While you were sleeping someone moved the door knob to your bedroom up 3 inches. You get up in the middle of the night to pee. When you reach for the door knob you notice it isn't there. Based on previous experience your brain predicted and moved your hand the most probably location of the knob. This seems a bit silly but it is based on probability. Our brains work and learn largely by prediction. Alpha zero is likely working backwards from similar positions it would like to achieve that it feels is probable. This may seem like a very subtle change but think about how much it influences chess play among humans. -David_Dennis
It has already changed the legendary game of GO, which it was created to tackle, called originally Alpha GO Zero. But it is a general purpose algorithm that can be applied to any board game with complete information.
I think it would be even bigger improvement to partial information games. In those using just fied math and other algorithms to find good move candidates can be even harder. But neural network would potentially provide much better heuristics than human developers have figured out so far.
I've played against the Berlin many times but I never go into mainlines, always anti-Berlin systems. I think the Berlin is an opening not really playable by anyone that isn't at least master level strength as the opening seems to achieve one of two results where either white will win the game or it will end in a draw. This is a tough thing for weaker players to grasp psychologically (me included) as never having any winning chances means that there is going to be relentless pressure placed upon you during the entirety of the game. Often my opponent will simply blunder or allow a strategically winning idea to unfold because white can just play comfortably and naturally while black is always sitting back and defending.
Shreyas Bhatt e4 e5 Nf3 Nc6 Bb5 Nf6 d3 is the usual but there is ... Nf6 O-O Nxe4 Re1 ... Nf6 O-O Nxe4 Qe2 ...Nf6 Nc3 transposes into a Four Knights Spanish ... Nf6 d4 exd4 e5 can lead to some exciting positions as well and this is just to name a few. It is a very rich opening ripe for exploration and experimentation 🙂
CLUCKY CLUCK I am not saying it is impossible to play for a win. It is just not an opening particularly rich with black counterplay if white knows what they are doing.
There are also some chances for Black in the Berlin to win. Even Anand demonstrated in his 11th game in the World Championship versus Carlsen that his approach (...b5) could have lead to very interesting play with 3 possible results. Of course you're right with master level strength! :D
When my dad taught me how to play chess, he told me E4 was the strongest starting move for white....until you're a master! he said, when you get better, start with D4..and yeah pushing with C4 for black later in the french defense never worked out for me either so i stopped doing that
For anyone interested, the "AlphaZero was trained for 4 hours" bit isn't quite correct. AlphaZero reached Stockfish's level after 4 hours, but these matches didn't happen until AlphaZero trained for 3 days.
I thought the same thing about 3 days, but it turns out that it was the previous Go program that was trained for three days. The original paper is clumsily worded in this regard, but table S3 from the paper states 9 hours of training time. However, looking at the graph of improvement (Figure 1), from the 4-hour mark to the 9-hour mark its progress nearly flattened out and it gained very little. We really we need more complete information. I hope it's forthcoming. Seeing all the games would be also nice rather than just the cherry-picked ones.
Stupid question: alpha zero trained itself to play, did that learning process stop before the matchup with stockfish? Was it allowed to keep learning during it's 1200 game tournament? Or was it's learning ability frozen in time to avoid creating a database of games and becoming like stockfish? If it was allowed to keep learning, is there a way to see if zeros rating improves if we compare say the first 100 games to the last 100 games?
Thank God for human players, all this electronic/digital/technical whatever is anathema to the spirit of 'human' chess. And I will never stop playing e4 either.
I have a question: Can pure brute Force still beat AlphaZero? Lets say a nation state gave Stockfish 100,000x the CPU power and challenged AlphaZero in 2020. Who wins??
S L CPU power doesnt mean anything unless SF code is scaled properly to use all those extra cores. At some point it simply.doesnt make a difference how many CPU's SF has. You would have to ask SF author about CPU limits. And how many cores he can realistically code for.
Dave van Laren - When most say "CPU power", they mean the time it takes to complete a task, not the actual voltage applied across the chip. Well, kasparov9 is right, that time can vary greatly as the amount of cores, threads, and cache changes. Most applications aren't programmed to make efficient use of more than 4 cores. If the application is designed for 2 cores, having a quad or octa-core chip would make little difference.
I have a question. I may be missing something obvious, but at 9.40, wouldn't 'Nxe6+' grab a pawn for A0 for free? Judging from the later moves, I don't see how A0 would miss something if it grabbed than pawn and then returned the Knight to f4 and continued as it did for the rest of the game or similarly. Unless it already knew it would win the position by that move, in which case ok. :P
I'm not sure that AlphaZero is the greatest entity. It was a test experiment to prove that AlphaZero can play like a superhuman, not that AlphaZero is better than Stockfish. The conditions of the match were not fair for Stockfish to consider a real winner.
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.
AlphaZero runs in TPU, that's hardware specifical made for neural network. Stockfish needs proper testing and adjustment to run with 64 cores as apparently it run. However I'm not saying that Stockfish is better that AlphaZero, just that the match isn't a conclusion.
Stockfish usually has access to some sort of database of games and openings, and did not for the purposes of the match, while Alpha did!! It could be totally unfair unless I’m somehow mistaken
Stockfish usually has access to some sort of database of games and openings, and did not for the purposes of the match, while Alpha did!! It could be totally unfair unless I’m somehow mistaken
lol @stephen0793 aplha zero is a deep neural network based on General Reinforcement leaning algorithm its doesn't needs database to work with it learns from itself
Their is no chance for u to say "a better move was"
There*
They're*
Shut up
👍
Lol
Stockfish: “1. e4”
AlphaZero: ‘LOL - mate in 176.”
Lol😂😂😂
Humorous
hahahahahaha
lolol
hahahaahaha
Botvinnik was once quoted : If Tal offers a sacrificial, take then think. If I offer a sacrifice, think then take. If Petrosian offers a sacrifice, resign.
Haha
Love it!
Wasn't Tal quoted saying "Some sacrifices are sound. The rest are mine."
@@somebodys7404 I don't know whether he did or not, but I like the quote. Botvinnik's quote about Petrosian is brilliant.
Yes! Botvinnik was more human than he appeared beneath that stern facade. The Petrosian quote is indeed brilliant. Petrosian wasn't known for his sacrificial play. He was probably rather underestimated. He was a very strong player. His win in the second game of the candidates match against Fischer in 1971, (or was it early 72?) was hugely impressive, considering what Fischer had done to Taimanov and Larsen.
*There will come a time* when AlphaZero makes the first move and Stockfish immediately resigns.
This would require that Stockfish has attained extraordinary insight into Chess, so that it can see its position as black is SO hopeless after the initial White move that it can't even expect to draw. Without this insight, it would logically try for the draw. It is very possible neither humans nor AI will ever achieve this.
Nick Lamb Hey now, he wasnt being that serious lmao. Its just a joke xD
funny joke but chess is def a draw
asdfjk; only when both players are kinda of the same strength :)
Yes if you tell Stockfish that it is playing AlphaZero.
1:50 you might want to consider an exorcist for your dog.
hhahaha he should
I knew someone else seen that
Rofl
Lol I think that's just the dog scratching himself on the couch
I laugh too hard on this
1:46 After dog to K9, it was at this point that Agadmator had lost his mind.
"After 1. e4?!, Stockfish is already at a disadvantage"
stockfish didnt have opening theory on when it played against A0. So really i didnt know what move is considered as best for specific openings.
The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening.
Dan Kelly but that's because alpha played better, not because the opening is bad.
Disagree. 8... a6 is still opening theory, but it's a bad move. When I played 1. e4, the only move I feared was 1... e5 into the Berlin Defense.
i never thought i'd see the phrase "1. e4?!"
Stockfish:"What a shitty game, I can't find out what my mistake was!"
Alpha Zero:"Your first move..."
I think alphazero also thinks that winning chess simply involves paralysing the opponents queens bishop - mostly when playing white, rendering the light-squared bishop useless, but also as black treating the dark-squared bishop the same way. This might be distorted by so many queens indian games, but it seems universal in the games we have seen. Though it doesn't seem to have the same problems liberating and using its own queens bishop.
thisnicklldo you're kinda right kinda wrong. It is more about the difference between how stockfish and alphazero values the activity of their pieces. The thing is it isn't all too simple to determine the impact of a piece's lack of activity, how bad it is depends heavily on the position. which I believe alphazero understands better than stockfish because a machine learning approach lends itself to discovering and "understanding" nuances better
Well, I'm only partly serious. I don't think the evidence of what we have seen is enough to imagine that alphazero has a particular view about a particular piece, in general. Though with a neural net, it's not impossible - its learning may well have ended up with neurons related to the activity of individual pieces, but probably all of them or none of them is likeliest. But maybe the queens bishop really is the easiest piece to limit, and alphazero knows it. I'm not sure you are saying anything different to me - of course the relevance of the activity of a particular piece depends on the position, but alphazero seems to be generating positions in which the low activity of the queens bishop is relevant. It would have to be like that wouldn't it? Else it wouldn't bother trying, and would never have learned that it helps it to win.
In the much less complex neural networks I have played with, this could prove to be a problem, as they tend to be self-reinforcing (well, mine did). So there must be some risk that alphazero progressively gets more and more concentrated on generating positions where the activity of the queens bishop can be limited, and where it matters (this is a feature of 'the local minimum problem', which alphazero might be too sophisticated to suffer from, for all I know). This could, just possibly, prove a weakness in the future, since perhaps it does not learn the value of a similar strategy against the kings rook, or whatever, or at least, not as well. I absolutely would expect it to have specific views on specific pieces and structures, unlike a brute force engine like stockfish which knows nothing except the outcome of all the deep sequences it tests.
I don't think alpha zerois evaluating piece values individually, rather the entire position for patterns. It understands the value of pinning, forking, blockading activity, sacrificing, etc... of all pieces relative to all other pieces to understand from a fundamentally positional foundation. I don't think it is targeting the queen's bishop, but rather this occurrence is a consequence of the pitfalls of stockfish's specific style of play.
Maybe it's as general as you say, but don't understimate the ability of a neural network to get quite specific. Your vision may generally interpret lines, areas, intersections etc and learn many objects in a generic object-neutral way, such as cars, pencils etc etc, but it also respond quite specifically and non-generically to faces, and a baby specifically learns its mothers face very early indeed. I'm not suggesting that alphazero has the general power or structure of the human brain, but it is a very, very big network, with many layers and will operate with much more than just forks, pins and other high-level concepts. One of the points of neural networks is that they are simultaneously working on both high level categories and low-level detail.
I agree it must be operating its value system on many deeper more dynamic layers of concepts than the surface level. Forks and pins may be a second layer concept relating 2 or more pieces at once.. General activity might actually be represented as a much deeper layer, and how many open squares there are available for your pieces in the future. Everything is simultaneous and we can never know the exact purpose of any single node or function.
1 bishop is worth 1 bishop. Sacrificing it is up to the remaining layers to determine the entire value structure of the whole position once every piece on the board has been evaluated in reference, through all of alphazeros layers.
My point was, the bishop was never assigned a value. The next move was assigned a value based on the whole board.
The computers have convinced us they're playing chess when they're really just using it to communicate and planning to monopolize time travel!
Stockfish: Terminator, the good one
AlphaZero: Skynet
@9:40 why not capture pawn on e6? Forces check. Looks like if it tries to take pawn on g7 afterwards it will lose the knight due to a pin, but it could always just retreat after taking that pawn for free and be right back to the same position instead, no?
"stockfish resigns the game...: idk whys that funny
Its surreal and cathartic
Hmm, when I think of scifi movies, I always think the AI as being on the same side. Here we have one defeating another.
@@RobertCleggRC Well if AI truly goes evil then the obvious solution is to make another AI to fight it inside a giant robot or something
Engines can't resign
Your dog going nuts in the background is great 😂😂
7:51 "for a human this might be a bit troublesome to convert into a victory, but not for Alpha, and Stockfish knows this...."
Very interesting statement....I'm sure that there is no way for Stockfish to know that it's playing another computer, but are you saying that Stockfish might accept this endgame if it "thought" (computed) that the opponent was a weak player..? Does it have such alternatives built in..? Or were you anthropomorphizing...? :-D
Great channel, keep it up, I'm addicted... B)
+andybroon2007 Let's go with anthropomorphizing :)
Stockfish does have a "contempt" setting for how it estimates its opponent. When playing against a human or a weaker engine, a higher contempt setting will let it win faster, while the standard contempt setting has it assume its opponent is "another one of itself" and will make it play conservatively and take less risks.
Stockfish does not have a system to estimate its opponent as the game progresses, this is a setting you need to choose manually.
Google "stockfish contempt"
Hmm interesting, thanks......this whole endeavour is so fascinating, now I wonder if a different contempt setting would make for a better chance for Stockfish, or just even more of a beating... :-D I hope they release the rest of the games (and play some more...) soon, Agadmator will be analysing them for months... :-)
Anthrpomorphizing? Jeez. Just say 'personifying', that's much more regular everyday English.
aw don't be like that, at least you've learnt a new word today......if only you'd learnt how to spell it too.... :-(
Thanks for another great AI game. Please do more.
"I knew it!" ~Richard Réti, from the grave.
1:42 gotta love the dog in the background
Stockfish: Plays a move
Alpha 0: You made the biggest mistake of your life
1. I come.
2. I upvote so I don't forget after.
3. I start watching.
Same here bro
How do we updoot here?
This ain’t reddit bud, it’s called a like (btw I do the same)
1. I start watching
2. I come
👁👁🐝🐝🐝
What if you wanted to play 1. E4
But god said: Mate in 176.
176 likes
underrated comment
Lmfaoooo
"and Stockfish knows this."
4:30 you named him Alphie 😂
Question: at 9:40 after ...Rc1, why can’t white play Nxe6+ then after ...Ke7, Nxg7. Isn’t this completely winning for white? The black bishop is pinned and can’t defend e6. Why does white play Ra8?
Thanks!
after the Ke7 by black, the black bishop is no longer pinned so it can defend e6. Also even if the bishop was pinned to the king (it isn't), you still would not be able to do Nxg7 because the knight is pinned as well.
Last December/January I watched several videos about the Alpha Zero-Stockfish match, some better than others.
Yet somehow this one evaded my attention until just now. It's an excellent video, up to your usual high standard. It happens to have been worth watching purely for the Ian McDonald quote which, I agree, is very powerful.
I shall read the book, now (I just downloaded it to my Kindle).
Of course the game (which I had seen before) is also very powerful, and your analysis succeeds, as always, in being both insightful and succinct.
Thank you.
I love your dog in the background of all your videos. One of the many things that makes this channel great.
omg I fucking love ur dog I always watch him when he's doing something in the background :D
I'm disapointed that Agadmator never did present his dog in a video
He looks like he's having a seizure
1:42 lol
@@africanfartingfrog he's a happy dog. 🐕
thanks for another great video. Be sure to say hello to your dog xD, he usually gets my attention and the great chess games of course.
His dog is so cute
@1:50 puppy having an absolute blast behind you @Agadmator - thanks for producing more quality chess content, definitely enjoying the coverage of a0!
Is it swimming?? lol
He's self learning :P
I didn't even know that I am interested in chess, but I'm binge watching your videos, great content!
So... what is the refutation? and idk if alpha zero does probability like when it plays go but if so whats the probability of winning from such position?
AlphaZero:
Stockfish: **starts sweating**
Just wanted to let you know that I’ve seen a ton of your videos and I really enjoy them! I love how you explain the thought process behind each move
Never knew one quote could keep me awake at night
Bruh what's your dog doing at 1:55 lmao
scratching its back. u never had a dog did u?
Its in pain
I think it's dreaming. Dogs often do weird moves in their sleep. It's probably dreaming it's being chased by an Alpha (Zero) Dog.
I love it when that dog does his upside down paddle moves
As a note, we can see the latest versions of Leela chess 0 (the other self learning ai) are now also "refuting" E4. She also will always play the Berlin against e4. This leads me to believe that BOTH self learning engines believe that the Berlin is indeed the refutation of E4. Now this is unlikely to be significant in terms of human over the board chess, but it is indeed, very interesting. Leela now seems to be enjoying playing the Catalan, (Leela ID42916 as of May the first, 2019) as well as other d4 openings. However it should be said that it will take much longer to get a conclusive outcome as Leela is far from being "completed"
I don't think the first few moves changes the course of the game...I believe AlphaZero out manoeuvres Stockfish
its never about the opening its just making the best move everytime
it's probably more about the sequence of the opening moves. Depending on how your oponent responds to your move, then you move to counter that makes the better response
opening sequences leads to distinct middle game advantages and a slight variation might transpose into other lines if you play 1 e4 as white, black could play the scandinavian d5 or other lines that A0 might consider equal..
There are likely many ways to draw, but in my research, black conspicuously can mess up the Sicilian. Many engine vs engine matches I played resulted in a loss for black when the Sicilian was played by my opponent. Yet, when I played it, the game was always a draw. This could be because the opponent had a second rate engine, but more likely, the opponent chose a bad opening sequence.
The most important factor when choosing the best opening sequence is to see the average rating of the player who chose each move. Secondarily, the quantity of players that chose a move should be taken into account. Although, a move overwhelmingly preferred by masters, is better than one with a slightly higher average rating.
In my experience, most of the time a player lost after playing the Sicilian, it was because they chose the most popular move instead of the one with the highest average rating in a masters database.
i know that but you can use any opening u want but if you dont make the best move in every variation it dontmean nothing cuz you cant keep forcing that line
Love your dog and how he plays while your taping. Thank you for helping me understanding chess more.
I think AlphaZero's move order also tells us 2 more things:
1. Best way to meet the french is Nc3
2. Winawer might be better for black than Nf6 (otherwise AlphaZero would go 2.e4 which is more logical than 2.Nc3 as it occupies the center immediately)
By using this very specific move order Alpha got the Steinitz variation which it may consider the best way for white to play against the french.
Makes sense to me.
Hello agadmator, is there an app of Alpha zero in play store? I only see stockfish, where did you get the Alpha zero chess engine? I'm curious, please reply..;)
Why Alpha didnt captured the pawn on e6???? 9:40
Did i miss something?
I think it can be bcause that will pin the knight to the king.. making the knight unable to move. So next turn stockfish king just moves to c7 and you have to move the rook away thus enabling the bishop to take the knight with check
That's actually quite an interesting position. I know it would become an illegal move for the knight to move after capturing the pawn on e6, but should it? The bishop is pinned. It makes more sense that a pinned piece shouldn't be able to put the opponent's king in check, until it's unpinned. It would also be more confusing.
@@hishamamir8186 king cant go c7 he has to go e7...then the rock cant take the bishop ...and the king takes knight cuz its pinned ...so alpha loses a piece
@@mayedalshamsi a pinned piece can check the king for sure...if u think about it...like u could capture kings it would make more sense to u
at 9:40. why doesn't the knight take pawn on E6 to check the king and fork. then capture G7. The bishop is pinned so can't capture the knight
Watched it again, just a brilliant AI game!!!! More please!!!
why doesn't alpha take the pawn on e6 at 9:39 in the video and put the king in check while forking the g7 pawn? the bishop was pinned to the king
THANK.YOU! I noticed too and just realized I asked the same question once before but no1 noticed. Had to go through all the comments to look for any1 else seeing it.
I cant really see how black would get out of the position, though I'm not stockfish or AG..
I found it!!!! Its cause you will lose either the knight or rook of you're not careful.
Ne6, kd7 and your knight is pinned (bishop)
AlphaZero has liberated the sacrificial players like me by letting us know that Stockfish is not quite the all-knowing Oracle we had presumed. Many lines which should be theoretically sound (particularly positional sacrifices) are dismissed as a forced loss a pawn down when it's now clear that up to fifteen points can be sacrified for a king hunt. This opens up everything theoretically, creating a true golden age of chess, with the "spellchecking" of the computers now tapered by the uncertainty of the pre-computer era, where we never really knew if our moves were best. Get ready for some of the wildest chess in history from humans, who now know how to beat Stockfish.
Humans will never get anywhere near beating Stockfish
That is a GREAT quote Antonio from 2.5 years ago! The robots are coming and ever closer now! Then again they may already be here...
Alpha is not telling us that e4 is bad. Alpha is telling us that based on positioning and how you started the game, if you take what looks like "free" material, even just a pawn you will lose the game
grahamyodude In this game A0 took an offered pawn and went on to win, which counters your assertion.
That makes "zero" sense man.
No it doesnt, because alpha zero ended with 2 pawns, the one alpha zero took, and the one he offered gave him 1 pawn. The point Im making is that if you play like stockfish, then you will not understand how Alpha Zero plays. Im making the point that even if it seems like there are lots of variables in chess based on positioning, there are actually very limited variables. Notice how halfway through alpha repeats the same moves over and over until stockfish does something different. Its because alpha understands that he will not give up the positioning for anything. You need to think like Alpha zero to understand, you are thinking like stockfish. The fact that alpha both took AND offered a pawn, and is playing against stockfish, an engine that doesnt think like Alpha zero, actually backs up my assertion. Get on alpha zeros level
grahamyodude so you understand alpha zero? Impressive
I didn't know Alpha Zero is a man.
At 9:40, why didn't you bring up knight takes e-pawn with check? The bishop cant take since its pinned, or am I missing something?
I never play e4... My favorite is d4 like Alpha Zero. :)
You are way too humble, but I saw that interview where Alpha Zero said: "My favorite is d4 like Rishendra deo shakya."
Wow, Alpha Zero follows Rishendra' s games. 👌👍
AGEMO hahahaha... ;)
Human 3015, I have been playing chess for last three years and never played e4 in any game. And I am damn serious.
You are so clever ...
It's amazing how good Alpha Zero managed to get itself at chess. Whoever thought that playing thousands of games against your own powerful processing power and placing the results into a ram, creating a library of the good and bad could do that?
I think it's fascinating how it completely annihilates the likes of stockfish and has complete grasp of complicated endgame situations requiring scores of moves.
It plays them like a person MIGHT play 3 or 4 move scenarios.
Not your usual boring computer that's for sure. It feels so absolute that you are often on the edge of your seat and stopping the game to look at what's happening
Great video as always!
Just one doubt: why at 9:40 not playing Nxe6 with check and chance to win immediately a pawn?
I was thinking the same, but Deep Mind probably calculated some long line where it needs more moves to win or simply leading to a draw. Who knows..
No I just realised: after Kxe6 then king e7 and you lose the exchange on c8, because the knight is pinned and can't take the pawn
the knights moves are noted with an ´N´since the ´K´stands for King
Sorry you're right, I blundered
If Nxe6 Ke7 pins the knight
Love the quote and your analysis of Alpha Zero so far
His dog felt the true power of Aplha Zero
Best chess channel on RUclips by far!! I have abandoned most of the others because they get into the game and talk about 100 possible options instead of getting on with the game. Agadmator mentions maybe a couple of lines when appropriate and otherwise presents the game as it was. MUCH better than watching a half hour of "coulda, woulda, shoulda" rhetoric. Hope my broke ass gets some money so I can contribute at some point!
Alpha Zero be like: "In time, you will know what it's like to lose."
Alpha will probs lose to the next ai chess project
@@numbdigger9552 Then Google will best that engine with AlphaOne.
Why in minute 9:40 the white knight didnt grabbed the pawn on e6 if the black bishop is pin and after grabbing the pawn it comes with a check??? Help pls
Nxe6+ Ke7 then white loses the knight because black has unpinned his bishop and white's knight is pinned and attacked twice.
Shout out to the dog in the background. YOU'RE A GOOD BOY!
+Allofthegoodnameswheretaken Yesterday I prepared lunch for my girlfriend, went to pick her up at work, and when we got back, we saw Medo had devoured lunch all by himself... somehow he grabbed it from the table... But yes, usually, he's a good boy :)
lol what engine did you use to analyze this game tho??
One day, alpha zero will come up with it's own quote that goes on top of the video
Must kill humans - alpha zero
How long does it take for ai vs ai match to finish? Is it intantaneous or takes awhile?
Today I have my last final before I graduate college. Please wish me luck.
Goodluck
go alphazero!
u a physics major? whatever i'll wish you luck
Good luck!
Finance with a minor in statistics
Love the dog seems so happy x
anyone else noticed his dog having the time of his life on the couch in the background starting at 1:40??
come back and see the dog again. it's been 3 years
There’s something pleasing about just listening to agadmator just reading off moves in his Russian accent
@9:40 why does alpha not move Nxe6 with check? The bishop is pinned and I can’t see why it doesn’t go for it
Because the knight will be pinned to the King and it will be taken soon
Is there a limit to the computers potential? Will every game eventually end in a draw?
I cut my teach programing chess engines while in highschool in the 90's by 2000 when I had finish my education I had started work on a theoretical AI project using hidden Markov models and Bayesian tables. I made a life change to focus more on family life, but reviewing these games makes me believe that the learning mechanic's behind Alpha zero is based on a combination of those models.
Deep Mind I support you. -David_Dennis
I haven't seen anything concrete about Alpha Zero the chess engine, but there is a lot of material about AlphaGo Zero the Go engine. I imagine they are very similar, if not in fact the exact same architecture.
they are based on neural networks and General Reinforcement leaning algorithm
it's neural networks. These 80-x machine learning stuff isn't suitable for AI. Throw away and forget about it. Every time u'll end up with hard-coded evaluating function with only pieces cost undefined. It's not alphazero and any modern AI approach.
gopher neural nets are ai. Don't talk out off your ass.
There's a difference in calculating all possibilities and predicting outcomes. If I was going to speculate I would guess the engine to work similar to this.
Human players will often visualize a winning position based on how the current position is. More than likely what Alpha zeros is doing is looking ahead for winning positions then comparing them to the current position and calculating the probability of reaching that position.
An AI that uses probability is an AI that has the ability to produce a plan.
Something as complex a deepmind will have several layers that are checked against each but the real Holy Grail is in the layer that manages probability. The learning aspect of it or the "Deep neural network" or "modelled after the human brain" is just another way a saying the computer learns like we do. So how do we learn?
Here's a thought experiment, imagine the lights off in your house. While you were sleeping someone moved the door knob to your bedroom up 3 inches. You get up in the middle of the night to pee. When you reach for the door knob you notice it isn't there. Based on previous experience your brain predicted and moved your hand the most probably location of the knob. This seems a bit silly but it is based on probability. Our brains work and learn largely by prediction.
Alpha zero is likely working backwards from similar positions it would like to achieve that it feels is probable. This may seem like a very subtle change but think about how much it influences chess play among humans. -David_Dennis
It has already changed the legendary game of GO, which it was created to tackle, called originally Alpha GO Zero. But it is a general purpose algorithm that can be applied to any board game with complete information.
I think it would be even bigger improvement to partial information games. In those using just fied math and other algorithms to find good move candidates can be even harder. But neural network would potentially provide much better heuristics than human developers have figured out so far.
Reallyncool videos man, keep this up
dog almost killed me haha, had to pause. great vids.
I've played against the Berlin many times but I never go into mainlines, always anti-Berlin systems. I think the Berlin is an opening not really playable by anyone that isn't at least master level strength as the opening seems to achieve one of two results where either white will win the game or it will end in a draw. This is a tough thing for weaker players to grasp psychologically (me included) as never having any winning chances means that there is going to be relentless pressure placed upon you during the entirety of the game. Often my opponent will simply blunder or allow a strategically winning idea to unfold because white can just play comfortably and naturally while black is always sitting back and defending.
circle VIII please shre the anti berlin line that u mentioned. It will be very helpful and I guess also to a lot of chess players.
Well its not true that black cant play for a win in the Berlin. It is often used as black by stronger players to beat weaker opponents.
Shreyas Bhatt e4 e5 Nf3 Nc6 Bb5 Nf6 d3 is the usual
but there is ... Nf6 O-O Nxe4 Re1
... Nf6 O-O Nxe4 Qe2
...Nf6 Nc3 transposes into a Four Knights Spanish
... Nf6 d4 exd4 e5 can lead to some exciting positions as well
and this is just to name a few. It is a very rich opening ripe for exploration and experimentation 🙂
CLUCKY CLUCK I am not saying it is impossible to play for a win. It is just not an opening particularly rich with black counterplay if white knows what they are doing.
There are also some chances for Black in the Berlin to win. Even Anand demonstrated in his 11th game in the World Championship versus Carlsen that his approach (...b5) could have lead to very interesting play with 3 possible results. Of course you're right with master level strength! :D
You do realise that your dog is acting up funny in the background! I love your chess videos but I like that dog even better! :D
@9:42 why’d alpha ply rook to a8 instead of knight captures on e6 with check? The bishop is pinned and it looks like two free pawns to me.
TJ Galloway Rook to G1 check wins the knight
At 9:40 why does Ne6 picking up 2 pawns not work?
1:28 I am solely paying attention to the dog in the background.
@1:50 That dancing dog though! 😂😂
I will download thise vedio thank you brother you Surprise me I played e4 more than d4😯
Samir Aljubory Play 1. a4 bruh
Le Shryder I played a4 when I was 5 years old
Le Shryder I play e4 and d4 but e4 my fav
I never play d4
And I always open the game with e4 lol Queens gambit
When you make a joke, it sounds like you are not making a joke and my brain wont function properly anymore and i laugh. Every time. Love the vids man!
When my dad taught me how to play chess, he told me E4 was the strongest starting move for white....until you're a master! he said, when you get better, start with D4..and yeah pushing with C4 for black later in the french defense never worked out for me either so i stopped doing that
your dad is not the smartest guy right?
@@Rastoropny at least he raised me to be more polite than you, sorry you had a shitty upbringing and don't know how to even talk to people
What did you place in the description for those of us wanting to enjoy the two pass pawns in the center? Yeah, I'd like to see the technical
"Stockfish resigned the game" 😂😂
I like your dog spazzing out on the couch in the backeground at the beginning
It's funny, as soon as I read that quote, I got the goosies too. Powerful stuff indeed.
9:40, can someone explain why alpha didn't capture e6 pawn and fork the king and the g2 pawn?
For anyone interested, the "AlphaZero was trained for 4 hours" bit isn't quite correct. AlphaZero reached Stockfish's level after 4 hours, but these matches didn't happen until AlphaZero trained for 3 days.
I thought the same thing about 3 days, but it turns out that it was the previous Go program that was trained for three days. The original paper is clumsily worded in this regard, but table S3 from the paper states 9 hours of training time. However, looking at the graph of improvement (Figure 1), from the 4-hour mark to the 9-hour mark its progress nearly flattened out and it gained very little. We really we need more complete information. I hope it's forthcoming. Seeing all the games would be also nice rather than just the cherry-picked ones.
Stupid question: alpha zero trained itself to play, did that learning process stop before the matchup with stockfish? Was it allowed to keep learning during it's 1200 game tournament? Or was it's learning ability frozen in time to avoid creating a database of games and becoming like stockfish? If it was allowed to keep learning, is there a way to see if zeros rating improves if we compare say the first 100 games to the last 100 games?
AZ created no games database.
At 9:40, why not Nxe6+ instead of Ra8?
Yeah I also wanna know, I found it weird
Le Shryder Ke7 pins the knight, I just figure it out :(
Ohhh indeed woah 😂😂😂
the follow up is Nxe6+ Ke7 - Rxc8 Rxc8 - Nxg7 with two passed pawns... a draw
@@infoD1 If 58.Ne6+ then white will lose.
Lmfaoooo I love your dog in the background at 1:48
1:40 Your dog though
Brickhouse9398 😂😂😂😂
Thank God for human players, all this electronic/digital/technical whatever is anathema to the spirit of 'human' chess. And I will never stop playing e4 either.
I have a question: Can pure brute Force still beat AlphaZero? Lets say a nation state gave Stockfish 100,000x the CPU power and challenged AlphaZero in 2020. Who wins??
S L CPU power doesnt mean anything unless SF code is scaled properly to use all those extra cores. At some point it simply.doesnt make a difference how many CPU's SF has.
You would have to ask SF author about CPU limits. And how many cores he can realistically code for.
kasparov9 He said CPU power not CPU cores
Dave van Laren - When most say "CPU power", they mean the time it takes to complete a task, not the actual voltage applied across the chip. Well, kasparov9 is right, that time can vary greatly as the amount of cores, threads, and cache changes.
Most applications aren't programmed to make efficient use of more than 4 cores. If the application is designed for 2 cores, having a quad or octa-core chip would make little difference.
yeah agreed, that quote should give humanity the shivers.. just waiting for alpha-zero to play 1. a4 ... 2. Ra3 and put the lid everything
Fischer "E4, best by test"
AlphaZero: Really?
I have a question. I may be missing something obvious, but at 9.40, wouldn't 'Nxe6+' grab a pawn for A0 for free? Judging from the later moves, I don't see how A0 would miss something if it grabbed than pawn and then returned the Knight to f4 and continued as it did for the rest of the game or similarly. Unless it already knew it would win the position by that move, in which case ok. :P
I would ask DeepMind one question: how to play chess for win :)
u must have the brute force in ur inference engine.... :P
Doesnt knight x e6 checking black queen is a better move at 9:40 ? U can get both pawns without a lost piece
I'm not sure that AlphaZero is the greatest entity. It was a test experiment to prove that AlphaZero can play like a superhuman, not that AlphaZero is better than Stockfish. The conditions of the match were not fair for Stockfish to consider a real winner.
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.
AlphaZero runs in TPU, that's hardware specifical made for neural network. Stockfish needs proper testing and adjustment to run with 64 cores as apparently it run. However I'm not saying that Stockfish is better that AlphaZero, just that the match isn't a conclusion.
Stockfish usually has access to some sort of database of games and openings, and did not for the purposes of the match, while Alpha did!! It could be totally unfair unless I’m somehow mistaken
Stockfish usually has access to some sort of database of games and openings, and did not for the purposes of the match, while Alpha did!! It could be totally unfair unless I’m somehow mistaken
lol @stephen0793 aplha zero is a deep neural network based on General Reinforcement leaning algorithm its doesn't needs database to work with it learns from itself
I love the pup in the background. He’s just goofing around while we think critically about chess. My kinda bro...