Google Deep Mind AI Alpha Zero Devours Stockfish
HTML-код
- Опубликовано: 10 дек 2024
- Download Mproov and Improve Your Chess Today! app.mproov.me/...
Follow MprooV on Twitter / mproovapp #agadmator Read more about Deep Mind here arxiv.org/pdf/...
Link to the other games lichess.org/st...
A chess game between Deep Mind and Stockfish
Google Deep Mind Alpha Zero vs Stockfish
One of the games
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Be7 6. O-O O-O 7. d5 exd5 8. Nh4 c6 9. cxd5 Nxd5 10. Nf5 Nc7 11. e4 Bf6 12. Nd6 Ba6 13. Re1 Ne8 14. e5 Nxd6 15. exf6 Qxf6 16. Nc3 Nb7 17. Ne4 Qg6 18. h4 h6 19. h5 Qh7 20. Qg4 Kh8 21. Bg5 f5 22. Qf4 Nc5 23. Be7 Nd3 24. Qd6 Nxe1 25. Rxe1 fxe4 26. Bxe4 Rf5 27. Bh4 Bc4 28. g4 Rd5 29. Bxd5 Bxd5 30. Re8+ Bg8 31. Bg3 c5 32. Qd5 d6 33. Qxa8 Nd7 34. Qe4 Nf6 35. Qxh7+ Kxh7 36. Re7 Nxg4 37. Rxa7 Nf6 38. Bxd6 Be6 39. Be5 Nd7 40. Bc3 g6 41. Bd2 gxh5 42. a3 Kg6 43. Bf4 Kf5 44. Bc7 h4 45. Ra8 h5 46. Rh8 Kg6 47. Rd8 Kf7 48. f3 Bf5 49. Bh2 h3 50. Rh8 Kg6 51. Re8 Kf7 52. Re1 Be6 53. Bc7 b5 54. Kh2 Kf6 55. Re3 Ke7 56. Re4 Kf7 57. Bd6 Kf6 58. Kg3 Kf7 59. Kf2 Bf5 60. Re1 Kg6 61. Kg1 c4 62. Kh2 h4 63. Be7 Nb6 64. Bxh4 Na4 65. Re2 Nc5 66. Re5 Nb3 67. Rd5 Be6 68. Rd6 Kf5 69. Be1 Ke5 70. Rb6 Bd7 71. Kg3 Nc1 72. Rh6 Kd5 73. Bc3 Bf5 74. Rh5 Ke6 75. Kf2 Nd3+ 76. Kg1 Nf4 77. Rh6+ Ke7 78. Kh2 Nd5 79. Kg3 Be6 80. Rh5 Ke8 81. Re5 Kf7 82. Bd2 Ne7 83. Bb4 Nd5 84. Bc3 Ke7 85. Bd2 Kf6 86. f4 Ne7 87. Rxb5 Nf5+ 88. Kh2 Ke7 89. Ra5 Nh4 90. Bb4+ Kf7 91. Rh5 Nf3+ 92. Kg3 Kg6 93. Rh8 Nd4 94. Bc3 Nf5+ 95. Kxh3 Bd7 96. Kh2 Kf7 97. Rb8 Ke6 98. Kg1 Bc6 99. Rb6 Kd5 100. Kf2 Bd7 101. Ke1 Ke4 102. Bd2 Kd5 103. Rf6 Nd6 104. Rh6 Nf5 105. Rh8 Ke4 106. Rh7 Bc8 107. Rc7 Ba6 108. Rc6 Bb5 109. Rc5 Bd7 110. Rxc4+ Kd5 111. Rc7 Kd6 112. Rc3 Ke6 113. Rc5 Nd4 114. Be3 Nf5 115. Bf2 Nd6 116. Rc3 Ne4 117. Rd3
----------------------------------------------------------------------------------------------------------------------------------------
If you realllly enjoy my content, you're welcome to support me and my channel with a small donation via PayPal or Crypto.
Link to PayPal donation www.paypal.me/...
Crypto handle @agadmator
BTC address 3J7WigeEa95mNtZ8yJ26BBYexNz4r7XAUH
BCH address qzx5tu8uelq7s4tavsnk628f2t2s3g8gdvvnvcrvuq
LTC address MDrZdNjZm7btVkuLRdtrge9rwZn5TtjppM
ETH address 0x5840140C59C8b25AB59f0a3F6248c1f9cF1F2A0C
Nano
Check out some of the books I enjoy www.amazon.com...
Check out ALL my videos here • You Snooze, You Lose |...
Facebook: / agadmatoryoutube
Twitter: / agadmator
Instagram: / agadmator
Lichess: lichess.org/@/...
Chess.com: agadmator
League of Legends: agadmator (EUNE, my friend is using my EUWE account for a couple of years now)
Blizzard: agadmator #2992
Join our Discord Server here / discord
p.s. if you work in Twitter or Instagram, help me get Verified :) "Watch me without ads on your Amazon devices (bit.ly/Agadmat...) and Roku TV (bit.ly/Agadmat...)
Right after Stockfish wins the computer championship, AlphaZero shows up "Oh hai guys, whats this game called? Chess?", 4 hours later he is like "Here stockfish, munch on these 2 pawns and get in your corner like a good boy"
I can't click like because it is at 64.
Like that scene from the Matrix... "I know Kung- Fu!"
Just FYI, a bunch of people, including GM Nakamura, and myself believe Stockfish had an unfair handicap. Not to mention, SF was running on a older, weaker engine without a book. Just be aware of that.
Alpha Zero also had no book and it only "trained" for 4 hours.
They will most likely train Alpha Zero much more (perhaps many months) and then do official chess competition against all the strongest engines.
the 4 hour is also a lie, it was training for at least twice as much longer + it didn't only play without an opening book, it had no access to tablebases either
This is incredible... pieces hanging everywhere that can't be captured because of complicated tactics. This is definitely a game that no human would play.
Arya Well tal played couple of games like that
The difference is that most of Tal amazing moves has hidden mistakes. Stockfish would destroy Tal tactics because of its deep understanding of complex positions.
The scary thing is that they would, only 100 times worse.
Tal was the best !!
Julián Soto I was not questioning or saying that tal is on same level as alpha zero. I ONLY said that it's "like" that in the way that it has pieces hanging everywhere that can't be captured
11:44 "Stockfish resigned the game." Google should use that phrase to market AlphaZero.
Not many people would understand that though - unfortunately -.
*cough* leela chess zero *cough*
AlphaZero is not for market, yet.
I don't think Google have any say in the matter. Alpha Zero will be putting Google up for sale soon.
I've never seen an engine resign before
*Me: Moves E4
DeepMind: mate in 33*
😂😂😂😂😂
I like to imagine this as DeepMind being another player who looks like Ben Shapiro, whispering this under their breath after you play.
Don't overestimate yourself.
Before the game: +/=
You: 1. e4??
Alpha: Mate in 4
Lol
u'd get mate in like 7 bruh
I imagine Deep Mind with Hal 9000's voice. "I'm up the exchange, Dave. You will lose the endgame."
+Brandon Nobles :D So far, I imagined a woman's voice, but Hal works :D
Haha!
>Stockfish wins a piece
>Deep mind:"This game is too important for me to allow you to jeopardize it"
>Proceeds to crush him anyway
"Mate in 73. You lose".
"I think you missed it, Dave"
"Get me out of this corner!"
"I'm sorry Stockfish, I'm afraid I can't let you do that"
Take a look at your history.
Everything you built leads up to me.
@@NaoshikuuAnimations lMAO
Naoshikuu So fucking underrated
erbh
"I'll beat your ass in chess and jeopardy..."
Magnus Carlsen: "I often sacrifice a pawn to get a better position"
DeepMind: "singular?"
ahhhh smite 1v1 legend king
maybe i am an idiot but the most fascinating part was the opening. in 4 hours it practically had come up with a classical main line of the queen's indian. in a weird way validating human effort of decades haha, but that is scary in it self, thinking that ok now we know we did well, because this algorithm knew it in 4 hours.
Νυανγκάο Athanasios Makris But just think, the only reason the computer was able to come up with these solutions was because humans created it. It might be a lot better at chess than any human being, but it shows that humans have that capability to improve.
I can't argue with that. We can certainly improve. And I won't act like I know how these algorithms work, but Aga said that it was given nothing but the rules. So I took that as it played d4 because it quickly figured out that controlling the center is important for example and then built up from that.
But maybe the true culprit is the opponent, Stockfish, that knows the openings. A0 was finding the best moves to continue from what Stockfish was fishing for in its database. So we still have done well in chess as a species haha
It knows all the openings now, btw. And wins all of them too, the PDF is pretty short, take a look.
womplestilskin Thanks, I will
I think this 4-hours thing is a bit delusory ... it probably played more games of chess against itself then ever played by humans in the history of mankind or some crazy stuff.
This whole game is as if Deep Mind just decided where it wanted its pieces and Stockfish was powerless to stop it.
Arya it's like being able to solve a rubik's cube without an algorithm
Just like my games with me as Stockfish in this scenario. hahaha
Great way to put it, yeah truly out of this world chess
A0 combines the best human strategies + cold hard calculation of a computer + it's own unique insights. It's PERFECTION.
actually if I understood it correctly it's just cold hard calculation of a computer + it's own unique insights
Maurice Ashley: stockfish any comments on the game? I mean you were making some not smooth moves against Deepmind. How do you feel?
Stockfish: I mean what do you want me to do?!?
+Cameron Smith Nice :)
I laughed pretty hard at this.
Loving the Maurice Ashley's comments. Hahaha
haha epic comment
Cameron Smith ahahaha
When I woke up this morning, the last thing I expected to see was Stockfish being completely roflstomped into oblivion.
You still haven't
You could have told me that N. Korea launched a nuke onto China and it would have shocked me less than this news. Chess is just the beginning. Soon AI like this will use human-like intuition to replace our doctors and lawyers and even inventors. I can't wait.
Ottis...wake up, it's over.
roflcopt0r
@@Garroxta What if it replaces humans itself? Scary
It is interesting that Google's deep mind does not just calculate moves but also has a plan of steering the games into positions to immobilise the pieces no matter what it takes and then bust open positions by sacrificing some material to gain advantage(kinda like Ivanchuk) but maybe this is the greatest strategy possible in chess and what stockfish has done so far is to just calculate by brute force without any strategy
Exactly, A0 combines the best human strategies + perfect calculation of the best engines + it's own unique insights.
Deep mind also uses brute force.....
@@paultan5419 yes but it does immobilize the enemies pieces which is very advanced
@@firstnamelastname7298 The AI has no idea what chess is or strategy it's using. It's merely a naive algorithm using a trained machine learning technoque to evaluate all the possible states (with pruning) and pick the best move. It has no idea what chess is, or any concept of what it's doing at all.
@@firstnamelastname7298 my point is that whatever techniques it uses, no matter how advanced, it's still brute force (from a computer science perspective )
2:11
agadmator: deep mind made stocfish like a little...
me: bitch
agadmator: kid
me: kid
+Maulana Harjuno :D
😂
:🤣
😂
Be honest-you were going to say 'bitch' but censored yourself, didn't you? :D
Agad, all I gotta say is, when you reach 100k, 200k, 300k, and so on, don't ever change the way you operate this channel! You're simply the best RUclipsr I've ever watched, since RUclips started in 2006 or whatever. It's okay to tweak things here and there to try new approaches, but the very basic premise of your channel is hands down the best on the internet. Thank you for bringing so much joy and laughs to my life man! I been subbed to you since 5k, from an old channel I used.
+BayToTheVeg Thank you for such wonderful words. I'm not planning on changing it, maybe I'll fit in a vlog here and there :)
I agree!
Which is fair enough since we all don't (want to) pay for this content!!
More like one of the worst chess channels. I guess beginners like this kind of rambling without any explanations
Buranazi that's not very nice! If it annoys you so much, don't watch or comment silly :)
Stockfish = Usain Bolt
AlphaZero = Jet Engine
No no no! It is Jiren vs Goku!
Stockfish = hamster
AlphaZero = Lamborghini
It's more like:
Whoever the best human chess player is: Usain Bolt
Stockfish: Ferrari
Deepmind a0: F-22
What this game shows me, is that human opening theory is too conservative. Deep Mind used its knight aggressively and early, to tie Stockfish in knots and achieve a decisive advantage. Of course, you have got to be be damn strong to get away with this!
Learning It Quietly well also humans wouldn’t play the way stockfish would IN REACTION to Deep Mind’s moves.
The problem is that some of the lines these engines may play are so risky that if you aren't 100% of accurate you will blunder the whole game and even the greatest GMs will struggle to find those moves every time.
Can we have more Deep Mind games please?
agad you said stockfish was rated around 2400 in the beginning of the video, isnt it 3400?
Yes, it is :) I meant 3400 :)
Looking at the way game went ... we so understand why you said 2400. Scary! 😳
Hello agadmetor,nice video. By the way can chess openings be mastered without knowing theory? If no, what about chess 960 ? All that is required is creativity.
agadmator's Chess Channel ya right..because when ever i make stockfish and magnus carlsen app play....stockfish crush him...so it can't be 2400
The gams lasted for 7 seconds
2:12 "deepmind makes stock fish look like a little...b..kid"
😂😂😂
Damn I thought I was the only one
I laughed so hard at this.
Too weak, too slow.
Like shooting stockfish in a barrel.
this is gold
@@duckgoduckgo this is finegold
Thanks for making this game available, it is the first I
heard of it.
It appears that 100 games were played between Stockfish 8
(S8) and the Google entry, AlphaZero or DeepMind (DM). It seems we are
comparing apples/oranges in terms of the hardware on which the games were
played. DM was played using 4 TPUs (Tensor Processing Units which are little
super-computers that do specialized work). S8 was said to run using 64 threads
(does this imply CPUs?), and 1G of hash. S8 isn’t the most recent version, and it
also plays endgames better when it has Syzygy endgame tablebases available. The
paper doesn’t mention this. It does say that S8 was searching at 70 million
positions per second, which is about right if 64 CPU’s are available.
DM searches 80 thousand (that’s right) positions a second
and according to the paper, is highly selective in its search. This has been
the holy grail of chess engines since they were first conceived. Humans don’t
look at every possible move, but how do you design software to think like
humans. S8 and other modern beyond-human (they use superhuman way too much in
the paper) engines do it by pruning the hell out of the search tree. All
captures and some checks are examined, but after that, only a few “quiet” moves
qualify for examination. This is how S8 can search so far ahead in a short
amount of time.
We also don't know either side had an opening book, though I’m
guessing no.
I agree with your Tal comparison. DM does appear to always be looking for the
series of attacking moves. Very aggressive. But here we must consider a couple
of things. Tactical move sequences, where both sides are trading pieces can only
be resolved by tree searches using current state of the art engines. Has the AI
of DM found a better way? After all, it is only searching 80K positions/second.
Is this enough speed to resolve deep combinations? They paper doesn’t describe
what goes on in their searching, but we know from experience that most chess
games are won/loss by the balance of material.
This is especially true at the early stages of the game, where if you
are down a minor piece, it is a tough battle to catch up. Searching a tree and
only considering the captures can be fast, since the problem self-reduces by
the nature of removing a piece from the board with each capture. The paper
doesn’t indicate how “deep” their search is.
I’m sure more will come out of this as more papers are
published.
The claim was made that DM was only given the rules of the
game, which I assume means it was not given information about the merit of
various elements, such as is having a Bishop pair vs a Knight pair is good or
bad, but it might have been told to consider that feature in its learning.
This is how chess engine evaluations work. they measure many
"features" of a position and then come up with a number that
represents how good/bad the position is. Regression analysis is usually used to
combine many features. Again, not sure how Google did it with DM, but I’m curious.
Their algorithms must measure something, so what are their features?
AI techniques have been around for some time, and there is nothing new about
self-learning chess engines.
I would take some claims, such as it only was taught the rules of chess, with a
grain of salt until we know more about what the neural network is measuring.
Show me the details!
Back to the game and the pivotal moment.
21. Bg5
WOW! I agree, where did this come from? I let SF chew on this for 90min looking at the top 5 moves and it got up to 41 plies (20 moves by each side) and found Bg5 as the best move. SF is only searching 1.3 million positions/second on my slow laptop, but according to the paper, each side was only given 1 min per move. So, based on some simple math, SF should have found Bg5 in 90 min on my laptop if the move was that
decisive. I then had SF search the top 5 moves...
So what this really means is, DM found a 20 move ahead move and SF8 agrees.
well beyond what we mere humans can figure out!
You ask what DM’s rating is? Based on the paper, it is just slightly better than SF8 at 3300-3400.
They played 100 games, with each side playing white and
black equally. DM won 25 games as white and rest were draws. DM won 3 games as black, the rest were draws.
DM did not lose a game.
Amazing….
Alan Kilgore I don't think DM was given any openings, and I don't think they are that close if DM never lost once in a hundred games but won many games, with the rest ties. I could be wrong about the opening books though
The way AlphaGo was developed, I completely trust that DeepMind was given nothing but the rules. It is amazing what trial and error can achieve without anything to prejudice your perspective on how a game ought to be played.
lsduuhyik ygytyuui
1i thought they just played 100 games.
There are videos on how deepmind works, or how the newest one which uses reinforcement works. Reinforcement learning basically means it plays against itself over and over again, recognizing new patterns over and over again that lead to a win more often. It starts just playing randomly at first, and recognizes patterns that lead to winning( like castleing). The thing that is amazing about this is that there is no human involvement whatsoever. It has learned from complete scratch, and has deduced the best possible lines and theories and ways to play the game in 4 HOURS. This is also kind of sad as well. If you run deepmind longer, it will eventually get to a point where it is playing chess the best way it can possibly be played. If they do that, chess will literally be complete, in my opinion. There will be no further discussions about what the best move is in a position. Sure, you don’t have to play that move, but the fact that you know it is the best is kind of sad to me.
Grayson Parker this is an assumption.
Perhaps let it play against itself even longer and it will realize that last move previously thought best was actually not.
Even though it was only 4 hours, do you know how many thousands or millions of games it went through?
Very moving. NEVER thought I'd see something hand stockfish such a profound loss. Damn! What can I say.
Amazing. Just a note of the variation shown where Stockfish didn't capture on g5 - actually Stockfish considers it a very bad position for black to capture, and the reason is clear -
The variation you gave contains a small error - instead of the incorrect hxg7 in the middle of the variation (which indeed leads to a balanced position according to Stockfish), just Qd4 leads to a winning position for white.
Amazing game :) Thanks for posting
I watch these games and sometimes zone out for a second or two, then I get taken back by stockfish somehow being in even deeper trouble than just some 4 moves earlier. I continually have to pause the video and go back to look how it happened.
There are some very seamless transitions from position to position and stockfish is fumbling in the dark. This is incredible.
Deep Mind is the name of the Google company that developed the software, AlphaZero is the name of the program.
There was probably a mistake regarding the ratings; I think stockfish is 3300+ and not 2300+ :)
+mike s. Yes, I meant 3300 :)
Also, they played 100 games of which Deep Mind won 28 games and 72 draws were made. The supercomputer they used to train it takes 4 hours, but that would take around 70+ years on a regular PC, so the question is, when will it be as light as other engines out there, and when can it be used for training humans?
Look at how Deep Mind and Stockfish only push the minimum number of pawns necessary... This is very interesting and educational
Arya once you push a pawn, you cant un-push it.
Alan M. In human play you can see pawns marching forward in order to anticipate a good piece placement or to protect a well-placed piece. But Deep Mind just tossed his pieces in the enemy's camp with little to no support.
Not only in the enemy camp, he puts his pieces straight under attack. I mean, I tend not to put my bishops in the diagonal of a pawn.
I love pushing pawns. I think I gain tempo and space. Usually they get too far ahead and cannot be protected. Divides the army. Checks with no purpose.
In the middle of watching this game and I just want to say, in reference to your survey earlier, this is the length of video I am most content with on this channel - you give enough information on each move and it's direct implications, and you should keep up the great work. Now back to watching, and I'm thoroughly enjoying this game so far
Wow. Deep Mind *obliterated* Stockfish. I have never seen Stockfish get so utterly *demolished*. Unreal.
Nathan Xaxson o
AMAZING! I was studying Reinforcement Learning, (which is the same framework from which Deep Mind's AI was trained) when I saw this video. At the end I just got back the enthusiasm I was loosing for the lack of sleep. Amazing time to be alive! Reinforcement Learning/ Machine Learning / Deep Learning is starting to make a great technological revolution. What a time to be alive!
When I play against strong Engines my Position always looks likes Stockfishs in this game. Scary
Great that you posted the arxiv article! Very professional of you, thanks :)
Dude...you are the top youtuber on this particular AI subject!!!
Hi agadmator, in the article you shared with us there are 10 games played between these two AI. The article is really good and I recommend people to read it
I know you probably don't want to do 2 games from this set, but example game 9 is a crazy game that I would love to see you review.
1. d4 e6 2. e4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 cxd4 7. Nb5 Bb4+ 8. Bd2 Bc5 9. b4 Be7 10.Nbxd4 Nc6 11. c3 a5 12. b5 Nxd4 13. cxd4 Nb6 14. a4 Nc4 15. Bd3 Nxd2 16. Kxd2 Bd7 17. Ke3 b618. g4 h5 19. Qg1 hxg4 20. Qxg4 Bf8 21. h4 Qe7 22. Rhc1 g6 23. Rc2 Kd8 24. Rac1 Qe8 25. Rc7 Rc826. Rxc8+ Bxc8 27. Rc6 Bb7 28. Rc2 Kd7 29. Ng5 Be7 30. Bxg6 Bxg5 31. Qxg5 fxg6 32. f5 Rg8 33.Qh6 Qf7 34. f6 Kd8 35. Kd2 Kd7 36. Rc1 Kd8 37. Qe3 Qf8 38. Qc3 Qb4 39. Qxb4 axb4 40. Rg1 b341. Kc3 Bc8 42. Kxb3 Bd7 43. Kb4 Be8 44. Ra1 Kc7 45. a5 Bd7 46. axb6+ Kxb6 47. Ra6+ Kb7 48.Kc5 Rd8 49. Ra2 Rc8+ 50. Kd6 Be8 51. Ke7 g5 52. hxg5 1-0
+CubesAndPi I will probably do 1 more, checking all of them now
3 wins, 47 draws with Stockfish on white. 25 wins, 25 draws with AlphaZero on white.
Démian Janssen that's crazy
Oy that bishop sac...
@@iNeedPhone That shows how strong the starting position actually is for godly players. 😂
I think one comment you made is extremely insightful, and it is something I thought as I was looking at the endgame: when a strong human player is up the exchange in the endgame, it doesn't take long to finish matters off; and this is due to a collection of inaccuracies by your opponent. What this endgame shows is how much technical prowess is required in an endgame when no glaring inaccuracies are being made. Study the endgame, folks.
This is so beautiful! Stockfish plays like a engine BUT Deep Mind plays like a Human who is 10 times stronger than the strongest human! Crazy and also shows the scarry future of AI's
I saw in someone else’s comment you said you might throw in vlogs at some point. I think that would be cool. Congrats on your channels success. Definitely my favorite chess channel!
Stockfish: "Your move, D."
Deep Mind:
it is still worth watching even after almost 2 years , thanks for sharing the most beautiful end game in the history of chess
6:33 Deepmind to Stockfish: "Look at me! I'm the boss now."
I can't even describe how amazed I am; I mean, self-teaching for 4 hours and it destroys the strongest engine? o_O Thanks for this mister!
I've recently seen an AI similar to this called OpenAI. It learns exactly like you've described Deep Mind - by continuously playing against itself given only a rule set. Now, OpenAI has been used in Dota2, a computer game, so it's quite different in that regard, but still very similar. And yeah, it usually beats the professional players very easily. In the case of Dota2 and many other video games, the games themselves are too complex to make a successful equivalent of Stockfish, but this new method seems very potent.
If people tell the program what to do, it will only make "people-moves" so to say. If it learns by itself, however, it will make up new things, and people can in turn learn new ways of playing and exploiting things in the games from the AI.
Very exciting to me! :)
Agadmator you analysis has really really emproved the longer you have done this
and deep mind plays french defence....agamator sheds a tear
Lmao
Sends "sometimes to win a war, sacrifices must be made" to a whole new level. 2 pawns for 2 pieces that are not in play.
I love how you self-censored at 2:14 lol ;)
blamtasticful. I missed that 😂😂
This are the things that I love to hear about science and technology. Just incredible!
3:48 "I mean . . . this is Stockfish." lol
Symphony. This game is so beautiful that it has the same effect of a masterfully composed symphony which leaves it's audience speechless. I am deeply moved by this game.
I'd be interested in seeing some of the self-play games that occurred near the end of A0's training. How do you beat something like A0? Only A0 knows.
re the slow conversion: another thing to consider is that alphazero works by optimizing the winning probability (unlike stockfish which optimizes the evaluation in centipawns), so it happily chooses a slower conversion if this means that it's more sure to win. Also in the Go version, conversions were often strange to the human eye
117 moves, unbelievable
it is maybe the best ending game in the history of chess but this is indeed the best commentary about alpha zero in the whole Universe , thanks I think I watched this tenth times and it is still magnificent and has a lot of magic into it. and I also have an idea and I am going to do the exact moves against stockfish 8 and see what will happen :)
Me looking at the thumbnail: THATS CALLED THE WOODEN SHIELD
This just highlights the amazingness of a fact that human actually managed to pull a victory against Deep Mind in Go.
Go is just so much more intangible and complex than Chess is..yet so much simpler at its core.
A true Beautiful Game.
2:11 deepmind makes stockfish look like a little...
U wanted to say Bitch haha😂 I feel you it pronounces smoothly.
+Robb V. Of course not :D
I thought he was about to say a little boy but changed it to kid. I don't think it is that common to use slang such as "little bitch" if you're not from the states.
Victor Udd-Peterson i aint from the states
I want to thank you dear Antonio for such a delightful channel - you’re doing it so vividly and interesting .. p.s. I laughed so many times from your comments on some moves and tactics .. keep up the good work !
What do you need to win a chess match?
Alpha Zero: 3 pawns ad a knight.
Damn, you're quick!
Skipped watching your stuff for 3 days (busy). I get linked to the paper on arxiv. I check your channel: already the case :o
Just setup your channel on ROKU. AWSOME!!!!
Cool. Let me know how it works :)
I love how chess is such an interesting measure for these things
I must be the only one here who doesn't know anything about chess other than how the pieces move lol
To me, the guy reciting the letter and number coordinates immediately after a piece moves was unreal XD
Back about a decade ago a chess engine, RomiChess, was starting from an empty learn file and winning 100 game matches against the world's best including Rybka. And winning them with a higher score than AlphaZ beat stockfish. And RomiChess with learning disabled was only a 2200 to 2300 elo engine back then depending on who's rating agency was consulted. After 20, 20 game matches against Glaurung, the predecessor to Stockfish, RomiChess was scoring 95%. But no top engine author cared and most were dead set against say a 2250 elo engine trouncing them because it had advanced reinforcement learning. It would be the same way for AlphaZ today except AlphaZ is getting a billion times more publicity than what RomiChess received. And RomiChess did it on equal hardware! RomiChess was a far greater success story than AlphaZ by a mile. And yet the story never got told.
RomiChess is rated 2425 today and after just 102 games of training got this solid draw from 3400 rated stockfish.
[Event "Romitest"]
[Site "MASTER"]
[Date "2017.12.10"]
[Round "103"]
[White "RomiChess64P3n2"]
[Black "Stockfish_8_x64_popcnt"]
[Result "1/2-1/2"]
[BlackElo "3495"]
[ECO "C67"]
[Opening "Spanish"]
[Time "14:54:01"]
[Variation "Open Berlin, Pillsbury Variation"]
[WhiteElo "2425"]
[TimeControl "10+1"]
[Termination "normal"]
[PlyCount "139"]
[WhiteType "computer"]
[BlackType "computer"]
1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. O-O Nxe4 5. d4 Be7 6. Qe2 Nd6 7. Bxc6
bxc6 8. dxe5 Nb7 9. Be3 O-O 10. Re1 d5 11. c4 Na5 12. b3 Be6 13. Nd4 Bc8
14. Rd1 c5 15. Bd2 cxd4 16. Bxa5 dxc4 17. Qxc4 Be6 18. Qxd4 Qxd4 19. Rxd4
Bc5 20. Ra4 f6 21. Nd2 fxe5 22. Nf3 Bd5 23. Rc1 Bxf2+ 24. Kxf2 e4 25. Bb4
Rf7 26. Rc3 h6 27. h4 exf3 28. gxf3 Bb7 29. Kg3 a6 30. Ba5 Raf8 31. f4 c6
32. Re3 Rd7 33. Rae4 Rd6 34. Re6 Rd4 35. Bc7 Rd1 36. Kf2 Rd2+ 37. Re2 Rd1
38. Be5 Rf7 39. Rd6 Rxd6 40. Bxd6 Rd7 41. Be5 Kf7 42. Ke3 g6 43. Rg2 Rd5
44. h5 g5 45. Bb2 c5 46. fxg5 Rxg5 47. Rxg5 hxg5 48. Kd3 Bd5 49. Bc1 Kf6
50. Be3 g4 51. h6 Kg6 52. Bf4 Be6 53. Kc3 Bg8 54. Kd2 Be6 55. Ke3 c4 56.
bxc4 Bxc4 57. a3 a5 58. Kf2 a4 59. Kg3 Be2 60. Be3 Bd1 61. Kh4 Bf3 62. Bf4
Be2 63. Bd2 Bd1 64. Bc1 Be2 65. Bf4 Bd1 66. Bd2 Be2 67. Bg5 Bd1 68. Bc1 Be2
69. Bd2 Bd1 70. Bc1 1/2-1/2
The revolution started 11 years ago!
This is frightening.
good work.. best chess tube by far. regards, from istanbul
I am working on my masters in mechine learning and I can tell you we are very far from skynet level ai. We are talking like probably not even my life time kind of time. Then again technology is growing at exponential rate so we will see.
That end game reminded me of the scene in The Matrix when Neo is blocking Smiths punches with one hand.
Mr. Anderson
Someone has gotta say it: Skynet!
At 3:03 some engines recommend 10. Na6, developing the other knight to an inferior square to c6, but then one wonders about the future of the light square bishop.
It's amazing but here is a comment of a machine learning expert (Pedro Domingos): "AlphaGo Zero is great, but hold on: self-play is one of the oldest ideas in ML, and humans take far less than 5 million games to master Go."
Here is a comment from a machine learning guy who is probably not as much of an expert as this guy you're talking about: AlphaGo Zero is so much stronger than any human could ever hope to be. If that's what it takes to "master Go", then it would take a human infinitely more games to master Go, because a human would never reach that level.
Yeah, but it is also a few thousand times better than any human who has mastered Go.
you can't possibly master go.
I think it's very important to keep in mind how difficult it is for a software to play Go on a professional level. They only achieved that last year with alpha go (was it last year or two years ago?) before that there was no engine that could beat a top level go player not even the lower ranks of professionals. Agadmator brushes over this a bit but I don't blame him, most chess players don't know anything about Go.
Nice touch with the Roku!!! Will subscribe :D
Not a word about go, agmator. I am disappointed. But your surprised reactions are just like the ones in the go community, when they were confronted with AlphaGo: "This is beauty! Wow."
+Ger Hanssen I never tried GO. Don't think I would contribute much. Is there a good online server to tryout go? I am planning to learn the basics some time soon and maybe do a short video on it
Ah, that would be interesting. I am willing to help in one way or another. What I wanted to say is that this whole development of AlphaZero came out of the challenge to play go, and that would have been worth mentioning.
I don't play online go very much. Pros nowadays prefer Tygem (www.tygembaduk.com/ehowto/tygemgo.asp), but there are at least 10 other ones, some of them 100% Asian. I know that from the videos of Haylee (ruclips.net/video/VBv7XpiEY44/видео.html). She is a top pro that played as a strong amateur commenting during the game while thinking. She picked the idea up from a chess player. She never mentioned who.
Sensei's library (senseis.xmp.net/?About) tells you just about anything about go and the go community.
3:18 if C5, the bishop on B7 is protecting the knight. Other than that, great analysis as always! Very interesting game, it's one of those games where you only realize you're in deep trouble when it's too late.
'This could be the greatest end game ever played in the history of chess!' - you nailed it with that line.
Everyone. Mark this day. A paradigm shift in the application of AI in chess (and beyond)....
This is BRILLIANT!! So BEAUTIFUL!! Wow!!!!
“I often sacrifice a pawn”
Deepmind: “Singular or multiple?”
“What”
“What”
Hi Agad I really love your vids great analyzation more power :D
Wow..what a match...im very happy to see stockfish lossing a game as i play against it and he crush me like hell ..so it feels nice....😉
Hey agadmator when i saw that game of bobby fisher in which he player evans gambit(if i wrote it right) from that day im in love with that Openning...but as usual i can't play it that good...so can you show us a video fully analysing that openning or can you tell me a book to refer for mastering in it..??
I was hoping for Stockfish to win. So were most of the chess programmers who have spent some considerable time of their lives learning and developing conventional algorithms. It was a bit sad to see some AI from Google come out of nowhere and defeat the strongest chess engine. But we have to accept it, move on, and learn new methods. Although it's too early to conclude anything. More testing must be made: different hardware, different time controls, etc. Stockfish did not have an opening book! It's important. And also, I don't see many people mention that most of the games were DRAWN. Agadmator gives the impression that Stockfish lost all of the games. It's misleading.
pahom, so you are a programmer and you claim that Alpha0 is not AI? How is it not AI then? I'm genuinely curious. Because to me it totally seems like AI implementing unsupervised learning. I'm not familiar with details of Alpha0's implementation but we can all agree that the core of it is machine learning which is a field of AI. And even if the gathered data is then interpreted as evaluation it doesn't really matter, that's not the point.
pahom, I see, so you're a smart-ass. Good to know. Many competent developers worldwide use the term AI and they're completely fine with it. I'm not pretending to be an expert here: AI is not my field. But this doesn't suddenly make me not a programmer. I work with .NET platform, if you're curious. You can go and convince someone else in your superior intelligence. I'm not interested.
pahom so you are saying that they are misleading us............ and how you know all this???
pahom AI stands for Artificial Intelligence. This term means a program that can teach itself to do something really good from a basic level without human input. Think about it as you would about a human mind: you teach your child, say, the basic chess rules, and after that you give him a chess set and tell him he has some time to play against himself and improve. This is what they mean by "intelligence". If it's "artificial", them it's a human made program that can teach itself. Google's Alfa Zero can teach itself to play chess, just like a human with real intelligence can, only way faster (like 1000 times faster). It also has more computing power for complex calculations of course, just like humans have a brain for that. That's why Alpha Zero is indeed an AI. It is not the final step of AI development by any means of course, but denying it is an AI is strange, as you know that it managed to teach itself, so what is the problem? If you were expecting AIs to be able to teach themselves to play chess without any inputs like basic chess rules, then your expectations may need to be altered, as no real world entity can operate with magic.
Google Deepmind opened with queen's pawn. It was at this moment that Stockfish resigned
2:10 “Deep Mind makes Stockfish look like a little b- uhhh kid”
We all know what you wanted to say...
"This is madness!" "This... is ALPHA ZERO!"
3:47 stockfish? More like stuckfish
5:52 that image of stockfish with two knights, a bishop, rook and pawns all crammed into the corner completely useless, while the king in the opposite corner with the defending pieces getting overwhelmed...
sometimes I wonder why spend years learning about a game that an algorithm can master in just 4 hours?
Because, in the end, it's still the humans who have the creativity to do those things, it was human intelligence that created AI, even though AI is vaaaaaastly superior in speed processing and effeciency, we still have creativity and a lot to be expected from :D
It's worth noting that the hardware setup was favourable for alpha zero. Stockfish was given 1GB hash table size, when 32GB or more would be expected. Also, the time control was 1 minute for a move and no endgame tablebases were given, despite the current norm for computer chess engine chess games.
How is hardware favorable for AlphaZero? Why should it have endgame databases when AlphaZero doesn't need it.
Niels Liljedahl Christensen the time control was 1 minute for a move. Obviously the hardware could make great difference. We shouldn't say which engine is better basint on such games. That's all I wanted to say. Nevertheless It's an amazing work, but I think all the praise went into wrong direction. It's astonishing that after just 4 hours of training that engine could play so well and I think that was the main point of this work. The endgame part was a side note to pinpoint further differences from regular chess engine matches.
I agree that the most amazing thing is how fast it learned. However, I don't think hardware was favorable for AlphaZero. It just uses completely different hardware. Stockfish uses a 64-core, and AlphaZero uses 4 TPUs due to its evaluation function being a neural network, which is just a lot of matrix multiplications. They probably have similar power consumptions.
I'm mostly disappointed about the hash table size given. Usually is a lot bigger. Hardware has a great impact on Stockfish performance and if we wanted to actually compare two engines we should focus on that matter more. At the moment any comparison between these two (in terms of their overall strength) could be easily discarded.
I think they're just comparing evaluation functions and search, which is what AlphaZero brings to the table. AlphaZero is not a chess algorithm, it's a game learning/playing algorithm. If they actually wanted it to be a competitive chess engine, they would add a lot more, like books and endgame tables etc., but I doubt they really care about that. DeepMind is interested in AI, not in chess (or Go or Shogi or whatever).
Nah deep mind would lose to beth’s queen gambit
Creating symphonies at the peak of human understanding, GMO food that taste better then the best flavor profile of anything ever tasted before, artwork compiled in dimensions never thought of; all while talking to itself in language beyond our comprehension,.. Creating viruses, phages, vampires, and fairies then going to war with itself.
I played a move AlphaZero played, 1.d4
Love your channel so simple
These two players need to study more openings
For example, at 3:02, why didn't white take the dark square black bishop?
Floyd Maxwell that would just give up a powerful knight and allow black to develop their queen
Aren't bishops worth more than knights, and the dark square bishop the more valuable of the two?
Mate, they are beyond studying petty human openings
human opening cant even compare to these, it's on another level
Given that neural networks can be trained and theoretically produce different results, unlike traditional chess engines which are purely computational and bounded by the information fed to them, I wonder if we could identify some distinct "styles" that emerge from many iterations of Deep Mind. What if we trained 10,000 of these and found one which was more "tactical" and one which was "positional"? What if they played each other?
What would happen if we fed the rules of the game into Deep Mind, but started with a set position? GMs are already pretty adept at arriving a set position 15-20 moves into a game, so having the best chess "entity" in the world play out moves from one of those set positions as a next level of preparation does not seem too far off in the future.
The key aspect here is the 1 minute / move time control, as due to neural networks and search algorithms AlphaZero is far superior to Stockfish: 80k vs 70k evaluation speed (positions/second) in favor of AlphaZero! It would be interesting to see a 100 games match in standard time controls between these two... I would bet then the evaluation speed won't count as much and the result would be way more close than today!!! :)
According to the paper, AlphaZero evaluates 80 thousand positions per second, while Stockfish evaluates 70 million per second, so in fact AlphaZero is vastly superior in terms of efficiency in the number of positions evaluated. However, each of those evaluations comes at a much higher computational cost.
In some of the papers I have read, they have stated that the time control arguably places a greater handicap on alpha zero and that stockfish is stronger on quicker time controls
Neural network evaluating is really slow compared to the evaluation function of Stockfish. The 70k you mention is actually 70 million, but it's always nice to try and confuse people by pretending you know something when you really know nothing
Everyone talking about how impressive it is it taught itself to play at this level in only 4 hours, but remember that's 4 hours of running on Google's astonishingly large supercomputer cluster! It's still super impressive of course but since we're talking about chess computers, mentioning the processing power involved is important.
There were actually two different models used: one that used 4 hours, another that used 9 hours of training. The 9 hour version won a 100 game tournament against Stockfish 8 at 28-0-72.
It's worth saying there's good reason to think Stockfish would've been much stronger had it not been handicapped though - they used an old version and fixed move time to 1 minute, instead of allowing Stockfish to do its own time management which probably had quite a big difference on its strength. It's still important research and there's no question AlphaZero was stronger, but the research conditions were still a bit dubious.
Ehm, Deep Mind developers? Hi. Just wanna say... if you see a woman with a shotgun, a very big man and a child, just run.
Hasta la vista, baby.
HahahHa
More deep mind games please
I don't know if it is just me being a paranoid person, but This Alpha Zero has got me pretty worried, tbh
Alpha Zero took in to consideration run the world but at the end it seems prety point less for him, you know he calcula tes very far away and well at the end It's not interested
You know deepmind is good when even stockfish recommends resignation
When I play stockfish, 30 moves in
My pieces: my time has come
Kasparow: To me, pawns are trading material for activity, coordination, position and attack.
Alpha0: I agree wholeheartedly