Google Deepmind's AlphaZero Chess Engine Makes "Inhuman" Knight Sacrifice
Вставка
- Опубліковано 5 гру 2018
- AlphaZero is back with dazzling new games from a fresh 1,000 game chess match against Stockfish! Don't miss this brand new game analyzed by grandmaster Robert Hess as Alphazero pushes the boundaries of understandings of chess, chess engines, and artificial intelligence!
Read our report ➡ www.chess.com/news/view/updat...
Sign up for FREE online play: www.Chess.com
Follow us on Twitch: / chess
Like us on Facebook: / chess
Follow us on Twitter: / chesscom - Ігри
You have to go forward like 20 moves to explain what alpha zero is doing
AlphaZero clearly never listened to GM Finegold. Never play f3.
It's often weak against the KID, but not against the Sicilian Dragon. You need to know how to play it.
Stockfish was white
I take that back f3 is strong against the KID as well.
As a Go player who also know chess, I know the feeling of the inhuman move.
Why does being a Go player make you know that feeling?
"INHUMAN REACTION"
Watch the AlphaGo documentary, it may offer some insight
@@lucassganderlla7305 what does that have to do with him he doesn't play vs alphago
@@eyeofhorus1301 you're a real sack of sunshine, aren't you?
Inhuman, it's just a move from Tal's universe!
Even tal can't play this move, that's a move with no visible benefit until it all made sense 10 moves later, no human can play like that
4:32 even 3300+ rated engine is silly
Isn't it technically Stockfish vs. Alpha Zero? I thought chess matches were always listed white player vs. black player .... was confused for the first several minutes ...
This is correct, usually the white player is listed first. I’m guessing they put AlphaZero first since it’s probably much more likely to grab someone’s attention compared to “Stockfish”.
Magnificent attacking chess. Similar to Kasparov's energetic style of play
This is a great idea they should make an AlphaZero version that learns the human short sightedness in the game. Then AlphaZero plays less optima moves but in a stylish way. It would wow spectators even more. It is like a GM that gives a handicap to an amateur to show off fancy moves
Human: *Write that down, write that down*
Alpha Zero's flanking pawn.. every high level game. Crazy stuff.
I seriously think that even grandmasters can't really analyze these computers' plays. These computers must be looking hundred of turns ahead that human's conscious minds can't hope to comprehend.
The boldest move of all is the king sacrifice.
Even if the king dies, it's mate in 15 moves
King sacrifice is possible by losing the first couple of games then making a huge comeback.
LoL yOu CaNt SaCrIfIcE tHe KiNg So StUpId
Never play f3
Terrible
f3 is not the problem
@@mentalrectangle Incorrect
@@VACatholic f3 did not cause any problems in this game.
@@mentalrectangle The truth hurts
i remember a game in which tal played kings indian with black pieces ...against samich variation ...and tal sacrificed the knight within 10 moves .....that was a game i remember is similar to this game
Fascinating game and excellent commentary!
Another great idea with AlphaZero is that they should allow AlphaZero to take away a piece on the board on it's first move. AlphaZero may learn and show that the certain pieces like having more pawns make you weaker not stronger. Therefore it would be stronger to play a game starting with 7 pawns instead of 8 pawns. This goes against the concept of early material advantage.
Agreed
such setup would be useless if not counterproductive since it's an illegal board and cannot be encountered during normal play. Yeah, AlphaZero might win such a scenario but it's meaningless in terms of a normal play.
@@infinummjb technically can be encountered if we start with a bunch of knight moves!
Never mess with ALPHA ZERO
How deep does Stockfish search in before n-f4?
Prawns, horsies, castleheads, relishops, crownhead, crownheadV2.
Different. Knowledge is infinite.
What a monster!!!
at the imaginary scenario of white defense at 3:19 after black Re2... instead of White Rf3, isnt Nf3 stronger so it's set to take the pawn at h3 later which is crucial. If still Black Nxf4, then Kh1. If second black Ng4 pins the white knight and threatens h2, then white Rc2 and an exchange of rooks somewhere. With no immediate forks or checks to worry about, so the White queen can chase the black N-R pair out of there. And with no light-squared bishop for Black, the white king is temporarily ok in the corner instead of being chased on the run as when white Rf3 was played.
I would like to watch the AZ-SF 8 games with both eval bars displayed. I would pay money to see that.
It’s funny that we categorize Alpha Zero’s moves as “standard theory” seeing as how it was never taught any form of standard chess theory. It’s quite simply picking the most pragmatic tactical move.
when he says that it shows that chess knowledge is incremental it cant be a coincidence that humans over a hundred years and this computer play some of the same openings.
Mustapha D
True, and it shows that Artificial intelligence does think along the same plane as humans do. Yet, at the same time, is not constrained by presuppositions and previously held notions of what the best decision would be in any given situation. Whereas Stockfish has been given every move played in every position of every chess game in history. Plus it thinks of 60,000,000 moves per second whereas Alpha thinks at a rate of 60,000 moves per second. It completely changes the way we can look at chess and any subsequent problems we may face as a society.
@@mustaphad1319 Try 500 hundred years lol, A0 was even playing the Spanish
@@risskitt well, in fact it is "constrained by presuppositions and previously held notions of what the best decision would be in any given situation" simply by virtue of learning from experience. If it never encounters some bizarre random board to learn new tricks from it just won't learn them. AlphaZero thinks roughly as humans do, it's just quicker and hence better than humans.
AlphaZero's knowledge of chess is in essence a huge matrix of probabilities that return mostly great/best moves for each chess board you throw at it. Humans just aren't able to build such big "matrixes" in their head and certainly can't process them as fast as Alpha0 can.
6:10 why can't white play Qb1 or Qc2 and replace the bishop after the rook sac
❤bebeautiful lesson and demonstration.Thank you very much
What "train rides" are these people talking about? AlphaZero train picked up Chess one stop then dropped it off the nearest stop on its journey towards science and other important things.....geesh!
Wait, a0 is not technically engine, or is it?
It's scary to think that in the future, computer engines will teach humans to play like this and we will have super GMs who are even more super than modern super GMs.
At 4:33, I would've played knight capture on g4 instead.
but then alpha zeros bishop can take knight, and pawn would have to take bishop and then alpha zero could still knight e4 but without white having a knight that could defend it
@@piAustin But trading piece when you're up piece gives you more chance of holding on to the game.
@@swordwaker7749 But prior to 4:33 aren't they even in pieces? Also I would consider Alpha Zero to have better piece positioning if white did go for that trade.
But then again, I'm not a chess master, so maybe you are right lol
@@piAustin Not a chess master either, but my stockfish analyzed that even if it couldn't defend the b2 pawn, it would still win. The stockfish saw bishop takes b2 coming, but refuted that line with some crazy tactics into a winning position. I would not argue with stockfish on tactical outcome, but alphazero's queen to e2 was not accounted for. At position 5:34, stockfish might've seen this already. (Stockfish agreed with a0 on most moves leading to it.) At full depth of 23, stockfish still sees this as winning for white.
@@swordwaker7749 thanks for the analysis, I see what you mean now
4:13 I hate doubling up my pawns so much that I would have taken the black night at g4 with my white night out of habit. Would have turned out better.
such crazy machines
only time stockfish played like a human. meaning it realizes after the game it was not winning but not while playing
Why aren't strong engines able to calculate those series of moves like the position at 3:10 ??
I don't get it. Does it only seek immediate compensation?
A0 is looking deeper using it's own evaluation. It seems to look at the entire game (opening, middlegame and endgame) as a whole instead of trying to concretely analyze which move gives the best immediate results within the next 30 plies or whatever
what if instead of Nxb5 stockfish 8 played Bxb5??? someone explain please!!
its even worse because tje bishop doesnt block the checkmate on g2
How do you analyze Alpha Zero`s game? Using Stockfish?
Using his own brain....
hindsight
What was being used to analyze this game?
Deion Bonner stockfish
I've seen this guy somewhere before. I wonder where.
Probably on Twitch
I know, I felt the same
aren't all the moves he makes inhuman since it's an engine lmao
Most of if not all engine moves are learnt through libraries. The point here is this move doesn't exist in any known human game, i suppose.
Most humans wouldn't do this
@@vitoryugojsm pretty sure alphazero learnt chess only knowing the rules and not through libraries tho
@@JakYTTV Sure. All other engines usually learn through librairies though, maybe that's why alphazero has a tendency towards these type of unconventional moves.
@@vitoryugojsm They learn through libraries? I don't think so... They may use opening book and endgame table bases, but that's it.
Classical engines mostly use minimax algorithm with an evaluation function. The "learning" you're referring to all occurs in the evaluation function, and maybe some pruning parameters and such for minimax. Learning is not really the right term though, humans adjusting knobs is more accurate.
this is insane calculations. light years ahead of carlsen or any human
Carlsen could beat Alphazero or Stockfish easily, he just needs a baseball bat and access to the server room =P
@@seangraner297 LOL good one
@@seangraner297 lol
It does not make sacrifice. It already knows every possible outcome and proceeds deterministically.
AI is already the chessmaster and is teaching Humanity how to play chess ❤
So, a magical move that isn´t winning, nonetheless creates something on the board and gives white many many chances to blunder.
Of course, we are talking Stockfish here so it didn´t. It wouldn´t accept the sacrifice in the very first place if there was a crushing refutation awaiting.
I want to see AlphaZero V.S. Stockfish 10
@Deep Mind X bruh Leela lost 15 and won 1
So why don't they play the same game each time?
The key move in this game 18...Ng4!! was also found as best in the position by Stockfish 10 app in only 9 seconds - running a whopping 7 cores on my little cell phone(samsung s9+). Score -0.19 depth 35 at 4:00 minutes and 0.00 forced drawing lines are first seen in just under 7minutes. Clearly Stockfish 10 on faster hardware (s9+CELLPHONE!!!) could mind muscle that Alpha Zero into AI migrainaritaville, and that is why DeepM not showing any games this AZ vs. SF10. They most obviously have played them against each other, just not showing any of those games...yet. There is probably an OmegaZero coming soon.
Leela chess zero beat SF 10 in SuFi 15, (Leela chess zero is the same type of system as Alpha Zero).
its not inhuman i make deez sacrifices all the time, but i get nothing in return
"It's Not Human Yet"
Game 5 watched ❤
Just for reference, Stockfish 10 is +112 Elo to SF8.
From what we see with Lc0 project, Leela under-performs against weaker opponents so it might very well be that A0 could be as strong as SF10 (+/-10 Elo or something similar).
Taking a look at your sheets, Leela ID 11258 is 44 Elo stronger than SF8, but tested against SF9!
What is 11258 "performance" against SF10 and what is performance against SF8 (and maybe even SF7 and SF6)? Is this tested, I'd really like to see those results as it might give us better comparison on where Test30 currently stands!
Just for reference, AlphaFold: now using AI for scientific discovery.
This seems more important. Think about what it would mean to give the same time and importance to chess only, as Stockfish does!
@@LawrenceGreco Indeed! AlphaZero's primary purpose is to make the world better through using AI to solve practical purposes. And for that, I deeply respect Deepmind.
@@zlac I have tested SF10 against ID 11258 and found an Elo difference of -68
@Bryan Smith Because Deepmind's homepage specifically states "Solve intelligence. Use it to make the world a better place."
You don't need to be a top-level Alphabet executive to understand the primary focus of Deepmind's AI projects.
So far they have used AI in practical purposes such as protein folding and eye disease detection. This holds with what is said from both Demis Hassabis, Deepmind's CEO, as well as what is said on Deepmind's main website.
These are not humans, lol you said it
Alphazero chess engine please sand link
lol nice calculator
In future we found alpha zero chess engine in play store .plz rply❤️🙏🏿🙏🏿🙏🏿🙏🏿🙏🏿🙏🏿
How is the draw decided?
Draw if no pawn moves in 50 moves or no piece taken in 50 moves
this guy was 'muddling' up a perfect game with his interjections...i kid you not.
Is it possible to play against AlphaZero?
No, it's proprietary code that google won't release. Which sucks because Google relies on open source free software for a lot of their products, but they don't want to give back.
Ass Whole Ah. Yeah, that sucks... Would love to play against Alpha and arrange matches against Stockfish. Hopefully in the future it will be possible. Thanks for your reply.
@@user-yf8il6we2z Well, due to AlphaZero's nature as a deep learning algorithm and not a programmed engine, it doesn't make sense to ship it out. There is no one "proprietary code" that is AlphaZero.
@@sentient4248 the lc0 project has replicated the neural network engine from alphazero and believes they have surpassed the strength of the publically shared alphazero games. Have you had a chance to experiment with it?
@Ananthakrishnan K both lc0 and stockfish from today have greatly advanced past the strength demonstrated by alphazero. Lc0 poked some holes in SF and the SF devs learned from the experience. Same for the other direction.
Do humans even stand a chance against sf8?? Or is it too strong?
Well, put it this way: Stockfish 1 had an ELO of about 2900 (which is Magnus Carlsen's current rating). Every other version was 3000+
Humans can draw sometimes if they have an exceptionally good game.
y stockfish lvl 8 only?
Look at the date
In future deeep mind releases alpha zero chess engine in Android plz rply ❤️❤️❤️👍👍
Fact that A0 allowed perpetual check and draw in position where he had an extra queen , is still very disappointing . Looks like he didnt get what he wanted . ( assumed machine can be "male", but Alpha plays like male player )
Eh. As black most players (and even engines) would be content with a draw. Of course, AlphaZero may not be as tactically proficient as Stockfish so for all I know there may have been a winning combination that it missed, but I kinda doubt it.
until alpha zero wins in an event like tcec or the cccc i will never belive that it is the best engine
In that case, just keep your eyes on Leela Chess Zero. That program is based on, and for all practical purposes *is*, Alpha Zero. So if Leela ever wins either of those tournaments, that is the equivalent of Alpha Zero winning it. Basically, Leela Chess Zero = Alpha Zero.
Matthew Viramontes Leela is probably stronger than this version of alpha zero now. (It took way longer). I’d like to see more professional help on leela and letting it use tensor cores
@@TheHomicidalTendency no, that's what Deep Mind is trying to avoid. This gets beyond chess here and into science, particularly the ethics of science, for the point of Leela was to *independently* verify Deep Mind's results, and so if Deep Mind allows us to use it's TPUs to better Leela, that taints the whole thing. It wouldn't be pure verification from a scientific standpoint. People just need to have some patience. Leela is almost at the point where she can't be defeated, getting either wins or draws, and the next step is for her to get more wins, enough to win a tournament. It's gonna happen. Leela isn't even a year old yet, and she's almost there. Perhaps just a couple more months and she will be.
You are implying that DeepMind rigged the match. You have no right to say that unless you have evidence.
no I am not implying that the match was rigged, that would be pretty stupid, since everyone could analyse it with Stockfish. But Alpha Zero is running on purpose build hardware, while Stockfish is running on generic hardware so it can not possible be a realy fair match.
All that just for a draw lol
That was a weak Stockfish playing!!
I mean... not at the time the testing was done, no.
And that was a weak alpha zero playing!!!
@@DrunkMelon It wasnt a weak alpha zero...
@@DrunkMelon Would you think Google would use a weak alpha zero? lmao
@@iamlegend3927 the point is it gets better every second. so by now its really weak. and yesterday's alphazero is weaker than todays
‘Twas a draw! Why’s everyone always kissing up to Alpha Zero?
erm
Because AZ beat SF, one match was AZ 155 wins, SF 6 wins.
I can beat alphazero and stock fish easily
Was Stockfish limited in this match as the previous one?
it seems like they used stockfish 8, stockfish 10 (current one) is about 100 ELO better. Maybe these are just some of the unpublicized games they played already last year or whenever that was.
@@Quitchy you're probably right that self-play ELO isn't a very true representation of the performance against AlphaZero, but it should give a hint.
Regarding the cores, there could be more but 44 is already a lot. However, they said they used just 1GB of ram. Since hash tables are very important for classic chess engines that could have been a performance bottleneck (on my notebook with 8 cores 1 GB are already at 99% capacity after 60 seconds).
@@dariusduesentrieb They used the TCEC settings so it should be good.
@@mtriptube TCEC allows up to 16384 MB, at least currently. I have to admit I am not sure how important hashtable size is, in very short time controls (bullet) a smaller hash table is even better (at least for my engine).
@@dariusduesentrieb the error here is one is comparing processing power to an iterative self learning machine
Bam! xD
Doesn’t seem as impressive when it doesn’t win
If both parties are playing perfect moves the whole game the only outcome is a draw.
It's impressive because what stockfish's calculations based off perfect play from its own pieces showed that what wasn't supposed to be a draw game nevertheless ended up being one. Eerie.
@@chistopherr7536 the perfect moves are those moves what the engines ,for it´s own ,calculate as best move,so for sure stockfish would never have claculated Nb5 as best move
Alpha Zero's final score was 290 wins, 886 draws and 24 losses
Blank is still stronger XD
Who would win in a Chess match....
Current Chess champion Magnus carlsen or Alpha Zero?
My pick Alpha Chess
Defends on what version of AlphaZero
we know stockfish being rated 3600 versus magnus carlson at 2800 that it can by definition beat magnus carlsen as easly as a he can beat up on 2000s
the top chess champions do not stand a chance against even the mediocre computers these days. Carlsen would get obliterated
this version is sf9 according to update on chess.com article
SF 8 (old at the time) on crappy hardware and conditions, not an open event (not announced to anyone), just a closed AlphaZero promotion basically on the very best hardware imaginable. Very fair indeed
This is different match. And also anybody who actually looked at the games in the first match could see that AlphaZero was doing something special.
@@Haltung You mean the games they chose to promote. This is a great new technology and no doubt it is promising, not only in chess but in a lot of areas; but still if you have a sense of fairness you would like a fair match
@@VictorAlexanderFiltenborg did you read the new paper? It was a fair match this time, A0 killed it
@@carrottoponcrak Anyway it's match with old SF
@@Haltung this is different match? Which one then
outstanding move
I think people should enjoy the beauty of the game rather than getting upset about which chess program is better.
AlphaThanos
Is it just me or does Robert Hess look like Bryan Callen from MadTV?
#2nd 😛, I was kinda waiting for a0 to crush s.fish
I am used to using computer assistance when playing chess online at chess.com. fortunately until now my account was not banned by chess.com.
It will be sooner or later
Lol why the hell would you do that
I like chess videos, but I really don’t like how they skip through the openings without explaining what’s going on and what they are aiming to accomplish.
MalikEmmanuel look at agamator he is great to watch
Spoiler alert it's a draw
stockfish drew, give em some slack, against arguably the most advanced engine, nice move yes but alpha didn't get the w, so they didn't play well enough
Elie Obeid I agree with this. Unless alpha zero plans on drawing with black and winning with white as it’s goals with each colour then two engines meeting at a draw shouldn’t be too much praise for AZ.
What is so impressive is that Stockfish valued white's position at 1.5, which should be winning. AZ knew it wasn't nearly 1.5 in white's favor
The amazing part is that alpha zero can draw games with piece sacrifices, which in the long run means we will learn so much more about chess from alphazero than we did from engines.
@@ElliotSveum 1.5 is not winning by any means generally speaking.
obviously it depends on the position and such, but you cant just go and say its eval was 1.5 so it was winning...
the amazing part is alpha zero is self taught . Even if A0 had lost a few games and drew other it still would be very impressive
AlphaZero is clearly not human, No human can play like that.
I can already see the amount of people missing the joke 😂
So many of my doubters in the comments section.
Just a program. And winning all the tiem for Yout marketing ofthis. No special characteristics,))))
Alpha zero has to be cheating!!!! JK
Chess is such an elementary game. You need to discuss what scary, crazy, and amazing shit that AlphaGOZero did when it comes to GO
#first
No offense ,but could you show the complete game *first* and *then* do your analysis?
Fawn pawn !
FAWN PAWN HAS BEEN INSTALLED
Stockfish=caruana
Alphazero=carlsen
looking forward to the draw
CANT SPELL CHESS CHAD-HESS
The match result of Alpha zero vs Stockfish over 1000 games shows Alpha Zero is only 52 elo stronger then the version of Stockfish Alpha Zero played. And this is with Alpha Zero playing on much faster hardware. The rating difference between Stock 8 and Stockfish 10 is more then 52 elo. I am getting off the Google hype train.
Thanks for understanding! The results of those 1000 games may look impressive, however once you calculate the Elo difference and take into account that Stockfish 8 (-112 Elo to SF10) was used, it's not that impressive.
its still a self playing computer that can learn any game better than humans. it will be amazing if they can make one better than alpha zero
There's a big difference though between Alpha Zero and Stockfish in terms of how they learn and evaluate chess positions...
Stockfish is essentially an algorithm developed by humans that analyzes chess positions based on how humans evaluate them (piece count, squares controlled, development, tempo, etc)... while Alpha Zero developed its own algorithm without any input from humans aside from the basic rules of the game.
So not only is Alpha Zero better at chess than Stockfish, but it's also better at teaching itself chess than humans are at programming a computer to play chess... which if you think about it is quite amazing.
Mistaken view. AlphaZero plays to its opponent. It crushed SF9 the same way it did SF8 and SF used the preferential hardware for its kind of engine.
@@Add9Sus4 AlphaZero does not develop "its own algorithm". AlphaZero improves through the gradient descent of its neural network.
Stockfish on the other hand is a handcrafted heuristic.
Based on evidence so far, AlphaZero is weaker than Stockfish 10. As an example, AZ's performance against SF8 on the 1000 game match was +52 Elo to SF8. Stockfish 10 on the other hand is +112 Elo to Stockfish 8.
Anyone else not impressed? Even if AlphaZero didn't lose a single game, and the results were 161 wins / 839 draws, the elo difference between the two engines is only 56 points. Stockfish 10 is apparently 50 points above Stockfish 9, and about 100 points above Stockfish 8. I'd like to see AlphaZero play against the most recent iteration of Stockfish on identical hardware.
Stickfish 10 can easily beat Alpha
@Ricky AZ *is* playing against SF 10 on similar hardware and in a public tournament, in Chess.com's CCC tournament, under the name Leela Chess Zero. Leela is Alpha Zero. The only difference is that Leela wasn't trained using TPUs, and she's not playing on TPUs. It's ironic, for Leela is exactly what everyone wants, Alpha Zero training and playing on layman's hardware, but no one sees it because it's named "Leela". Think of it this way: Alpha Zero and Leela Chess Zero are identical twins, but AZ went to expensive private schools and Leela went to public schools. That's it. Otherwise, they're the same thing.
@dididididid Inform yourself AZ was not running on a super computer.
@@matthewviramontes3131 leelah is to weak, to slow compared to A0
@@shezmu5324 Well I don't think Leela has played almost 1B games yet haha
S T O C K F I S H 8
"To evaluate performance in chess, we used Stockfish version 8 (official Linux release) as a baseline program. Stockfish was configured according to its 2016 TCEC world championship superfinal settings: 44 threads on 44 cores (two 2.2GHz Intel Xeon Broadwell CPUs with 22 cores), a hash size of 32GB, syzygy endgame tablebases, at 3 hour time controls with 15 additional seconds per move. We also evaluated against the most recent version, Stockfish 9 (just released at time of writing), using the same configuration."
these are 11 month old games. not 'fresh'. it's a mystery why they are published so late. yet, all 1000 games are not published.
It was always going to be hard for something selflearning, lowish calculative to outperform something in a game so closely solved as chess is. It doesnt bend the laws of physics.
Superhuman ai is around the corner, and if we get it right we have a universe of potential before us. Global warming and whatever else is a mirage, a successful superintelligence will be able to solve any and all problems. We just need the one chance to get it right
What are you trying to say? Alpha crushed stockfish, just read the article
That racist thumbnail though. No wait...
Stockfish 10 is still stronger than A0.
1. Stockfish 10
2. Stockfish 9
3. Alpha Zero
4. Leela Chess Zero
5. Komodo
6. Houdini
7. Stockfish 8
How do we know stockfish 10 is better than AZ if they've never played? AZ is good and so is SF10. I feel like it'd be 100% draws.
Wrong. It is believed to be stronger, until they test it there is no way to know. Also alpha stomped stockfish 9 on this same test.
@@shezmu5324 They also said in the paper that A0 also was able to beat SF8 TCEC settings by an even higher margin than the official results shown, when Deepmind allowed A0 to play a lesser move by a tiny tiny tiny amount. No doubt it would crush SF10 in my mind. They can always choose to have it play itself for a month
@@softan This is not my conclusion. It's a fact based on the actual rating comparison against SF 8. Stockfish 10 is 110-120 ELO stronger than Stockfish 8 compared to A0 which is only around 50-60 ELO stronger. A0 real rivals are SF9 and Lc0. They have about the same ELO rating against SF 8. It's like SF 10 is Magnus, SF 8 is Wei Yi, and A0 is Nakamura.
@@Nuhyamin1 Leela chess zero, (which is the same type system as Alpha Zero), beat SF 10 in SuFi 15. 53.5-46.5 :)