ฝัง
- เผยแพร่เมื่อ 12 ก.ย. 2024
- IM Danny Rensch is back with another AMAZING Leela Chess Zero game! In this one, Leela offers an incredible piece sacrifice on move eight! Watch this epic breakdown and check out the Computer Chess Championship at chess.com/ccc!
Sign up for FREE online play: www.Chess.com
Like us on Facebook: / chess
Follow us on Twitter: / chesscom
Follow us on Instagram: / wwwchesscom
#leelachesszero #computerchess #chessengine
Watch more computer chess games here!
th-cam.com/play/PL-qLOQ-OEls7uZ_Kr0zY8AgfYavPtUaH_.html
Please give me pgn
Both engines are open source, but LC0 is an evolved algorithm. I doubt there's lot of selective pressure to make perfect moves when you're so far ahead of the opponent, which explains the sloppy endgames.
Stockfish: *plays a normal-seeming opening*
Leela: so you have chosen death
This is insane
Tbh: Giving up the centre, with ...e:d4, while simultaneously moving two minor pieces twice (by retreating) by move six, with black, in an open position, only to follow up with the rather crude ...c7-c6, smells a bit fishy. ^^
😂😂🤣🤣🤣😅
@@Badbentham A bit Stockfishy.
To be fair after analyzing with SF for 23 min the eval increased from -1 to -0.17. So SF doesn't think taking the bishop is much better for black.
Stockfish and everybody else: "Study the endgame"
Leela: "lmao piece sack on opening, troll endgame moves yeet"
You don't need to study endgames if you don't have to reach it.
@@AdityaPal_sciencepal Big brain. Just get mated or mate in the middlegame.
"sac"
My god that was stunning. Lc0 gave up a bishop just to get a pawn on d6, which essentially trapped two pieces on the queenside for the entire game. The very pinnacle of positional play.
The power of the THORN PAWN! Not just one of them--two of them! Black did manage to get rid of one, but the thorn pawn on d6 was worth more than a passed pawn, since the black pawn on d7 gave it shelter.
@Mitchell chess Yes alpha does it as well. They play chess like a boa constrictor, slowly wrapping you up until you cannot move.
Paul morphy is the pinnacle of human positional play
"This is really inspiring, might play a game myself"
*Proceeds to get slaughtered by a 600elo*
"Engines dont sacrifice" -Gary Kasparov
"Hold my bishop"- Leela
But leela is not an engine she is an AI
An AI which learns chess by self play
@zamsu gowasu i wonder how boring is your life.
@@imroseuddin4680 Projection
@@immortalbrolyblack2891 Yes, she is an artist. She moves in new ways. Her games are poems, unfolding ever exploring possibility.
One minor suggestion: When going down those analysis lines, it would be so nice if there was some kinda filter on the board i.e. the board would be a different color. This way it would be much easier to follow the game and know which moves were actually played and which ones weren't (especially when you rewind etc. a couple times it's so easy to lose track).
Very nice game and good analysis :)
Fantastic idea. Although in this video i had no trouble understanding which moves were in the actual game and which were not
futurefox128 I love that idea, or even moving the position of the main game onto a mini board in the corner...
Yes futurefox, same here. Tried to play along with Arena on Stockfish 10, just to watch the evaluation Stockfish gives ;)
It's hard follow up and to recognize at what positions the analyzing starts leaving the actual moves....
Peaches. This.
I agree. It is not always easy to tell when in game or analysis.
Thanks Danny! Chess games by Leela: "There is no spoon, only forks."
@JediNxf7 go fork yourself
Sure when I sacrifice a bishop its a blunder but when an engine does its double exclam 🤔🤔
EDIT: Peaches
yeah same
idk, if you feed this move to Stockfish it would probably still tell you it was a blunder even if its actually winning :)
He said pidgey tho
@@ademiw5467 lol he didn't say pidgey he said peaches
I make double exclams sacrifices all the time. Sadly my opponent usually wins.
Imagine if some grandmaster discovered this novelty years ago and was saving this for later but Leela showed his novelty to everyone
I think, but this is just one of my thoughts I may be wrong... If you let for 10 days leela vs stockfish and then collect the data you pretty much end the way of playing chess
@@malekagribi6284 It took almost 20 years to "solve" checkers, which is a far simpler game. Chess will not be solved without quantum computing.
Chess won't be solved with quantum computing. Chess will (probably) never be solved. Not enough elementary particles in the universe. It will certainly not be solved within some reasonable human time scale.
Sebastian Dierks Never is a strong word. Point is that once it's solved, all you gotta do is tweak a rule or add a piece. Chess has evolved and will continue to evolve so that it doesn't get solved in the near future. Also, I should mention that solving a game doesn't end it's ability to be played. Stockfish is "better" than Magnus, but hasn't put him out of business.
"Never" might be a strong word but in this case it is not too absurd to use. I also put probably in there just to be safe. At the moment, chess is solved for all positions with up to 7 pieces (ie two kings + 5 more black and white pieces). This endgame table base, created in 2012 with a supercomputer, is 140 TB large. It's huge for a text file! The numbers grow exponentially with every piece you add. The 8 piece database is assumed to be 10 petabytes and you need a supercomputer with 50 TB RAM working for months at an expense of millions of dollars. Imagine creating the database for all 32 pieces...
Also, of course solving chess wouldn't make the game unplayable. You don't even need to add a rule, no human (or cyborg or whatever we are then if we manage to solve chess) could have memorized all the variations to play the perfect game.
"Behold, the hatred and anger of a million innocent chess players losing to the vile Stockfish, they cried tears of despair! And those tears gathered into the firmament, and brought forth a new artificial being: Leela. And that being would go on to TROLL THAT [BLEEP]ING PIECE OF [BLEEP] ENGINE TO OBLIVION and it was GLORIOUS! VENGEANCE IS OUR'S, OH AMATEURS OF CHESS! For Leela hath said 'Eff it. You can have my Queen. I don't need it to stomp you, bitch.'"
... or something like that.
Haha
@@sawaria123 HA
😂😂😂
@Bard Erland too bad it takes brain power, not to make subjective, self preserving, biased(probably on being an incurious idiot), judgements based on an entire group; were not just talking about race, were talking about individuals from all walks of life.
Bard Erland only a moron would say something like that ...
World: Chess is dead
Leela: Hold my Bish
Hold my peaches
what is a bish? im sorry im german, i never heard of a bish xD ^^
lategame she just wants to rub it:
"dude, im like 2 years old 😂 - i havent even had enough time to study the lategame yet"
"who needs accuracy when it's still +M12?"
Leela: So you like theory?... HA
PEACHES
Leela's games are starting to look like alpha zero's, except alpha zero was playing a stockfish 3 generations back
Brett Spivey and how many generation back do you think that is for Google...Respectfully this kind of engine analysis is sadly not close to the infastructure on ML monsters like Alpha.
@Harsh Raval AlphaZero hasn't been improved for years and both SF and Leela are considerably stronger than it now
@@okibelieveyou7553 i'm pretty sure your statement is not at all true....
@@demolition5000 well what do you mean? AlphaZero was a Google project for machine learning. It was a proof of concept they don't care about improving it any further. Once it beat Stockfish 8 shortly after Google stopped the alphazero project. Since then the current stockfish and leela can beat Stockfish 8 by a greater margin than AlphaZero did and also you have to understand computer engines progress very fast. If Google continued with AlphaZero who knows how strong it would be but the point is they have better things to do and AlphaZero is now just outdated as a computer engine.
@@demolition5000 AlphaZero hasn't improved since it was first introduced. If you are not convinced, just look it up yourself. Both Leela and stockfish have improved greatly since alphazero was first introduced.
31:36 "I mean what is going on Leela?? Can you just stop playing with your food!" This one killed me 😂😂
"I'm not a GM but I play one on TV." lol. Well played Danny!
😉
Wkwk
This was such a crazy game, kind of mindblowing how such a simple sacrifice can lead to so much suffocation for black. Loved the video!
Also a big fan of peaches too 🍑
@Mitchell chess Why don't you let Leela play it? It's open source.
Great game, great commentary, Danny! You could actually cut 10 minutes from this video by stopping to apologize for the length of the video )))
Peaches!
OMG!!! Leela appears to have literally turned into Alpha Zero with its play...this is mindblowing.
Absolutely
She is stronger than a0
@@dilyan-2904 Perhaps. But if Leela faced Alpha today, Alpha would be retrained against the latest version of Stockfish or Leela itself...
Leela might not be stronger after that...
@@dilyan-2904 Leela is not stronger than Alpha Zero if DeepMind has a few hours to spare on their proprietary GPUs. Leela is very strong, but Alpha Zero runs on custom hardware.
@@Varney_of_London Training against Leela would be cheating, because that's what the Leela project is -- learning by playing games against herself. Anything that trains against Leela is by definition a Leela clone.
The last word of the video is always "Peaches".
Really? lol
This is amazing. Although for some people who havent watched this video, they will see this in Agadmator's channel.
My prediction.
can you put up a sign that shows when you're showing us what didnt happen in the game, like a big [DEMO] so that we know that the moves youre showing us arent the ones that are in the game because i feel like i cant trust what im watching cause any second you could turn it back like 10 moves
@@AgustinCaniglia1992 he talks fast
@@badmanjones179 Maybe make the board black and white during variations.
Danny: Analyzes a game brilliantly. Also Danny: “the last word is peaches”
Danny: has a shirt that says "My kids play outside"
Danny: is the CEO of a chess website
Awesome! The sacrifice makes all the sense once you see it on the board.
I knew this move a long time ago but havent had a strong opponent to demonstrate it since stockfish likes to play 199 moves just to lose a losing game
I honestly believe that if Bobby Fischer were alive to witness something like this he might've actually been really stoked about it. He often complained he hated what chess had become, a numbers game, purely a calculation. He had a deep dislike for traditional chess engines, as well. He had felt it had lost it's "creative spark" in the professional field. True, some of it was probably his own deluded pessimism in his old age, but after viewing this demonstration, I believe there were actually some credence.
Think it's worth bearing in mind that you can only play these kind of long term piece sacrifices if you have Morphy/Tal/Nezhmetdinov type of tactical genius to back it up.
Yes you need perfect precision to convert the advantage. For example I totally didn’t know how to handle queenside development in the case of white playing b4 (1… a3 2. bxa3 b4! was brilliant)
Lela: makes a piece sac
Stockfish:*cries in tarter sauce*
😂😂
we mortal can do a piece sac too, like I always did but then most of the time, I resigned after few moves.
That was like getting seduced and then hypnotized, geez. peaches Danny
lol who remembers when danny criticized other people's thumbnails that portrayed Lc0 weirdly but did the same thing for this vid? i'm not hating on danny though i love him to bits, great analysis as always
Not sure but I think he was talking about agad's thumbnails(at least thats what I understood when be said that)
Love me some agadmator videos!
"Leela believes in it. So I'm a believer too." Morphy nods from the grave.
I wish i could play chess like Leela does. This is God level of chess right here. Truly amazing. Thanks for the commentary Danny, i really like you showing all the possible lines for the less experienced players that would think there would be other possibility to counter the initiative. What can i say ? Mind blowing ! 👍
I do not because I would have no human player to play I would crash the best players in the world very easy :D
Now it's time to organize the TRUE fight of the century...
Leela vs Alpha Zero.
Can you imagine how insane their match would be!? Both sides sacrificing pieces all over the place..
I would pay good money to see such a match.
All of the games would actually probably be boring slugfests with many draws. This game is an exception, and such sacrifices happen very rarely.
Leela playing with the food in the end!
It appears that these self-learning neural networks just find out that smothering your opponent's pieces, locking up your opponent's minor pieces behind his own pawns, unable to do anything, is a very strong and working tactic. AlphaZero was fond of that, and it appears that Leela is too.
This is the best chess game analysis video I've ever seen.
Peaches btw
The problem missed here is the move 4...Bb6 is a forced incorrect move order to intentionally put black in a bad position (the theme of that TCEC was to put positions on the board that are very bad for one side in order to produce more non-draw results). You can't transpose into the correct move order because white DOES get to play Nxe5 with advantage. The correct move order is 4...Nf6 first.
That was great analysis, Danny. I really like Leela wins. They're really cool.
♕
So we can Open a Business on this C7 Square 😭😭😭 I’m crying Lmao 🤣 @18:28 👏🏾👏🏾
Thank you for pointing this game out, and your excellent presentation.
This was a very enjoyable video.
The altered variation analysis was
just delightful.
12:48 screaming for Ne4.
Leela will revive the Halloween Gambit - I can feel it!
Pawn structure is very similar to Four Knights Game: Halloween Gambit. White down a piece, but has big center, pawn on d6, nice development, and black king is very suspicious.
Yeah I can see why you enjoyed this game Danny -- definitely worthy of the Morphy reasoning you mentioned. Just peaches, lol.
Yeah, and unlike every other chess engine I've seen or heard of, only Leela and her (correction) brother Alpha Zero play more like a human too -- overall with plans based on weaknesses and priorities over just material value! Thx -- great game and analysis!
The way they play really makes you feel like you could do it too if you were just a bit smarter. It's fun chess to watch
It's the brother Alpha Zero
@@vanessaingenhoff4343 thanks, I was wondering about that when I wrote it. Corrected.
"I don't care how long this video is" The power of online content.
But it's honestly hard to follow the line they actually played over your analysis going back and forth the whole time
These deep learning algorithms are so powerful! Leela analyzed the COVID19 pandemic and learned how to put into quarantine all black pieces!
The thing I constantly see Leela (as well as Alpha) better at then most engines is understanding that if you can keep a piece like black's light square bishop from ever entering the game that can be worth at least as much as capturing that piece. If you can keep the bishop from ever entering the game, and in most lines neither rook even enters the game, well you're up an incredible amount of tempo.
Mikhail Tal would be proud!
As always awesome analysis Danny. One nice point that could have been mentioned is that despite the LSB sac early on, a lot of the infiltration is done on the light squares as black never really got to use their own bishop before it was to late anyway.
Also peaches, there you go !
_Last word if the video is "peaches" do I get a prize?_
I think what SF failed to see (early enough) was how Leela would continue to double down on the major piece sacrifice to keep her attack alive.
alpha go did similar things. once it believed it had won, it wouldn't make progressing moves anymore, it would wait until it was clear that it had already won long ago
The chess is never ceases to amaze us. That's why we all love it.
I can't understand the 55 dislikes. You are getting a free 30 min complete analysis of the game between the best two engines of the world. Wtf else are u waiting for noob!
At first I grumbled when I saw 33 minutes. Then I started watching and couldn't stop. A peach of a game for Leela.
This is probably a better version of a Halloween Gambit because white has already castled and Qf6-Bxd6 isn’t a resource for black.
Liked before I even watched the video. Edit: Peaches.
Around min. 25.00, you suggest Bf6 for black, expecting Qf4. However, I think h4 instead is pretty strong for white. Such a typical Leela move! That's probably the reason why black went for h4 instead of Bf6. Beautiful game, btw!
can we have the pgn for the game please.
I followed the video by replaying the game and having stockfish analyze all the moves. I just loved that after almost every one of Leela's moves Stockfish's evaluation usually jumped to something like -1.00 but then slowly returned back to around 0 and even positive.
Infectious enthusiasm Danny. You’re a peach.
People be thinking the last word is peaches, but of course it's Pee-chess, a reference to sPeechless, which this move by Leela left us.
Beautiful chess! Thanks for the analysis. PS Peaches
I'd say Leelas plan for bishop to f6 was h4. (Edit: ... and not natural response queen to c5)
I’m a beleelaver, too.
About the last comment on Leela's endgame, probabilities is right. The way these algorithms determine the move to play is by creating a probability for each possible move. The better a move is according to the algorithm, the higher the probability. The move with the highest probability is played. However, to avoid the algorithm from always playing the same move, it is customary to add a small amount of noise to each probability as well. Normally, that does not matter if one move is far superior. However, if all moves are winning, the probabilities for each possible move are very near each other. At that stage, the random noise that is added suddenly stands out, and can lead to very strange moves that seem very sub-optimal.
Credentials: Master's degree in Artificial Intelligence
Alpha zero:" so you try to take my entertainment now.....you have chosen death"
Leela: I'm sorry
Alpha:" I'm sorry as well for your destruction.
6:50, actually bishop b6-f2 check, if rook takes it will be kind of fair trade as rook and pawn are lost in exchange for bishop and knight while king steped up to g2 opening up himself. If king retreats to h1, knight g3 check and knight taken by pawn h2-g3 and met by bishop f2-g3. Black lost knight for 2 pawns with 2 lines without white pawns...it seems acceptable for black especially after castles in a next move and trying to engage queen somehow in the attack...before that pawn d7-d6 and white square bishop on e6...or g4...but white would make useful moves in between for sure...so all in all...it was not so desperate situation for black there regarding "knight has nowhere to go on safe square"...but the point is ok, white had initiative..
Peaches lol, reminds me Ice Age
So I was dragged back to this video by a reply to my (surprisingly popular, I might add) comment almost two years later. I was interested to see how the newest iteration of SF would see the position after 7..c6. Bear in mind that since version 12 SF has integrated some Leela genes in the form of a neural network.
Lo and behold, 8.d5 jumps onto the radar after just 30s of deliberation by SF14.1. 5 minutes and 1B nodes later, d5 sits at the top with an eval of +1.31 and a WDL of 464/534/2. Second best is Bd3 at +0.86 and 227/768/5. For reference, Lc0 v0.28.2 agrees that d5 and Bd3 are the best and second-best, but is much more modest in its eval, announcing +0.40 and 455/362/183 after 13 minutes and 3M nodes. (Leela evaluating significantly fewer nodes is normal)
This whole thing feels a bit like Deja vu - I remember feeding one of the games that SF8 lost against A0 to SF10, and it was very quick to sense when the game started to drift in A0's favour, all while SF8 was announcing all goose eggs. (SF9 was able to see the position deteriorate but couldn't offer better alternatives; SF10 could.) And to think that SF8 was already much stronger than any human player but SF14 can make mincemeat (or rather, mincefish?) of it without breaking a sweat, it's amazing and terrifying at the same time.
This comment was made by "peaches" gang ;)
Ive never seen a 3000 rated player with nearly all their pieces still on their starting squares 15 moves in to a game, what a ridiculous idea, how has no one found it before?
How about them peaches.
Holy damn, what a game. Thank you very much for the great analysis IM Rensch. Appreciate the work you put into this. The video could have been even longer for me.
Yup. Definitely could've watched another hour of this.
Damn..I've become a fan of LEELA now...which left me wondering what i was doing not knowing LEELA Play till now.
Danny! It’s a 33 minute video. It could have been 3:33 and I wouldn’t get bored. Love your enthusiasm!! Thanks.
With the endgame it is like the algorithm is 'spinning its wheels'. All roads lead to Rome. If the algorithm was to rank a quicker finish as better than a slow one, then it would find the quick way.
my guess is that the weird endgame technique is due to the way Leela interfaces with the tablebase. So giving up material is actually a way of getting to a winning tablebase position. Once that has been achieved every move that leads to a win has 100% winning probability which is why the route taken to the win may seem arbitrary.
Similar situation is in go. There most programs go for highest win probability, so when you play them and they start playing sloppy, they are either completely winning or completely lost, so they basically stop caring about individual moves, because they wont change outcome of the game anyway.
At 16:38 you say the knight cant develop without losing itself. I think before even moving the king stockfish can develop the knight. Hes already up a pawn and a minor material wise. Giving back the minor is probably the best play. Likely Nge7 after leela played e5.
Edit: especially since leelas entire gambit is around having the pawn on d6, there is no way she takes on e7. e7 looks attacked but is actually more likely a safe square.
Stockfish after e6: I'm still okay...
Leela: you are okay when your pieces are out, now they are LANDFILLED!!!
The scary thing about this approach for white is I feel there is no room for error. One suboptimal move without tempo that allows black to move one of those paws and suddenly their queen side unlocks and it all falls apart. But Leela is a peach!
Peaches ? I think a mammoth somewhere is having a baby
underrated comment, dw i found you
After Qg5 the threat is not only Bb6-d8-e7, but the even strong Qd8#. Peaches.
Its really beautiful- these strong engines took chess preparation to another level
Quite astonishing. As I watched I imagined how a benevolent AI could help solve so many problems.
@Danny Rensh It is much harder than that remember that Leela is based on Deep learning architecture and Reinforcement Learning. All these methods are black box in nature meaning that we know that Leela is learning something but we can't explain(make inference) on how she does it. Reinforcement learning with Deep learning drops the inference notion in favor of being more precise about an anwser, so Leela is super powerful( I.E wins more) but we simply can't explain how she does it. Yes it is based on probabilities but no human can tell the way she learns
comment already did not make sense in terms of 'english' after the first half of the first sentence :(
Leela has made this positional sacrifice (to sink a pawn into the opposing Q3 square whichever color) a few times, although i agree it's remarkable this early in an opening.
Although the resulting position is surprising in this opening, it's not that unusual in recorded games, and I've done the same in other situations and yes... it's something that is easily worth a piece when practically the opponent's entire Q-side is paralyzed for the entire game. For the sacrificed piece at least a rook and bishop and sometimes more like in this game never participates in the game.
Also, your description of Leela's well known "trolling" in absolutely dominant positions and giving away excess material is not random and actually provides some insight to its algorithm.
To understand this, the first thing to know is that Leela is uniquely a chess engine that has learned everything through training self-play, even the games it plays in competitions aren't part of the learning experience. Leela starts off given only the rules of the game and then plays against itself zillions of times, and through trial and error develops its own approach to playing chess, completely on its own and without any human input.
This is perhaps how Leela gains "insight" into surprising positional moves and even sacrifices, by self-discovery.
No one told Leela the principles of good positional play, no one told Leela how to create a combination.
Everything you see is the product of a complex algorithm by incrementally learning from the beginnings as a true patzer,, but this is a patzer playing against a patzer (itself) until and as it improves.
So, now back to Leela's "trolling."
It shows how the simple directive to "just win" was given and not "win as soon as possible."
If Leela's trolling was truly random, then randomly Leela would choose faster winning moves sometimes, but that is never the case.
And, if Leela was given the directive to "win quickly" instead, it's conceivable that she wouldn't have gained the deeper understanding she has because numerous possibilities might have been pruned before their true worth was determined.
It seems that Leela is dead set on winning at any cost, and without any other consideration, even time.
And, this may even be a lesson for other applications of this chess engine, that left to itself a win is a win is a win.
And, why a Leela solution to a problem may be world class, but not the same as what you'd expect from an equal world class human.
Very similar (if not the same exact sacrifice in the opening) to of one my game, to me it's clear (evaluating the positon) that with d6 the position it's lost, resign ....as a human....
Ah. Compliments! Well explained IM Rensch. Appreciated video. Thanks.
Keep going.
I play games like this all the time. Like no joke. For real. Not all of them work, but those that do work like this. The essential idea is the same. Structural and organizational damage.
I think Leela's end-game winning logic could be based on the idea that if Leela initially examines moves in random order and finds a play that simplifies to a winning position in two moves, there's no reason to look at anything else. If no play would simplify to a winning position in two moves, then it may be necessary to look for a play that simplifies to a winning position in three moves, but if there's a winning two-move simplification, why bother? Since every pawn push or piece capture resets the fifty-move rule, a mate in three isn't really any better than a "simplify in two".
Danny Rensch asking the right questions. Very instructive!!
13:19 "I'm showing you every stone without unturned" 😂
What an amazing game! Wow. The sac makes sense from a human perspective too because white is essentially playing up a rook with black so tied up on the queenside. It’s amazing the way Leela handles the initiative, really beautiful stuff
Peaches i love how positional play beats material. Atleast in this game. But just the idea. It in a sense re writes the book.
This game reminds me of a game I lost to an 1800 many years ago. Piece up, but no development at all and no way to disentangle.
This game reminds me of the Halloween Gambit that emerges from the Four Knights variation, except it's a better version because black is even more cramped thanks to the white's pawns on d6 and e5 as well as the Bishop on b6 (Notice that at 10:50 the bishop actually prevents the Black queen from going to a5 and black cannot play b6 to free his c8 bishop, which means the rook on a8 will be stuck too)
28:18 Actually, I think that Stockfish was defending against Qd8 by blocking with f6.
Garry Kasparov : engines don't make sacrifices
Leela chess zero: Excuse Mister?
Lololol