24
Aug

Leela Chess beats Stockfish Dev on 43 cores – end of era near for Classical engines?


Hi all, I have an absolutely amazing game
to show you. Of Leela against stockfish dev from last month
in a bonus game that was played at T sack. Let’s have a look at this. So stockfish was on 43 cores. E4 from Leela, the set opening book just these
moves, these four play. So pretty standards and the book Leela plays
Nc3, Nf6. Open Sicilian, we go into the Knight off Bd3,
quite unusual Bd3. But it is in has been played before at high
levels. G6, h3, Bg7, Be3 both sides castle and now
e6, b5 has been seen before in high-level master game with black winning in tempura
against boggler in Dresden 2008. But e6 have from stockfish, Qd2 setting up
a battery to try and maybe exchange off bishops and some points b6 f4 aiming for f5, Bb7,
f5 and stockfish voluntarily opens up the G file here with G takes. That G file could that be useful? Bh6, fxe4. We have nail Qg5. Now black has to be very careful. Not Ne8, because if Qxd8 here but stock fish
has calculated Nh5 here to defend the g7 issue. Qxh5, Bxd4++,Kh1 and now stockfish gives up
the bishop voluntarily, Bxc3. You might wonder exd3, that fails to check
actually. Qxd4++ and Qxd3. This position is just winning for white it
seems all much better. If the rook moves, then there’s Rxf7 hitting
b7 and h7. So for example Rg6 we just take there and
Bxg2 we just play this and then rook have a check winning the Queen. So these lines are all very good. So Bxc3 again it seems as though in theory,
in theory stockfish has taken some risks. But to actually punish this in practice is
a different story. White is threatening now Qg4 and Qg7 with
the absence of that Bishop for example to put it on the board. So that is the threat and it’s parried with
f5. We have Bc4 now not taking the rock. Taking the rook is interesting. But it’s just a small advantage perhaps for
black, for white only. Because look at blacks pawns they’re quite
impressive. So Leela actually refuses to win the exchange
here. Bc4 instead, there’s another plan more ambitious
perhaps with Bc4 just to tease d5 which blocks in this bishop and the bishop goes here, which
has greater flexibility than on b3. The problem with putting the bishop on b3
concretely as an example, say this Ne7 try to open up the G file with g6 here, there’s
runs into Rf5 and there’s no Bg4 to try and kick that rook away and now in the game continuation
the bishop is more useful if there is a need for Bg4. We have Rf7 which might actually be already
a key mistake, just that moved 19. A key mistake here from stockfish. It seems as though Nd7 might be the best defensive
try. For example if g4, f4 and white might have
to just take the exchange here like this. But it is a small edge for white. If white doesn’t take the exchange here and
plays g5 here instead, then it could run into trouble actually with f3 in this particular
scenario. For example g6, Rf5 here and this position
is actually leading for an advantage for black. It seems as there’s no problems for black
there. So it seems as though Nd7 might have been
a key defensive resource. But in the game we have Rf7 slightly more
awkward and we have now g4 here and it’s under very different circumstances. F4 is played. If F takes actually done just Qxf7++, thanks
very much. So f4 on Nd7 here again this is very good
after G takes, E takes check and then Qxf7. So f4, we have now g5. Now here we see the value of the bishop actually. If rock f5 here, then there’s Bg4 kicking
now and if yeah this is just troublesome after Nd7 for example just taking in g6 that’s crashing
through. For example like this, its just crashing through. Yep, so yeah this is a very important resource
here, Bg4. We have now if Re5 we can just take on f4. So we have actually, if f3 in this position
instead, then g6 here is just in time hitting that rock which also is why that Rf7 seems
to be a little bit inaccurate. If takes them we just take their own crashing
through. So in the game we have Qe7 and then g6, Rf5
hitting the queen, gxh7++, Kh8. On Kxh7 there’s Bg5++ protecting the Queen
hitting the opponent’s Queen and taking here the bishops actually doing a great job protecting
the Queen in this line. So that’s an advantage for white. So the Ba2 very important for Bg4 for holding
the Queen, important role compared to b3. So we have actually Kh8, Qg4 frightens Qg8#. Qxh7 is played. If Nd7, Rg1 threatens mating with Qg8. If takes here, Qh4 threatens let’s see this
position to just double the rocks. Which is menacing. This is actually just crushing this scenario. Check-check and it ends up where black is
in big trouble here. Blacks just losing after Bg6 Qh4++. So yeah these lines are just terrible after
Nxd7 just to run that for again. Rg1 is just winning here and just give an
alternative in this, so black is just because white is doubling and increasing the intensity. Then it’s winning this line you might think
also King takes, those Queen d8++ here they seem like that. So anyway Qxh7 was played not Nd7. We have Qh4, Nd7, Rg1, Rf7 and stockfish realizes
here it’s totally busted. It just realizes that white can actually intensify
the attack and there’s no defense. So we have to now see, i’ll demonstrate why
concretely here. The game actually ended because of the adjudication
rules. Rg5 with the intention of doubling rocks is
lethal. F3, Rag1. Doesn’t matter about the bishop. F takes for example. Rg7 is crushing for example Rook takes, Bxg7,
Be5++ that wins the Queen and the game basically after Be6 threatening made that, that this
is just mating and if we look at this again. If Nf8 then Rg8++ takes Bg7 double check and
mate. If f2 then the rock can go here and bishops
actually another useful purpose is to be able to take on f1. So here the bishop is a hero in this line
as well. So this is just hopeless for black, this scenario
it’s just crashing through again. For example like that or let’s have a look
now at Rf7 and this line again with Rg7. So taking only two Rg7 you know Be5++. Yeah we’ve seen this actually. Okay so yeah we’ll go back to the final position. It seems as though stockfish underestimated
the power of the g file attack. Be2 is a really key move though to make sure
the attack could work as well as it did. So a great game there from Leela and stockfish
really doesn’t lose too many games recently as we’ve seen. So this is really quite a special trait. Overall stockfish did beat Leela. But there were quite a few draws. This was the only game in that bonus games
which Leela did beat the stockfish dev on 43 cores. But that really is once in a blue moon event
for stockfish generally to lose at all recently. It’s getting a kind of invincible reputation. However is this starting to be looking like
the end of the era for stockfish and classical chess engines in general? Comments, questions, like, shares appreciated. Thanks very much.

Tags: , , , , , , , , , , , , , , , , , , , , , , ,

52 Comments

  • kingscrusher says:

    Replayable game with indented variations: https://www.chessworld.net/chessclubs/ltpgnviewer32/ltpgnboard.asp?GameID=5023850&v=N4WCsRvBrtc

  • David S says:

    Incorrect evaluation of king safety can still be a "horizon effect" problem for the traditional engines, it seems.

  • Magnus Carlsen says:

    Thanks! I wonder what's the Elo for SF Dev in 43 cores though.

  • mrbdazz says:

    2:08. It always amazes me how the AB engines are so confident to enter a defensive position like this and that most of the time they're right. Anyways, it seems that the AB engines (esp. SF) are still getting stronger each year, and their ELO ceiling potential is not yet in the horizon (pun intended πŸ˜‹). So, rather than predicting the end of era any time soon, I'm expecting fierce competition and lots of interesting games!

  • Abu Abdu says:

    Thnks, nice selection as usual

  • jose raul Capablanca says:

    A fine attack by Leela, I don’t know if we are at the end for Stockfish, or the beginning of the end, but to paraphrase we are surely at the end of the beginning of the age of Leela. Thanks KC keep up the good work

  • Dimitrios Sarbanis says:

    At the 1 minute 34 mark when Stockfish takes the Knight at C3 why doesn't Leela Check with Queen on G4 looks like a forced mate to me

  • Adnan says:

    Stockfish got his but kicked,hard!

  • Bernard Finucane says:

    The self learning engines will get better and start consistently beating the engineered engines until we start to see analysis of what exactly the self learning engines learned. Once that starts happening, the results can be engineering into the "old fashioned" engines.

    Probably the hardest thing about building an AI is "feature engineering", which is identifying what the AI should be thinking about. Self learning automates that process by playing millions of games, and the results are hard to find because they are distributed around the weights of the resulting network, not written out explicitly. But eventually Leelas' ideas will be harvested by classical engines.

  • Julio andres gomez says:

    ItΒ΄s not the age of Ultron, itΒ΄s the age of Leela.

  • ANIKHTOS says:

    And stockfish loose the game. How satisfactory is to hear this words. Later i imagine we will be waiting to hear and leela loose the game lol.

  • realCevra says:

    grats, i wonder where that comes from, since sf6 5spply has still no problem even with the demo hardcore white leela (same old story, 22 moves in sf is up a pawn, at around 50 moves it's up pieces and announces mate in 15 which leela runs into in 10 moves).
    maybe nn power which drastically relies on simd scales better with gpus so suits leela better than multicore cpus suit sf, since whether you have 1 core or 40 cores, that's maybe just 1 more depth in AB

  • pazdziochowaty says:

    I would like to see Leela on ICCF. Best correspondence players are stronger than pure Stockfish because they cooperate with the engine and go through variations to see if evaluation does not drop after some moves. And they can use specialized applications (for example FinalGen) that calculate tablebases for specific positions which allows TB-quality analysis for 7+ pieces on the board
    If Leela without human help would beat them that would be even greater achievement than wins over "pure" SF or Komodo.
    I wonder if there are ICCF grandmasters that already use Leela. Quite possible there are not yet, most players invest in super-server CPUs so they might not have mighty GPUs that are required for neural networks. If you put dozens of thousands $$ in your hardware you will not easily throw it away to put another dozens of thousands in a completely different hardware

  • ΠœΠΈΡ…Π°ΠΈΠ» Π§Π°Π»Ρ‹ΠΉ says:

    What a BS claim to say the least. Why I don't see analysis of all leela losses happened in that match but see the analysis of THE ONLY win? Well, maybe because leela zealots will cry and you wouldn't be able to hype around it.

  • heffalump111 says:

    Beating the allmighty Stockfish? Whoooooow

  • Logan Mial says:

    Leela learns by playing against herself, so wouldn't the best games for us to learn from be Leela's games against herself?

  • Joel Perlstein says:

    Yay! More Leela games!!! πŸ’š and a great one!

  • U.V. S. says:

    I once beat Stockfish Dev with white after a book opening. The opening moves were 1. e4 e5 2. Bc4 Nc6 3. Qh5 Nf6, and we played from here. Stockfish Dev was on 8 cores only but I believe I would have beaten the 43 core version, too!

  • methylbenzodiazepine says:

    That was a peculiar loss for our beloved fish. Look at black's super strong central pawn island. King safety first. Thanks for the video, cheers.

  • panagranit pana says:

    What is stockfish dev??
    Try to set up ur videos
    Some times ur Audience may not know what you’re talking about . please you have a great audience all over the world but not all of us are computer geeks or have time to follow that. We just like the content that you have but a couple of times you have not explained the various computers that are involved here and taking it for granted that we know what you’re talking about.

  • Bobby says:

    Yes – it's the end of an era for normal chess engines.
    Leela will only get smarter with more time.

  • Talon Higgins says:

    Is this the strongest version of stockfish? And was it played on even terms? I want a engine to be consistently better than stockfish so i can play it and get stronger. But i dont want this to be unbalanced match with accommodations for one engine overshadowing another.

  • εˆ˜ι£žε„Ώε₯½ says:

    shes a rock star

  • Lawrence La Chapelle says:

    What,s time control?

  • Christopher Polzin says:

    How did Leela fare overall in the competition?

  • Matt Mackay says:

    Make way for the queen of chess!

  • Jacopo Pianigiani says:

    Ads every 5 mins?????

  • Fabrice Caron says:

    Leela 11149? Thank you !

  • Jose Herrera says:

    Masterful, beginning of the end of the era, indeed. Thanks for the analysis, great as usual!

  • modolief says:

    Wot?????????? Amazing. Top flight performance by Leela.

  • modolief says:

    I'd like to see the official world computer chess championship — https://en.wikipedia.org/wiki/World_Computer_Chess_Championship — use a Watt allowance instead of any specific hardware constraints. For instance: Your equipment is allowed to draw no more than ten thousand Watts per hour. Spend as much as you like for ASICs or CPUs or TPUs or GPUs or whatever, but the time control and the Watt allowance is the only hard limit. That would show us what the ideal chess software/hardware setup would be.

  • futurefox128 says:

    I think you forgot to mention which net id this is this time. Luckily it is stated in the game link. πŸ™‚ Wow, this is only 11089 (why didn'
    t they use 11250 for this?). Well… it's actually scary to think that there is a stronger Leela than the one here… which even has TB support!

    I thought after CCCC there will be a long wait until we see high-end hardware Leela games again. Very glad to hear that there is bonus games on TCEC AND a long bonus match against Komodo on CCCC for 3rd place coming up. <3

  • Pablo Sartor says:

    Amazing Leela game!!
    We are witnessing the dawn of a new era.
    Thanks KC!

  • Th3Wick3dOn3 says:

    Well, Classical engine love to play on the edge, No one in their right mind would open the G-File and give the dark square bishop, but stockfish gets away with it most of the time not on leela. Punishment on the Dark squares here.

  • Spiffo says:

    I sorta doubt we're nearing an "end of era" for classical engines, but this game does illustrate that Stockfish isn't perfect (yet).

  • KhakiGirl says:

    i think traditional engines can match or even beat the strength of self learning engines provided that it has stronger processing and cores.
    The reason self learning engines are getting stronger faster is due to it "remembering each positions" and choosing the best out of it. So its no surprise that after 1000 games against the best, self learning engines would become better.
    In the meantime, traditional engines are still formidable, but since it relies on the hardware specification, and the code to generate the best moves, the only thing keeping it from getting stronger quicker is the electronics technology and how efficient the code was composed. Id still believe traditional engines can reach the same level(or even solve chess) as self learning engines but it would take a longer time.
    As of now, I rate Leela a little stronger than stockfish but only in the middlegame. Endgames are the most complex and complicated part of chess, and as of now, the strongest traditional engines has the upper hand against Leela in the endgame. Although, if a decent advantage is gained from the middlegame, usually the victor of the endgame will already be decided.

  • bhgtree says:

    I can't wait for the TCEC Cup and CCC2. the LC0 devs have put a lot of work into Leela in the last few months and games like this show it like-wise the Many SF devs have put years of work in getting it to the top. I as a user and fan of both would like to say thanks to both teams of developers.

  • Chevrium says:

    just a small suggestion, but could you show some sort of indicator in the video of who is playing as which colour. i just realized having closed this video and came back to it later that it can be hard to tell who is who. thanks

  • hq cart says:

    back in 2018, Do you believe that StockFish was the best engine???

  • Casualchess says:

    faster mate at 6:42 bg7#

  • observeUS observeUS says:

    Small request – can the full screen be used for the chess board and not the speaker covering 40% of it? thanks We can hear the commentary which is all that is needed.

  • innertubez says:

    Thanks! Your reviews of these Leela-Stockfish games are very entertaining and detailed! I don't know if the end of Stockfish-type engines is nearing, but it will be fascinating to see these two types of engines play more matches vs each other.

  • El Supremo says:

    Be2! Deep!

  • Shantorey Wilkins says:

    πŸ“šπŸ“šπŸ“šπŸ…°πŸ…°πŸ“šπŸ“šπŸ“š
    πŸ“šπŸ“šπŸ…°πŸ…°πŸ…°πŸ…°πŸ“šπŸ“š
    πŸ“šπŸ…°πŸ…°πŸ“šπŸ“šπŸ…°πŸ…°πŸ“š
    πŸ“šπŸ…°πŸ…°πŸ…°πŸ…°πŸ…°πŸ…°πŸ“š
    πŸ“šπŸ…°πŸ…°πŸ…°πŸ…°πŸ…°πŸ…°πŸ“š
    πŸ“šπŸ…°πŸ…°πŸ“šπŸ“šπŸ…°πŸ…°πŸ“š
    πŸ“šπŸ…°πŸ…°πŸ“šπŸ“šπŸ…°πŸ…°πŸ“š

  • Michael Lubin says:

    Nice video, but go a little slower next time!

  • JBishop29 says:

    Very nice channel.

  • String_Dogg says:

    I'd love to see a video regaurding the origin and history of chess. Prehaps it would make a good video series? It's inventors clearly didn't understand the beautiful complexity of the game. I doubt they imagined it would still be puzzled over for years to come. It's crazy that even our strongest computers have yet to master the game. I find chess comparable to the bible – in that it's original creators clearly didn't understand what they had invented, or to what future it would create. The trible stories of the bible have a hidden value or lesson to be learned, something that wasn't consciously intended to have such complex meaning when originally wrote – but it still puzzles philosophers today. Even the greatest music, such as Bob Dylan's "All along the watchtower" is one of the most relatable pieces of art, but yet it's true meaning can never be explained in words. This is why chess is art. Because it connects with us on a deep unconscious-level. A level in which we do not understand. But yet we try our best to. I find it simply astonishing that it was created over 1500 years ago. Timeless, ever-logical, beatiful and complex – chess.

  • Montana Baker says:

    Can someone explain to me why a computer would be set up to have a prime number of cores? I always thought symmetry was important.

  • VITA kyo says:

    We don't hear about AlphaZero anymore …

  • ViewingIsland99 says:

    No, Stockfish can win obsolutly easy, Montecarlo system is perfection, if we put a 72 cores CPU there is nothing to do against Stockfish, AlphaZero was a fraud.

  • elevengiant says:

    what cpu has 43 cores??

  • Senad HamziΔ‡ says:

    Why do you always record leela wins against stockfish, we all see a lot of games when leela lost against stockfish, but you never record it, we see in TCEC and CCCC clear domantion of stockfish crushing latest leela, why dont you record that, be ralistic for once in your videos, you guys repeat leela games until it wins, but mostly loses from stockfish

Leave a Reply

Your email address will not be published. Required fields are marked *