End of Chess AB Era? | Leela shows remarkable pawn structure insights | Stunningly sacs three pawns

♚COURSES kingscrusher.tv/chesscourses
📚 kingscrusher.tv/chessopenings
📚 kingscrusher.tv/chesstactics || kingscrusher.tv/londonsystem
#KCLeelaChess
♚ Play turn style chess at bit.ly/chessworld
♚ Play Chess vs. Kingscrusher and others: www.chessworld.net/chessclubs/asplogin.asp?from=1053
♚ Subscribe to best KZbin Chess Video Channel : bit.ly/kingscrusher
🎥 Leela Chess Playlist: bit.ly/leelachess
FIDE CM Kingscrusher goes over amazing games of Chess every day, with a focus recently on games of Neural Networks which are opening up new concepts for how chess could be played more effectively.
👕 Kingscrusher Store: teespring.com/stores/kingscrusher
♚Donating via Paypal: goo.gl/7HJcDq
♚Channel exclusive perks: kzbin.info/door/DUDDmslypVXYoUsZafHSUQjoin
Game quality tags: amazing, awesome, astonishing, brilliant, classic, crushing, dynamic, elegant, exceptional, excellent, exciting, fabulous, famous, fantastic, finest, flashy, greatest, important, impressive, incredible, instructive, interesting, magnificent, marvellous.
Info about Leela Zero:
en.wikipedia.org/wiki/Leela_Zero
...
Leela Chess Zero (lc0) is a free, open-source, and neural network-based chess engine and distributed computing project.
Leela Zero's algorithm is based on DeepMind's 2017 paper about AlphaGo Zero.[3][6] Unlike the original Leela, which has a lot of human knowledge and heuristics programmed into it, Leela Zero only knows the basic rules and nothing more.[7]
Leela Zero is trained by a distributed effort, which is coordinated at the Leela Zero website. Members of the community provide computing resources by running the client, which generates self-play games and submits them to the server. The self-play games are used to train newer networks. Generally, over 500 clients have connected to the server to contribute resources.[7] The community has provided high quality code contributions as well.[7]
Leela Zero finished third at the BerryGenomics Cup World AI Go Tournament in Fuzhou, Fujian, China on 28 April 2018.[8]
Info about Alphazero:
en.wikipedia.org/wiki/AlphaZero
AlphaZero is a computer program developed by the Alphabet-owned AI research company DeepMind, which uses an approach similar to AlphaGo Zero's to master not just Go, but also chess and shogi. On December 5, 2017 the DeepMind team released a preprint introducing AlphaZero, which, within 24 hours, achieved a superhuman level of play in these three games by defeating world-champion programs, Stockfish, elmo, and the 3-day version of AlphaGo Zero, in each case making use of custom tensor processing units (TPUs) that the Google programs were optimized to make use of.[1] AlphaZero was trained solely via "self-play" using 5,000 first-generation TPUs to generate the games and 64 second-generation TPUs to train the neural networks, all in parallel, with no access to opening books or endgame tables. After just four hours of training, DeepMind estimated AlphaZero was playing at a higher Elo rating than Stockfish; after 9 hours of training, the algorithm decisively defeated Stockfish 8 in a time-controlled 100-game tournament (28 wins, 0 losses, and 72 draws).[1][2][3] The trained algorithm played on a single machine with four TPUs.
...
Relation to AlphaGo Zero
Further information: AlphaGo Zero
AlphaZero (AZ) is a more generalized variant of the AlphaGo Zero (AGZ) algorithm, and is able to play shogi and chess as well as Go. Differences between AZ and AGZ include:[1]
AZ has hard-coded rules for setting search hyperparameters.
The neural network is now updated continually.
Go (unlike Chess) is symmetric under certain reflections and rotations; AlphaGo Zero was programmed to take advantage of these symmetries. AlphaZero is not.
Chess can end in a draw unlike Go; therefore AlphaZero can take into account the possibility of a drawn game.
AlphaZero vs. Stockfish and elmo
...
►If you like my chess videos please consider become a full member here at my Chessworld Site. www.chessworld.net/chessclubs/asplogin.asp?from=1053
#KCChess game tactics
♞ Challenge KC and others for turn style chess at bit.ly/chessworld
♚TEAM kingscrusher.tv/team
♚TWITTER kingscrusher.tv/twitter
♚DISCORD kingscrusher.tv/discord
♚TWITCH twitch.tv/kingscrusher
♚REDDIT kingscrusher.tv/reddit
♚QUORA kingscrusher.tv/quora
♚DONATE paypal.me/kingscrusher
♚COURSES kingscrusher.tv/chesscourses
♚COURSES kingscrusher.tv/chesscourses

[@kingscrusher]

Q. Can Neural network Leela Chess imprison Stockfish pieces enmasse? 🤔😀😎🌍 Video: goo.gl/eM8yBQ Replayable game: goo.gl/VUSz2L #chess #chessgame #ArtificialIntelligence #AI #MachineLearning #DeepLearning

[@florianmangelberger1377]

Hey! will you do game 66? it was basically won by leela and she just blundered the win in the endgame. I think it would still be very productive to analyse the game. :)

[@florianmangelberger1377]

I mean 65. sorry

[@jimglass1290]

The AB era just got a reprieve in game 65 when Leela blundered away an ending that Stockfish and Bluefish both evaluated as 100% (not 98% or 99%) winning for her. One of the reasons people are fascinated by Leela is that she can play both so wonderfully and so badly - just like real people do.

[@u.v.s.5583]

How to pass the Turing test? Just play a dreadful endgame.

[@hoanghiepvu5078]

Even SF without a 6 man endgame cannot convert this endgame. Not sure how Leela with its MCTS algorithm can convert (with 6 man endgame) in this difficult situation. I think it is not a "blunder"

[@u.v.s.5583]

@@hoanghiepvu5078 Then the blunder might have been the move that led to that last pawn exchange. Leela underestimated the depth of that endgame. The position after the last rook exchange was absolutely won, and got really messy after that pawn exchange.

[@hoanghiepvu5078]

@@u.v.s.5583 yes, if she overestimates her ending skill then it is a thing to improve :)

[@Askhat08]

@@hoanghiepvu5078 I ran SF 10 vs itself and white sf easily won.

[@strangerz2]

The greatest game by far in tcec! Alpha zero style!

[@kingscrusher]

Great comment, K

[@herzwatithink9289]

Machine so far ahead of man, man so far ahead of machine. No one I've seen comes close to highlighting the dichotomy better than kingscrusher.

[@yasodhargouthambandaru6414]

NO, it is Leela zero's unique style..........

[@nicholasperkins4655]

I don't think I am jumping the gun by saying that what the Leela project has done over the pass year is absolutely incredible! Like I was saying earlier the history of all chess knowledge was surpassed by Leela in a little more than a year. Not saying that alpha0 couldn't have done that earlier, but deepmind wanted to pursue games of incomplete information like starcraft II. Many were skeptical of A0 including myself, but these results by Leela have made me a believer even though A0 for some reason practically ignored the Sicilian defense in self play. We are witnessing history and thank you kingscrusher for being ahead of the curve concerning this trend.

[@Hellstaff01]

The leela team should be rewarded with a thorn pawn t-shirt!

[@4merxtian432]

We love your work KC! Leela views the game holistically. SF lives in the moment while Leela views moves as a pattern for finding comfortable endgames.

[@kingscrusher]

Cheers, K

[@franklippert4278]

With Leela we are already accustomed to superlatives by now. But this game _is_ special, no doubt. There were moves who just didn't appear on SF's radar no matter how long it looked at it. _Everyone_ (human or engine) would have played 20.Bxc4 winning back 2 pawns, but Leela sees deeper.

[@wolffang21burgers]

Cover game 65...... Actually I'm not sure I can watch that again...

[@u.v.s.5583]

Epic fail! You are doing it wrong! Cavewoman fails to catch a fish and almost drawns in a pool!

[@ahuruglica]

You know it took Leela less than one year to reach the strength of Stockfish. And stockfish has been here for a decade. Stop being buthurt because this is the last match when they compete, next super-final if Stockfish is lucky to even reach, will get crushed by Leela.

[@kingscrusher]

Epic fail by Leela indeed - not sure I can cover it !

[@woffordwolf2071]

@@kingscrusher yeah that 65 game was the craziest thing I've seen in chess in quite a while....or maybe ever.

[@woffordwolf2071]

@Peter sturm Ok, but where can I find the game, I didn't find it in the archives?!

[@joseraulcapablanca8564]

A fantastic win and a brilliant concept imprisoning the opponents pieces. Thanks KC keep up the good work

[@hazelelloyd]

Thanks KC, this was amazing, I still get chills when watching these. The pawns I believe are an under appreciated piece whose strength actually gains depending on position vs tactic. And Leela clearly has tactics rather than depth searches, that is how NNs are supposed to work - they "think". Could you post the name of the book you recommended on pawn structure, I'd like to check it out, I keep missing the name, my hearing aid is not the best you can have :-) Cheers

[@naimulhaq9626]

Brilliant, educational too.

[@vargas2022]

This looks like an inferior variation of French Defense. However, your discovery that Stockfish doesn't eventually find this bishop "desperado" move that Leela used to imprison the black rook & king is truly remarkable.

[@jamesecarson5631]

King: just the best presentation. Congratulations.

[@Askhat08]

Really epic game. Poor rook never moved in the entire game ;)

[@kingscrusher]

Cheers, K

[@seasideman]

I find it fascinating how the neural net approach to chess has led to highly positional play. In this game, the classical ideas of space and piece activity positionally triumph over the materialism of Stockfish. Great game, thanks KC.

[@pablosartor6715]

As usual, this is a very good explanation of Leela's style of play Thanks KC!!

[@rachmattubagus6483]

i tought leela trolling with knight promoting

[@Durian_Defense]

Leela isn't "three pawns down". You forgot that a Fawn Pawn on g6 imprisoning the Black King and the Rook is worth at least ... what.... three pawns PLUS a win!

[@kingscrusher]

Great comment, K

[@paweljankiewicz1383]

Excellent analysis KC. As someone said gradient descent writes better code than programmers.

[@scottburger6224]

Hi KC! Do you have any book or KZbin playlist suggestions on the history of chess AI? Thanks!

[@paulbloemen7256]

Maybe the most impressive game I have ever seen: Bg6!! How relentless Leela played this out till the end. But, isn't the Mac Cutcheon variation really suspect? Like, when Kf8 seems to be the best move, with so many problems to solve, one must be a bit masochistic to want this all. Maybe c5 was better than b6, or? If the French opening at all, which variation gives Black some decent prospects, does Stockfish, or Leela, have a clue here?

[@kingscrusher]

Great comment, K

[@alienrenders]

I let stockfish run to ply 45 and it always wants to play Bxc4 as well. Stockfish seems to underestimate fawn pawns. Even after the exchanges that installs the fawn pawn on g6, Stockfish still thinks white is only slightly better (basically a draw) while Leela thinks it is winning. The line that Leela followed after Kg8, Stockfish considers it a draw when considering Kg8, c3 Qf8, etc. Another strange thing is that perhaps Stockfish has flaws that other AB engines might not. Stockfish is not stable. I'm not talking about crash/stability here, but the official definition of "stable" which means that for a particular set of inputs, the program will always give the same answer (at any given ply). Google "stable sort" for an example of stable algorithms. Stockfish will give different results on different runs for the same ply. As such, Stockfish does not use a stable algorithm. It is very odd. Perhaps some moves are relatively even and it has to pick one at random? I don't know. The first time I ran Stockfish, it absolutely would NOT play Qf8 even past ply 40. I'm running it now again and it's the only move it wants to play right from ply 1. It's at ply 45 and still won't reconsider. Truly amazing. This is not the first time I've seen this happen. This may be due to the low drawish evals it is giving these moves and that threading is causing a race condition to which move it considers first. Rb8 by stockfish is also interesting. It's a move it only considers after ply 40 on my machine. It also never sees the followup g4. It only wants to do exchanges for white to gain a pawn. Maybe your earlier comments about Stockfish being too materialistic are accurate. Leela had many opportunities to get those pawns back, but played alternative lines that Stockfish tends to ignore. Said in a plainer manner, these alternative lines would not be seen without Leela or NN. AB engines would never play them.

[@alienrenders]

Funny how in game 66, there was a disconnection (ISP issues). So they restarted the match. Then they restarted the match yet again (not sure why). In all three games, Stockfish played three different moves on its second move out of book. So a drawish game turned into a win for Stockfish. Some people in chat were now questioning why Stockfish was giving different answers. The excuses just don't fly. I'm a programmer and a stable algorithm is stable even if you introduce threading. There is something very odd going on with Stockfish.

[@abcdefghij9777]

Stockfish 10, 20181206: Hash = 16 GB, Overhead = 30, PV = 10, even number of plies > 30: {[32] 0.67:} 20. Bg6 fxg6 21. hxg6 Kg8 {etc.} {[32] 0.65:} 20. Bxc4 dxc4 21. O-O Kg8 22. Qxc4 Rc8 {etc.} {[32] 0.31:} 20. O-O cxd3 21. Qxd3 {etc.} {[32] 0.17:} 20. Kd2 Kg8 {etc.} {[32] 0.00:} 20. Be4 dxe4 21. Qxe4 Rc8 {etc.} ... Estimated time: less than 60 sec with the hardware used by TCEC

[@alienrenders]

@@abcdefghij9777 What are you trying to show here? That stockfish plays different moves every time you run it?

[@abcdefghij9777]

@@alienrenders By using 10 PV (principal variants) you will get a ranked list of 10 moves fully analyzed to a common ply depth, which should be an even number (in this case 32). Thus Stockfish is able to effectively broaden the search tree and produce a better ranking of the moves on that list. Doing so needs a sufficiently large hash (16 GB) and more time. With the TCEC hardware, however, this shouldn't be any problem. The last 5 entries on the move list above have been omitted (...).

[@Jaydelrey]

@@alienrendersI've been using stockfish for analysis since version 4 and to be honest this "stable" description that you expect from stockfish doesn't sound very desirable to me. (I personally like that stockfish can find different ideas/moves at the same depth). And to be honest as someone who has looked at stockfish's evaluation.cpp many times it has quite a complex evaluation that takes many things into account. Not surprising if it has trouble deciding between multiple moves that it thinks are all good in a given position. So an example if u were to analyze a certain position to depth 30 it may like 2 or 3 (or more moves) and score them similarly. So if u then clear the hash tables and run the same position again it might choose a different move at depth 30 (or whatever depth you set etc... In a situation like that it's best to just analyze deeper until it is more certain)

[@jeremykothe2847]

WOOHOO!

[@ahilanpalarajah3159]

Why e6 sacrificing the pawn first, rather than Rh5 right away? Bishop still comes to d4 regardless. What critical tempo gain does e6 provide?

[@alexmagor7538]

The funny thing is Leela probably would have eventually played Bishop G6 even if it wasn’t forced to.

[@Attlanttizz]

That was just nasty, like a boa constrictor squeezing ever harder...

[@pukulu]

In all of the games that LC0 is winning vs. stockfish, it has the white pieces. I don't know why.

[@noidelaplatja]

AlphaZero won SF8 only 3 games from 50 with black.

[@beattoedtli1040]

What could be the reason behind underpromoting a rook?

[@paulhowe9802]

What is the foam pawn??

[@dannygjk]

thorn pawn

[@ivanfig1]

Epic

[@Hawaiianislandweddings]

Did you say Bishop at 9:56? Don't you mean rook?

[@romankusnir45958]

I think, rook a4 even more precise than bishop h4 ;-)

[@enthusiastofopp7013]

What happened to your blitz videos? We miss them!

[@yasodhargouthambandaru6414]

Dear Leela,I love you soooooooooooooooo much................

[@bhgtree]

Thanks KC. It seems that Leela will finally have some competition on the trolling, at CCC6 new engine Allie (LC0 NN and custom engine github.com/manyoso/allie) is playing.

[@FirstLast-le6io]

Bishop G6 just screams to be played, im a relatively weak player but that move just seems so obvious I'm surprised stockfish doesn't see it at some depth...

[@kingscrusher]

Cheers, K

[@benjaminhartmann4522]

I wondet what kind of games we would see if the goal for leela would be a draw.

[@joelperlstein1255]

Kingscrusher>>>agadmator Leela=Alphazero>>>Stockfish

[@wolffang21burgers]

Watch game 65 though... Some refinement is still needed :(

[@julioandresgomez3201]

@@wolffang21burgers Indeed. At some point not Leela but Stockfish ev was at +152(!!!), and it was a draw! The kind of positions that any strong human wins with 30 seconds on the clock, but a NN without experience in similar positions is clueless and aimless.

[@wolffang21burgers]

@@julioandresgomez3201 Leela figured out the right idea in the end... But too late due to the 50 move draw rule, if she hadn't of wasted moves earlier on it would have been fine.

[@pkundrat]

@KC - did you maybe consider syncing with other youtubers to chose who covers which game? Duplicating efforts on the same game seems like a waste of time and energy, while we dont get coverage of other interesting matches. This way we could see much more games with human analysis.

[@JamesBrown-wy7xs]

♜♞♝♛♚♝♞♜ ♟♟♟♟♟♟♟♟ ♙♙♙♙♙♙♙♙ ♖♘♗♕♔♗♘♖

[@extollo]

good game, but the epic missed TB win game that follows means not quite end of era yet. lol.

[@u.v.s.5583]

End of an era: 50 move rule ends the era of Leela trying to mate with two pawns and a knight against a pawn and bishop.

[@mrbdzz]

U.V. S. 🤦‍♂️🤦‍♀️

[@ahuruglica]

It hurts when your beloved engine loses, lol. Internet is full of crybabies.

[@fredharte8677]

Leela was unable to find the winning line that had already been worked out in a 7 piece TB whereas SF apparently could. Every engine has its own characteristics. Were there to be a 10 piece TB produced it would be beyond the calculating powers of any engine to see winning lines contained in it. And no criticism could be levelled anywhere unlike that which Leela had to suffer in the TCEC chat. It is enough that she is two points clear as he sufi enters the later stages - despite her deficiencies in the strict calculating battleground of the endgame.

[@extollo]

@@fredharte8677 I understand the diff in engine approach and why leela couldn't see it, but this was a particularly entertaining blind spot. esp. because the game was so beautifully played before that. the evals just kept her from bringing the king around... future lc0 nets & search will be better, but games like this will always be possible with the nn/mcts approach.