Leela Chess Zero ID 464 vs Houdini 6

[@alphanaut8353]

Whenever Jerry asks "What move does Leela play?" I'm having good success by assuming it's the opposite of the move I would play...

[@kasurusak5168]

alphanaut hahahahahaha

[@noobmaster31]

When he asked that, my twisted first thought was exactly what leela played. But my rational side immediately dismissed that first thought as nothing more than my own inner failure. How could I think of a move that goes against the norm. Leela is beyond human understanding.

[@HipHopAn0n]

Do exactly this during games and you'll win every time 😉

[@billyb6001]

its always a pawn moving one move.

[@Jartran72]

I mean there are way too many moves possible to just go with an opposite. There are mostly dozens of choices, no opposites exist out of maybe choosing between 2 pieces to recapture.

[@Cscuile]

Very weird. Houdini 6 is known to have time management issues in certain positions so perhaps Rd4 was played out of time pressure on Houdini's behalf. Stockfish Dev sees the position after Rd4 as +1.84 (While before at +0.54) Depth 44 1.7 Billion nodes.

[@Adam-hj1hc]

Just started watching the video, and just wanted to immediately say thank you for all of the content you are currently putting out. Love your vids and analysis. Cheers mate.

[@ChessNetwork]

Thank you Adam.

[@davvves7977]

Top quality content, Jerry. To the point, calm, insightful, clean UI, just smooth as butter my friend. Keep it up!

[@ChessNetwork]

Thank you

[@lawdogwales5921]

@@ChessNetwork You're welcome.

[@bhgtree]

I've been playing computer chess since the Commodore 64 days (30+ years) and nothing can compare to the amazing way Leela (and Alpha Zero) plays.

[@lawdogwales5921]

So true. Like a superhuman Tal and Fischer combined with a supercomputer.

[@aimeduquet4879]

"fianchettoed knight riddle" sounds like a good movie title

[@musicianniaz]

Also, a great puzzle for Wheel Of Fortunes. Category: Black & White

[@ginoginoh]

Those positional exchange sacrifice always impress me so much.

[@michelemeriggi]

I think it was a mouse slip by houdini imo (;

[@steinglastra8545]

Michele Meriggi 😂😂lol

[@joemiller947]

Never thought I'd hear "GTX 1060" on a chess video

[@russellfroggatt]

These games continue to be extremely instructive with Jerry's logical way of analysing

[@r0bbi3]

The progress may seem convincing, but it's important that you look at the "Leela Ratio" as described on the LC0 blog; similarly to what Jerry explained at the beginning, it tells us of the hardware advantage that Leela was put at in a lot of her gauntlet games that have recently been played.

[@r0bbi3]

Nevertheless, Leela's progress is great to see, but these games shouldn't be taken so literally in terms of her strength

[@4merxtian432]

Leela ration was about 0.625

[@dtracers]

I agree it is really interesting to see how the chess communities and go communities are treating this differently. In go everyone is put the engines on the stronger hardware available (like multiple Titans) just so they can learn from the go games. Where in chess everyone is just trying to use the weakest hardware to play the games

[@Luredreier]

+dtracers Agreed. And since Leela scales better with hardware then traditional engines like Stockfish does...

[@RantyCat]

Thank you, Leela , for establishing my faith in Morphy. I always thought his decisions were the best ones .

[@Cscuile]

Thank you for the amazing Leela videos lately.

[@okiejoetex]

I am really looking forward to a Leela / Alpha Zero match up. That would be so great. Thanks for bringing these gems to us, Jerry. :)

[@mohd1nagyyyyy]

You really have a CRUSH on leela 😂😂

[@IronMaidenEE]

Try uploading a game of leela playing black..

[@robobrain10000]

or a game where leela loses.

[@ori5021]

even alphazero won only 3 games with black in the 100 games match against stockfish.

[@Pintkonan]

whilst stockfish managed to win 0 with black ( and white ), getting rekt 28 times :>

[@lWaterFlowl]

Ori shem-ur it just happens that in the 10 games released to public, just 3 of them are from A0 playing black, so you can't tell for sure that A0 won only 3 games.

[@asmo_1929]

We know for sure that alpha 0 won 3 of the games with black, and drawn the rest, why would they lie to us?

[@nimbleninja12]

really liking these leela chess series. none of those common openings, definitely learning a lot

[@HipHopAn0n]

These Leela games are the most interesting on KZbin. Amazing analysis and strategically this game is as clean as it gets!

[@jordanstack5982]

love your videos. i love watching leela's games, they are very instructful. keep covering her games!

[@suntzu3541]

Great video, love the commentary as usual

[@burt591]

15:39 I tested that position on my computer with Stockfish 8 and it realizes instantly that Rd4 is a bad move. It suggest instead moves like Rf7, Rf8, Qc6 or Qc7 and a few other moves (basically the same idea that you suggested). I think it was simply a blunder by Houdini, maybe the hardware it is running on is too weak or it was some kind of bug

[@AAntichrisTT]

Unless it was running on more powerful hardware than you're using and can see further down the string hence realising that Rd4 gave it a better fight than the alternatives Stockfish 8 offered? I don't know myself, just suggesting.

[@ailst]

I also analyzed that move and came to the same conclusion. There is no justification for that move. The position went from drawn to losing. It almost looks staged... Like someone snuck it in to make Leela win. If Houdini really saw it as the best move, it should not have lost afterwards. And as far as finding a winning plan for white goes if black doesn't do that blunder... If Stockfish doesn't find it, then clearly I won't find it either. I don't think there is any. If I let Stockfish Duke it out it ends in a draw.

[@vnarl]

I agree that Rd4 immediately gives up the game, perhaps Houdini was trying to evade a draw or some kind of glitch. after a correct Rf7 or Rf8 defense, White has no easy win, even after sacrificing the knight on f6, Black has enough resources to defend; and if you give back the exchange, Black could do it at the moment when the knight is on d5, plus take the pawn for exit to equal ending

[@mateipopescu4338]

"the black king is homeless!"

[@eyeborg3148]

Matei Popescu WE WUZ KINGS!!!

[@ryanhorenci6275]

Wakanda forever

[@maestro56777]

Even chess can't escape from this epic humor

[@mateipopescu4338]

Art of Letting go loool

[@robertbeuck7556]

He should be able to stay at he babymomma crib

[@ImaginaryHuman072889]

I always thought it was strange that most computers resign AFTER they make a move instead of before haha

[@lispcodinglive]

My guess? The algorithm has them get in the best move considering time constraints, a separate algorithmic decision. Thus limited, they do not get to "OK, I'm dead." But once they move they continue thinking on the opponent's clock and get to that conclusion. A human player would wait for a possible opponent blunder...do engines take a higher moral stance? Hmm...

[@ImaginaryHuman072889]

oh I agree with you, that the computer just didn't get a chance to fully calculate the outcome until after the move, just hilarious how they first move and then they're like "oh shit"

[@ariel_haymarket]

Man I wish I had that nice house.

[@Danicker]

Me too... I'm homeless ;)

[@mrmarkstv6585]

Wait what lella is soon gonna play stockfish 8 how is she learning that fast

[@38FerreroX]

She's estimated to reach sotckfish 8's elo at the end of november 2018

[@RodelIturalde]

I would like to see Lella get crushed by Stockfish! Long live the brute-forcing engine!!!

[@RodelIturalde]

Even though stockfish isn't really brute-forcing, it is more brute-forcing then lella.

[@ima7mad]

She got her father Jerry doing all the heavy lifting behind the scenes.

[@fahimp3]

Sadly they r planning on restarting with a new net, it will get worse before it gets better. :(

[@Rikku_]

Hey @chessnetwork, what is the source of the AI picture you use for Leela? And is there any Magnus Carlsen games coming up for review? Thanks for making these btw!

[@Twas-RightHere]

Leela vs AlphaZero would be very interesting to watch.

[@mikechambers9129]

A Rocket - LC0 v A0 would be interesting in 2 or 3 years. Today it would be a bloodbath unless A0 was running on a very wimpy server.

[@ocudagledam]

Isn't Leela A0 without the A0's experience? If so, on the same hardware, Leela would get crushed at first, but she would soon start to catch up.

[@Twas-RightHere]

Well according to Jerry's first video on AlphaZero it learned to play in only a few hours. Leela has been training for months, I see no reason why we should expect AlphaZero to crush Leela.

[@mikechambers9129]

A Rocket Leela is an open source, volunteer effort. Alpha Zero was developed by DeepMind, a commercial AI Research company purchased by Google. There's a difference. Right now Leela is around 140 ranked on the CCRL. It is nowhere remotely close to Stockfish 7 strength which Alpha Zero demolished.

[@jamieoglethorpe]

AlphaZero had no doubt played a lot more games beforehand that we don't know about. We did not witness any of the earlier efforts needed for debugging. We saw a demonstration of DeepMind's prowess on hardware that the rest of us cannot imagine, built by a dedicated team of some of the best computer scientists on the planet. Leela Chess Zero was built by a team of volunteers using computer power donated by a community. We are seeing all of this in public. We don't know how many training runs DM threw away before they were happy with the result. DM had it much easier than the LC0 team, because they only had to support one family of processors. Leela has to run well on AMD and Intel GPUs, as well as the whole family of Nvidia cards that people may have on their computers. It must also run as well as possible without a GPU. The Leela neural networks (NNs) get downloaded to volunteers' computers several times a day. This forces the devs to consider ways of making the NN as compact as possible, an issue DM did not need to consider. Running on much weaker hardware forces the devs to make the code more efficient. There is pressure to make the package "download, install and run" by non-experts. DM certainly did not need to worry about that. The LC0 project has now gotten to the point where it can run the equivalent of the several hours run by DM. When the training run is restarted in a few weeks, then a comparison can be made. The comparison will have to be on the number of games played against itself, not how long it takes. We are talking of a few months to play the 44 million games DM played. The LC0 project also faces more stringent requirements. It will not be able to do so in conditions under its own control. It will have to do so in public competition in conditions that must be seen to be fair by both sides. Stockfish , Komodo and the other top engines will certainly have access to opening and endgame books which DM did not give to Stockfish. The hardware will not be a supercomputer. It will have to be something in reach of a dedicated gamer. These days that would be something like this www.pcgamer.com/pc-build-guide-high-end-gaming-pc/: a 6 core Intel Core i7-8700K with a Nvidia GeForce GTX 1080 Ti graphics card, 16GB of RAM. Tournaments should also have a SSD for the end-game tablebases. I believe that such a rig should be considered fair.

[@seanpetermcdonald]

Thanks Jerry, very instructive.

[@camspianosink]

No comment on 2:07 when white castles leaving the knight open to be captured on H3 and break the king's pawn structure? Is castling into that position that normal that it doesn't even warrant comment?

[@ChessNetwork]

My brief comment about the damaged structure came before 2:07. If black gives up the bishop for knight, I imagine white's now unopposed light-squared bishop would help to repair the weakened white camp. Bf3 is a start, and maybe even Rg1 with an eye on the black king.

[@camspianosink]

My mistake. Thanks for the response.

[@bencheevers6693]

Edit: Took me about 5 minutes so I'm probably wrong but I would love some clarification on how I'm a very novice player but I came up with this reason for why the rook moved there, and I didn't think about it too deeply, Houdini only has a few pieces that he can move, and I think if he undoubles rooks then after knight d5 forces the queen to move he'll then drop a pawn and that he also has some reason why he doesn't want to move the king then the only other move is moving the rook and d6 is hit with a fork with the king as well but in that position he isn't defended by the same pieces? I just ran this down a couple lines but if he moves the king then knight D5 hits the queen forcing either a capture with the rook or if you move the queen then the rook can take F6 and be defended by the knight so the king can't move away from that square so it is zugzwang I think. I'm confusing myself but I think I forgot that if the rooks don't undouble the knight can't go to d5 but the pawn advance I mention below works, I think the king has to be in that postion to avoid that tactic and so it's the best of available possible moves. Edit: ran it down at least one more line (this is hard in my head as I'm not used to this) King can't move f7 or else c3 then d4 is played and the f6 pawn is pinned so can't defend e5. I think houdini saw all of this and played the best line that he didn't see forced moves to losing or if they were all losing then the one with the most chance of white blundering.

[@graemeweir1948]

Maybe the Houdini move is warning for us all. The machines are rising and Leela is leader! 😳 another great video Jerry, thanks!

[@ChessNetwork]

Than you Graeme

[@lukas74656]

At 10:04, why go for the bishop instead of the rook? thx in advance

[@SockTaters]

Yes! Of the videos you do, these are my favorite

[@stevenmcculloch5727]

At 15:00 you agree with black to close down the position with g5 but why is this best when black needs open files for the rooks?

[@huaweiandroid125]

I love this game analysis. Thank you!!!

[@sooooooooDark]

that moment when u realize that just 1 moment ago u thought black had the better position (from a human perspective) ..but now it is suddenly white that is in the lead lol

[@luisg.ontoriaalvarez2334]

last video about leela felt so tactic, and this was pure strategy long game. It maybe true what they say about perfection, It has no style! Thanks Jerry p.s. a safe king and two pass pawns against a rook plus a king completely naked... Seems a good reason not to play 51... Rd4

[@kasurusak5168]

willy ontoria wow, perfection has no style!! agreeeeeedd!!!

[@modolief]

Stupendous video, I watched it twice it was so good. A _fascinating_ game -- positional win by Leela Zero -- and butter smooth explanations by Jerry. I thought f3 by Leela Zero, Steinitzian restrion of Houdin's Knight, was an amazing strategic choice by the neural net.

[@eleutbenito3174]

Wow,very nice analysis.

[@kameronbourne84]

Jesus Jerry, chill out. I'm not even gonna have enough time to PLAY chess with all these uploads.

[@kasurusak5168]

hahaha

[@adarshnair2710]

Excellent analysis Jerry... The homework : I think the plan could be to push the pawn on d3 to d4. Let's consider overprotection of f6.. Scenario 1: if rook takes we take back the exchange Scenario 2 : if e takes, knight g5 with tempo and then e5 Scenario 3: if c takes, idk maybe again knight g5 queen d3 some rook manuever white king seems safe and a and b pawns look strong while black pawns eventhough passed can't advance clearly. I maybe totally wrong btw :D

[@sooooooooDark]

leela really values their pawns highly it seems i really wonder if we could ask her how many points she would give to central/non central/side pawns ..i bet central pawns would be valued like 1.7 points to her

[@zCrabOG]

She sacrificed a couple pawns last game though didnt she?

[@sooooooooDark]

maybe she values them as "they r not valuable in themselves" but rather "they r disruptive, especially when activated quickly" i mean look how leela develops her pieces...she doesnt exactly follow common opening principles (getting out pieces as quickly as possible/not moving the same twice etc)

[@curtisw0234]

bad players don't know the principles, good players do, great players know the exceptions

[@zCrabOG]

sooooooooDark but that's also completely the opposite of one of her last games where she sacrificed those pawns in order to develop quickly. I think leela makes exceptions and plays differently in every situation

[@andrew1717xx]

sooooooooDark I don't think Leala plays with her own "rules". That seems like a human concept.

[@AdrenilineMLF]

Maybe the rook move is meant to prevent Kd5 followed by Rxf6? The white queen moved from c3 to c4 just before, and I see no gains for moving the queen to b5 or a6, and its not like the queen can rambo behind enemy lines alone. I think the queen moved up to support the knight jump, which would support the rook breakthrough. Could you find a way to defend against this? I couldn't but I'm pretty new to taking chess seriously.

[@chrismcgoldrick3950]

It would be interesting to do a study of these engines using a fixed software revision (of each) but changing the available computational power to see if there are minimum capabilities to achieve a certain effect or if there are inflection points or discontinuities in performance curves.

[@vineetmoghe976]

Nice idea of flipping the board

[@siralfrednobel]

Can anyone tell me the earliest recorded game played till move 6?

[@piotrkantor7275]

Can You analyze some games where Leela lost after good game or when it's black? It would be really great.

[@Luredreier]

Is this the cuda accelerated client or the regular GPU client?

[@bluekeybo]

Jerry, where were these games played? Link? Thanks!

[@Tinybabyfishy]

Hi Jerry, That final question is a doozy. I want to bounce this off you: if Rf7 then Nd5 forces the Queen to move to c6 or d6 which gives us either Qc6 Nb4 Qb6 (=Qd6) Na6 Rc8 b4 or if Qd6 is played straight away then b4 can be immediately played In any case, the c pawn can't be maintained

[@Tinybabyfishy]

Which is to say, Houdini just got a better number for giving back the exchange and repairing the pawn structure vs. dropping a pawn and having an inferior structure and a tied down rook

[@matteogauthier7750]

In the games I saw played by those neural networks, they usually create a developpement and space advantage early on and brillantly use it to crush their opponent. In this one, it was interesting to see how Leela manoevred with less space and developpement than her opponent!

[@lovelight5337]

What is the basepoint of white pawn structure? Definitely c2. Houdini trying to open up the file against C2 point.. by giving an exchange sacrifice (Rd4) and also improve its isolated pawn (C5) that's the only possibility to attack Leela perhaps.. if Houdini can post It's Rook on C3 it would be a different story. I don't know... :)

[@dylanhunt5368]

51. Qc4 52. Nd5 53. c3 54. b4 Queen protects knight so black can't get back the exchange pawn+knight for a rook. Knight and c3 pawn protects b4. Black sends the rook in to block the knight from advancing to the nice d5-square. The white knight can also double attack the black f6 pawn together with white rook, and black falls apart.

[@mattroxursoul]

Did Houdini see a forced check mate somewhere in the next move? That is usually the only time these engines make moves like that.

[@illusoryveils]

What is this program you always use to analyze these games with?

[@lalalolo5666]

Loving the high upload rates. I also prefer game analysis to gameplay ( i still love to watch gameplay)

[@gabrielduchambon8746]

This position is very very interesting. The best I could come up with is 52.a5!? after Rf8 with the idea of activating the rook via f1-a1. I got into interesting endgames. I looked at ...Qxa5 53.Rf1 Kg6 and after for example 54.Qe6 Qc7 55.Ra1 Rd6 56.Qe8 Qd8 57.Qxd8 Rxd8 58.Ra6 Rb8 59.Nd5, the white pieces are optimally placed and it looks tough to hold for black (ideas of putting the king to c4 and playing Rc6). These kind of positions even with the queens off are a good try I think, with this trickiest of knight worth more than 3 points. I'm just a 1800 with an engine so I'm looking forward to other ideas from stronger players. :)

[@maxzhao2545]

Jerry, my take on black’s exchange sacrifice was that Houdini was gambling that its major pieces would be better placed on the queen side when the position opened up. Although I believe that the position could’ve been held as a draw for black, Houdini relies much more on evaluation depth than an engine like Komodo. Komodo could have possible gone for a draw. Houdini likely decided to try its winning chances against its better judgement. Its evaluation of its own pieces value was probably not entirely accurate.

[@borcudo]

Can you cover a Leela game where she looses?

[@mattroxursoul]

I love hearing about brand new original positions. I find them all the time, yet some how I feel less than satisfied with them )

[@zippymax1]

When you analyze the strategy of these harmless bots, my spidey-sense cries "SENTIENCE!!1!"

[@yogesh193001]

Leela has white privilege

[@TranscendentalTunes]

Winning comment

[@Ouvii]

Yeah, she made the black king homeless even.

[@jasonc0065]

He said, yes we can. And gave up the exchange. It's Bush's fault.

[@HospitallarKnight]

Black Kang homeless

[@taavettiihantola561]

I think in 13:05 Black should play Qb7 or something like that, so If the pawn is captured, Houdini has the H-file for the rooks

[@peterpetrov6522]

Hey Jerry, nice question but you are overthinking this! There is no win for white. Chess engines (including Leela) don't like draws and if their opponent is low on time, they will make a concession and mess up their own position just to continue playing. It appears they love drama even if they end up losing an easily drawn position.

[@iwersonsch5131]

I found Rd4 as well. It seems to inconvenience the white queen, doesn't it?

[@iwersonsch5131]

Wow. Houdini turning off 4 half-move search

[@lolwhatyesme651]

Why didn't the black bishop go on b7? Wat?

[@songsmadeforyou]

why does leela use graphics cards and other engines use cpu?

[@a.gindinson]

Thank you, Jerry

[@SayanMukherjee1995]

Hmm, my amateur guess is that black played Rd4 on move 51 to prevent Nd5. If Black doesn't want to give back the exchange he loses f6 and suddenly the knight, rook and the queen are coordinating around the black king, especially on f6 and f7. Soon e5 and g5 would be under pressure too. Looks scary to me! So I guess on Nd5 Black would have to give back the rook on d5, leading to white having a passed d pawn and a Qe4 Rf5 setup seems to be able to maintain pressure on f6. Not sure if this is the right way to go though, what are your thoughts? Please pardon this noob in case he said something stupid! Thanks for the great analysis; it was super interesting!

[@coot33]

11:33 this is why Qa5 happened to generate the royal fork in this variation. Simply Brilliant.

[@mstfici2685]

Leela keeps getting stronger

[@Sack_Zement]

Thank you! :)

[@INFINITY_99]

Great stuff, I really like the Leela videos! :) Sadly I can't provide any insight..

[@supercool1312]

The fried liver atack is dangerous, i love it

[@benjaminzaugg1127]

love the vid! could you explain a little what the two chess computers are building on? Why is it the GPU / numbers of CPU cores important for either? (I understand that generally they use calc. power but why especially the GPU?)

[@piotrgrzybowski8263]

The point is that Leela is based on Deep Neural Networks which can only be effectively run on powerful GPUs. CPU has not enough computation power to make Leela "thinking" in reasonable time. Neural networks are based on multiplying big matrices which can be nicely parallel on the GPU. On the other hand, the implementation of today chess engines do not allow to run effectively on GPU. Why? They are based on 64bit words which is perfect for 64 bit processors. Second point is that GPU is poor with recursion (tree searching moves). More CPU cores allow to make deeper search for best move in the same time.

[@benjaminzaugg1127]

What a great explanation! Thank you!

[@fleecemaster]

Imagine a tree search on a quantum computer! Man that would be good!

[@Luredreier]

CPUs are good at running complex code with a lot of branches (if A is true then do 1, if B is true do 2 otherwise do 3, that kind of things) and specialize in running the code in those branches very, very quickly with low latency, speculative execution etc. GPUs are good at doing a lot of math, but they're more specialized for the types of math used in graphics workloads, doing them in parallell, provided that the code *can* run in parallell. Neural networks like Leela *can* but makes poor use of the strengths of a CPU.

[@jamieoglethorpe]

Traditionally, chess engines use the Alpha/Beta (A/B) algorithm. With it, you assume that your opponent will make the strongest possible move against you, so you consider its position assuming it thinks the same way as you do. Critical to this algorithm is the cut off. There are just too many to go all the way to the end of the game. When you have gone far enough, you evaluate the position using chess knowledge to decide how advantageous the position is for you. You then go back through the moves, choosing the strongest move, at each point remembering its evaluation. You then make the strongest move. The top engines have very good evaluation functions, and very efficient searches. AlphaZero and LeelaChessZero use a different algorithm, called Monte Carlo Tree Search (MCTS). In it you calculate the chance of winning, by looking at each possible move in a position, and playing a random game. You count up the points (win, draw and loss). If you recognize the position, you can use the previously calculated value. You then choose the move that leads to the most winning possibilities. In addition to MCTS, Z0 and LCzero use a deep neural network to predict the value of each position, then to use MCTS to refine the valuation. The network is trained by playing many games against itself, and valuing each move according to the actual result of the game. As more games are played, its strategic knowledge is improved. In unexpected positions it relies on MCTS which is less efficient than A/B, which is why it is prone to blunders. Back to your question, A/B algorithms are very contingent on the position at every move being considered. A single core is needed to process the sequence, but the different moves can be processed in parallel. So a strong A/B engine benefits from many powerful CPUs. The NN/MTCS algorithm works with matrices and uses matrix operations to evaluate the inputs given the neural network. GPUs are designed to do these calculations, but the individual processors are much less powerful. They make up for this by making the matrix calculations in parallel. Typically LCZero needs two CPUs to keep a GPU going flat out. This is typically equivalent to what a six-core CPU can achieve.

[@truthseeker10990]

To answer your question maybe you can let the computer answer it itself, isn't it easy to setup those engines with that same position and substituting the Rd4 move with another one, and then let the engines take it from there, just a suggestions that I think is possible.

[@RidingBones]

hey Jerry, if Leela played 10 games against Houdini and 10 games against Stockfish (or 100), would it play the same position different after learning its opponents strengths and weaknesses? Not sure this have been tried but Leela seems to play more like a super human than a super engine. Maybe the question makes no sense but I will leave it just here.

[@4merxtian432]

Leela only learns from training games against itself.

[@加州猫主席]

15:41 Thanks so much, got bored of summer so I can work on this now

[@justinc0336]

I LOVE this game! I got both ‘quizzes’ right. To answer the question of why the rook sacrifice. There always seems to be some relation between material and initiative. I believe black wanted to give up a weak room for a more powerful knight. This provides better king safety and also creates a weakness on some dark squares around white’s king. Perhaps if the black’s queen could infiltrate past white’s pawn structure, that could be a long term asset. What would I do if the exchange didn’t happen? First off, the board is too locked up to convert, so white must look for some sort of break. I think playing Ne5 followed by c3 is essential. C3 is preparing b4. The rook could reposition to support the b file and in this case, a5 doesn’t work. The knight already controls b4. The queen and rook can further support the b4 square if necessary. For this to be successful, white needs that knight to stay on e5 though. It helps shield the d pawn when c3 is played.

[@vnarl]

a5 is generally a bad move, instead my computer suggests a6

[@felipejhony6039]

Leela runs on the graphics card instead of the cpu?

[@ChessNetwork]

Yes

[@maxwelljann5462]

Homework: RookD4 is the first move I saw. I didn't see the knight fork. I was putting myself in the shoes of black and thinking . I need to liberate these rooks, set up alekhine's gun and get a tempo on the queen at the same time, but given the knight fork and the losing endgame, why not move the king somewhere else and then go in for it? It's almost like a move of frustration or oversight, which I don't think a 3300 is capable of. Definitely a curious move. Maybe it knew it was already lost because I can see some nice queenside breaks and the knight is always going to be a problem, maybe with sacrifices on the kingside when black tries to defend the queenside. I think black's general idea is to play queen d6 here after white retreats. It's almost like he didn't see the knight fork. Then black wants to sacrifice on d3 to get an attack, but that's just me thinking desperately. Weird game. This is chess, though, were opponents often kill themselves if you play patiently. Maybe that happens with engines too and is the nature of the game. Great video, thanks.

[@Tiriondil]

@Homework: My guess is that inside the search horizon of Houdini this was the least worse move, especially as this is a rather closed position. You always have to consider that chess programs like Houdini or Stockfish follow another paradigm. During the search process Leela and AlphaZero always play the game to the end, Houdini and Stockfish don't. So chess playing neural networks can do strategic moves, which is something between difficult and impossible for the classic alpha beta search.

[@morsmortis4114]

What's funny is that always when some guy in a chess game analysys video asks me to guess the move of leela one can actually get it right quite often since you just have to think of a crazy move like an exchange sac or a hyper aggressive pawn push :D

[@maxpheby7287]

In relation to the homework question maybe Houdini is set to avoid drawing lines and so played a line although risky had better chances of winning. Or simply the knight is worth more than a rook in this position. But yeah is strange. Also Leela played differently again to the pawn storm style of before. Its great stuff.

[@mattroxursoul]

Homework at 16:15 could you check king with knight, then the king has to protect the rook. so moves to f8 or g8. if its g8 the knight could go to 6h and check again. the queen is free to attack rook at this point it appears.

[@0DarkTime0]

Its nice how you swapped the view for showing blacks perspective at this position :)

[@TheJayMoses]

Tough position to advance for sure. Maybe knight f5 looking to try to strand the g7 rook for the queen or knight to take. It would all depend on how the king reacts to knight f5. 🧐

[@TreyMasta]

I honestly came up with Re4. Honestly, only after he set up their was going to be a unique move though.

[@KillianDefaoite]

10:25 For some reason, Re4 just really jumps out at me. It frees White's pieces and shuts down Black's attack.

[@markrobs2954]

I thought rook e4 but only because you showed a variation with the same move so I'm like why not.

[@aminvardi2326]

anyone has leelas number?

[@JB_inks]

It would be nice to see two engines playing on evenly matched hardware. A GPU versus 2 CPU cores isn't a fair fight

[@christoph4454]

But if both engines use the same Hardware it will always be a draw

[@niecierpliwy3195]

I enjoyed this video

[@HebiSnake]

I don't think the GPU aka graphics card would affect a chess engine. That's more of a processor type of thing as far as I know.

[@brosephjames]

Deep Learning chess engines like Leela use GPUs because the type of algorithms they calculate are more efficient on a graphics processor than a general cpu.

[@HebiSnake]

brosephjames Oh, that's cool. Today I learned.

[@xtreme1002003]

Fry must be really proud of Leela.

[@thom1218]

+ChessNetwork - CCRL rating for Houdini 6 is 3380 on given hardware is for standard match time controls. Its rating is about ~350 to ~400 ELO points lower at roughly 2 seconds per move, where Leela has a huge advantage for having to search significantly fewer nodes to reach her maximum ELO rating.

[@conimede]

thom1218 Leela is penalized a lot more by low time controls. Due to the nature of monte carlo tree search, she can (and does) blunder horribly sometimes, missing shallow tactics (even mate in 1!), if she’s not given enough time.

[@thom1218]

+conimede - you're right... only if she's searching 1 or 50 nodes (easy, and normal mode on the demo site)... on a GTX 1060, even running the non-optimized opencl version, leela can search 1000 (1kn) nodes per second - so she's averaging 2kn/s per move in the match... The point is you only see Leela playing these super quick time controls... because she can hold her own in them (i.e. because Houdini takes a bigger hit from time controls than Leela does)

[@conimede]

thom1218 I’m not convinced. Leela goes at 1000 nodes per second, but Houdini analyzes *millions* per second. And a/b engines’ iterative deepening ensures that shallow tactical mistakes will never be made. 1000 nodes is still really little: Monte Carlo Tree Search favors focusing on a few lines in depth, so there is still a strong risk that Leela will blunder by not analyzing obvious moves. And she does! A lot! Those games simply are not published because they’re completely uninteresting. I think the reason why so many games are played at low time controls is that it’s simpler to play many of them and to cherrypick interesting wins.

[@thom1218]

+conimede I don't disagree with what you're saying, but to convince you that I'm correct, simply do the following experiment (as I have done): set both Leela and Stockfish 9 to a limited search depth of 1 node (in Arena for example)... Leela DESTROYS stockfish EVERY TIME... with both white and black pieces. As you can see, her net (a.k.a. extremely complex eval function) is ... superior ... at shallow search depths (a.k.a time controls). There you have it. In other words, Leela "frontloads" her chess knowledge, and gains much less stregth from brute force, as compared to a traditional engine. Yet again, lower time controls favor her, not brute force.

[@thom1218]

+conimede To address your fixation on missed tactics, Leela's net can overcome tactics (to a certain point) with its "intuitive" sense of positional play (again, at lower time controls)

[@mikechambers9129]

It would be interesting to hear the homework answer from Houdini's authors.

[@tweschke3]

Thanks

[@identityvdreamwitch3557]

I got all of leela's moves! I don't know if that makes me good or bad since I'm only rated at 1000. For the homework. Imo black recognized after playing g5 the knight was worth more then his rook with the incredible g5 home and check that came with it. Besides getting rid of the knight, black gets to strengthen his own pawn chain, and has an open c column to eye down whites weakest pawn. For the trade, black knew the position would just be better

[@russellfroggatt]

Amazing game. Leela is really strategic.

[@ChessNetwork]

A universal player. Queen of tactics and strategy :)