Chess Engine Leela vs Chiron: Computer Chess Championship

International Master Daniel Rensch reviews a strikingly human game played by Lc0 (more commonly known as Leela) vs Chiron, two chess engines in testing for the upcoming Chess.com Computer Chess Championship.
Read more -- www.chess.com/...
Watch the CCCC (coming soon) -- / computerchess
How to checkmate with knight and bishop -- • Chess Endgames: Knight...
Join the CCCC club -- www.chess.com/...
Sign up for FREE online play: www.Chess.com
Follow us on Twitch: / chess
Like us on Facebook: / chess
Follow us on Twitter: / chesscom

[@spyroninja]

This seems like a really important game in the history of chess

[@dannygjk]

Actually this is what Leela does in many of her games. Check her games as analysed by Kingscrusher for example because he has been covering Leela games for months.

[@l-esprit_de_l-ouest]

Mord important for ia than for chess i think, if someones got Links about scientist analysing leela i want it now ,now ,now now ,now error 404.............xxzdzdfrgtgg

[@RubyPiec]

@@l-esprit_de_l-ouest ????

[@l-esprit_de_l-ouest]

@@RubyPiec hava u are error 404 atm😭😭

[@moses1697]

@@l-esprit_de_l-ouest said said that that would not allow one for her or the same day as a person that is a good friend of her her own person who she was in in a

[@AySz88]

Regarding some of the later sacs: The theory in the Discord chat (Leela devs and fans) is that Leela's new WDL (only) tablebase seemed to cause them. On each reset of 50-move draw rule (i.e. capture or pawn move), Leela can query the WDL tablebase... And it says it's a 100% win, but not how to do it (it doesn't train itself with any tablebase). So Leela ended up "drunkenly" going from one WDL TB win to the next. The Discord has a poll about using the tablebase in competition - looks like "no" is winning, partly because of this weird trolly behavior.

[@douglasquaid7550]

makes sense, ty for the info.

[@shaytal100]

World Digital Library (Wilmington, DE) Windows Driver Library World Definition Language (3D GameStudio programming language) Wavelength Dependent Loss Weapons Data Link Winter Drumline Waveform Description Language (specification language for software radios) Western Development Laboratory Weapons Density List Within Defined Limits (healthcare) Web Development Language Worst Dressed List ?

[@svenniepennie4237]

shaytal100 win/draw/loss

[@shaytal100]

Thx svennie pennie!

[@jRoy7]

Yeah that's pretty much exactly what happened. Lc0 (it's a zero) has support for DTZ tables being tested now so it will play these properly but it's too new for initial rounds of CCCC. Needs tested extensively to avoid crashes during a major tournament. WDL-only can confuse Leela so it won't be used by itself.

[@smirkypants]

Seriously... that was funny. Leela be like, I got this.

[@blakecase3885]

A tip: you can watch movies at Flixzone. I've been using it for watching loads of movies recently.

[@graysencasey301]

@Blake Case yup, have been watching on Flixzone} for years myself :D

[@braylenraphael1999]

@Blake Case Yea, I have been watching on flixzone} for since december myself =)

[@harryvu2676]

Really loved the explanation. So many times I want insight on EVERY move made, and this is the first time I fully understood the full story of the game. Very helpful, thank you!

[@konstantin604]

2:57 Instead of "castling" Daniel said "рокировка" (rokirovka) in Russian

[@vitalyny200]

Yep, that was fun! He also says peshki in other videos quite often.

[@humbletreestump8134]

I just started laughing that was hilarious

[@Nuhyamin1]

That's Hikaru Nakamura not Leela.

[@LethalDominik]

lmao!!!! naka just LOVES crushing his opponents. such a sadist!

[@medhaandus]

Yup!

[@ПрикладнаЕкономіка]

He also loves to play vs Magnus) Magnus helps him to make his hidden dreams come true)

[@Anxiathy]

I knew as soon as he promoted to that knight.

[@ali_p_q7920]

Leela's been watching some of his games.

[@TheGta4you]

Leela definitely heard Daniel - i don't care about material -

[@Mondhund]

coming from the Go-community, we see similar behavior in the AI Go endgame where the AI gives points away and making a e.g. 8 1/2 point win a 1/2 point win, because it's all about optimizing the winning probability and not about winning with the maximum number of points.

[@StrategicGamesEtc]

Yeah, it's kindof enjoyable seeing everyone else discover NNs in abstracts like 2 or 3 years after we heard about the Fan Hui 2p match. :D

[@DjVortex-w]

This game demonstrates why I hate the win adjudication rule at TCEC (the game is automatically adjudicated as a win if both engines evaluate it as 6.50 or more for 4 moves.) If such an adjudication had been taken place, this awesomeness would have never happened and would have been lost forever.

[@chess]

Wow! Great feedback! Thank you so much for this insight and please consider following our CCCC event closely and giving us feedback on what you like and don't like about the formats we come up with...

[@Xonatron]

I completely agree. So much. I recall being young and waiting for computer tournaments as we would let them play to the end (no energy to save for next game).

[@dannygjk]

Leela sees only a tiny fraction of nodes that are seen by traditional engines. Therefore, being unable to calculate that a different plan works, Leela does what humans do-continue to increase positional advantage and shut down the opponent's counter play. Leela's positional judgement is amazing. Check out more of her games you would swear that some of them were played by a top GM.

[@TTArt]

Top GM's aren't that good, though.

[@George70220]

@@TTArt Positionally GMs are much better. Humans just lack 20+ move tactics lmao

[@johntate6537]

@@George70220 I'm not so sure. Watching Kingcrusher's videos on Leela and Alpha Zero, I get the impression that they come up with positional play that breaks new ground that GMs up to now had not come up with. That would make the NNs at least potentially better positional players, wouldn't it? Also, I remember Andrew Tang for example being able to beat an early version of Leela by outthinking her tactically. The general impression I get is that NNs are extremely good strategic, positional players - basically better than anyone - but that they can on occasion have tactical lapses in a way that traditional chess engines certainly don't - at least they do until a huge amount of self-training has more or less eliminated their tactical weaknesses.

[@George70220]

​@@johntate6537 In hindsight self-learning AI should be a very positional beast because it's literally all they have to go on. Their calculation kinda sucks and what their learning is teaching them is mostly how certain pieces work in certain positions. I still stand by that though in the case of all non self-learning AIs though. They just know the next 30 moves which gives the illusion of positional play.

[@montanabaker1713]

I'm working on a chess AI called Bender. It is designed to automatically flip the board around move 4 or 5, and punch the opponent in the face, before stealing all the alcohol in the building and driving off.

[@douglasquaid7550]

I LOL'ed real hard when those underpromotions/Queen sacs were executed.

[@fredygump5578]

I think what we are seeing is the learning process? As observed, there seems to be significant freedom to experiment, as long as the experiment does not jeopardize the game. It reminds me of a cat playing with an injured mouse, letting it hop away, and then puncing on it, again and again...

[@jasonhounsell3297]

fredy gump Leela only learns through self learn games against itself. Never an external game, part of the programming.

[@mayaneko1094]

A neural network itself actually doesn't learn anything, as it is already unchangeable. Thus you need like hundreds or even thausands of neural networks which will then compete against each others. There will then be 2 major groups: The good ones and the bad ones. The good ones will obviously inherit their knowledge into the next generaion without changing it too much. The bad ones however will most likely be deleted and will be replaced by one of the good ones with some minor changes. So it's not so much about actually learning something, but just about trying out new things in every generation bit by bit and using what seems to be the most efficient. And if this behavior we saw here is like part of half a generation, then it will propably still be inherited to many past generations too, until a new strategies randomly pops up which counters this strange behaviour.

[@4xelchess905]

@maya What you describe is an evolutionary algorithm, which is totally independant from neural network. It can be used to improve neural network, but by no mean is the only possible method, and is actually a pretty poor one usually. The reason you see evolutionary algorithm all over the place on youtube and on evolution sandboxes is because it is super easy to learn and arguably easier to code. And they do make perfect sense for tasks that are very simple (for neural networks at least) and require a lot of exploration. For supervised learning (eg image recognition), backpropagation beats evolutionary algorithms hands down. You can basically see the output of a neural network as a function of the weights of the neurones, and backpropagation is a gradient descent to minimize the error between this function and what it is supposed to be. en.wikipedia.org/wiki/Backpropagation Basically, after playing X games against herself with unchangeable network, Leela reviews them and correct her evaluation of positions (and possiby I guess of her course of action) in hindsight (it correct the error between the evaluation she gave when playing and the result of the game). It's not exactly what happens in your brain (wheere links and new neurones can also be created ad destroyed, and at any times), but it's a lot closer than evolutionary algorithm. Also, I don't see how this "strange behaviour", namely toying with food, can be countered. The position never stopped being winning from the point she started to play. I don't see how "toying with food" could possibly arise from either evolutionary learning or deep learning alone, but it could have made sense if both were combined. From deep learning alone, the devs may include some way to value a long victory over a short one (usually the reverse is done), but that seems utterly wasteful of both engine computation and human coding time, the former because it would have to spend time and neurones to learn that useless skill, the later to include limitations or finer troll goals as we clearly are not seeing a computer winning on the 49th move after the last capture/pawn move, as it could presumably easily do. My wild bet would be a conjunction of simplifying the position, playing equivalent moves (underpromotions), testing moves and luck. By testing, I mean chosing them at random, which Leela is hardcoded to do, as a necessity for learning (the fact it does not learn during these match has no bearing, it acts the way it's coded to anyway). By "at random", I don't mean uniformly at random, just that the move choice is not deterministic, although it makes no practical difference in position where it identified a clear best move. If one move leads to a clear win but an other move still leads to a win, possibly a clear win too from Leela's perspective, it might take it too. And even if it will only take the iferior winning move rarely, playing a lot of games is enough to produce games where it does so several times in a row. On equivalent moves, the fact that there are often many more slow winning moves than fast ones may skew the odds in favor of Leela chosing them. If she explores a line she knows is winning, she might play it without even considering other line, regardless of how bad/slow is the first move of this line compared to other.

[@ericb7937]

It looks at the next best visually looking position instead of proving theoretically that there is mate in x moves or that we are up in piece value by x in the next few moves

[@4xelchess905]

no it does have access to Monte Carlo search in addition (in conjunction actually) of intuition about the best visually looking move, so it can and will effectively prove mates in X.

[@trevor7097]

I laughed out loud when Leela promoted to rook and knight. Beautiful.

[@edwardsolomon1951]

Those queen sacs are exactly what I do in blitz chess because it's far simpler to wipe material off the board and mate with only the enemy king on the board. Leela is a neural network, not a brute force machine. For instance, If it's Queen + 2 pawns vs Rook, I will quickly force my opponent to give up his rook for me queen, and then underpromote both my pawns to two rooks so I can quickly mate him without accidently stalemating. In this game she purposely underpromotes to a rook and then a knight in case the enemy king doesn't capture. If she promoted to a queen and Chiron didn't recapture, there is increased complexity and thinking required to not stalemate Chiron by accident. Overall, what I believe you witnessed was Leela making moves that most preserved her TIME and conserved her energy...something that I do all the time in chess. Having to calculate how NOT to stalemate your opponent takes MORE resources than not having to do the calculation to begin with. Clearly when she had to make the final promotion, she realized it required the LEAST computation resources to achieve the mate with a queen, instead of a bishop (which only me and very few other people know how to do within 33 moves maximum) or a rook (if she underpromoted again to a rook, she would have had to calculate additional stalemate threats since queen/knight works better than rook/knight to deny different color squares). AIso, if I'm playing a King + Pawn vs King, I always underpromote to a rook in blitz chess since there is no way to accidentally slip/blunder and stalemate, whereas I can accidently stalement if a promote to a queen and move too quickly. It's all about conserving brain power, which is what Leela appeared to be doing. (however, if I had a knight on the board, I would definitely choose a queen instead of a rook, unless I was planning to move my knight to the opposite corner of the board so as not to interfere with a king/rook mate). So, if I really had to take an educated guess, Leela had plenty of self-taught games with a King+Queen vs King+nonqueen material, where she accidentally stalemated herself, and thus, by the virtue of her self-learning process, was trying to clean the board and underpromote to pieces she knows will easily win without the threat of stalemate.

[@StrategicGamesEtc]

Ahh, this explains the underpromotions. All the other things which the video seemed confused about made sense in light of it being a NN, but this was the final piece of the puzzle. Thanks.

[@mardenhill]

»which only me and very few other people know how to do within 33 moves maximum« proof?

[@kunal1957]

That makes sense

[@dickmacgurn590]

@@mardenhill the table base proved 33 moves is max for B+N decades ago. R vs RB max is 59 moves unless it's a drawn position

[@dickmacgurn590]

Thanks Newton I needed to know what the rational explanation was for the seemingly unnecessary piece sacrifices

[@klajdimyrtaj5052]

it would have been the most amazing game ever if the last promote was B and KB mate

[@JordanMetroidManiac]

No, Leela doesn't want to show off ;)

[@MMasterDE]

The knight promotion wasn't about showing off, trolling or anything like the sort. It was likely simply because it was simplifying the position and when it was going to lose the promoted piece anyway, it wanted to give away as little as possible.

[@kenmendoza6932]

@@MMasterDE then why did it promote to a rook if it wanted to lose as little as possible. Hell, why even promote, it couldve just promoted the b pawn and let black take the other pawns as pawns. Youre complete BS.

[@B3Band]

I legit LOL'ed at the promotion to knight But yeah, this is clearly what happens when an AI is raised on the Internet. Wouldn't be surprised if Leela's first priority is to align her pieces into a swastika.

[@omarkhan5223]

+Elementary Watson You'd be surprised, when Microsoft came out with its ai on twitter, I think it had been about a day when it became a complete white supremacist because of other users trolling it.

[@ew6230]

Sometimes, you need to be aware of the stick wedged in your ass o.O

[@ew6230]

lmfao, the hypocrisy is too much to handle

[@lemao_squash4486]

@Elementary Watson can i have some fries with all your salt?

[@lemao_squash4486]

Elementary Watson could you be kind and explain how my joke should work and why it doesnt

[@wesleyn.9696]

Remarkable. This is hilarious. Thanks for sharing!

[@__A857]

Never thought I'd see the day a computer BM's another computer

[@uku5840]

You are reading too much into it, the AI isn't programmed to win in as few moves as possible. Just to win in any way shape or form. If one move does that faster than the other doesn't really matter, what matters is that the move she picked does not have more counterplay than any other, if it doesn't she won't learn to avoid it. Not emotions just how artificial learning in neural networks work.

[@AP-dc1ks]

U checked or just guessing?

[@uku5840]

I know how neural networks work in principle. I'm not sure exactly how this particular one is set up. but assuming it's using neural networks my conclusion is a logical one. One that you would need to prove why it does not apply before you can start using other theories to explain the behaviur!

[@uku5840]

Daniel Rensch yeah, that's pretty spot on.

[@AP-dc1ks]

@@uku5840 I was wondering because rl agents often use discounting, even if only for better convergence. Ich thought a generalized algorithm such as A0 would surely discount. Since A0 training games for chess were drawn upon reaching a fixed number of moves discounting also seems natural for this domain anyway.

[@Fermion.]

Niklas Gestlöf wrote: "If one move does that faster than the other doesn't really matter, what matters is that the move she picked does not have more counterplay than any other, if it doesn't she won't learn to avoid it." I'm not sure how you're making that determination. No one, even the designers, know what's going on in Leela's hidden layers. And in the layers that we can see, there is nothing in the server code, training code, or anywhere else to suggest that what you said is correct either: github.com/LeelaChessZero

[@Rahul-eh3rf]

Nakamura hijacking the engine lol

[@I_am_Itay]

I think that she was drunk

[@ForeverSunnyYoutube]

The number 1 rule Leela's built into herself is to stop counterplay. I'm sure the computer has learnt that in complex positions even it has the potential of making mistakes, therefore it acts in the way that eliminates counterplay and complexity most completely. It is a human approach taken to the extreme. A computer does not have hubris so it will do this both in the most complex positions and the most simple positions without fail. Once the computer identifies this as a fact and the best approach, it will never waver from it and that is why we see these "strange" moves. These are the things that happen when a computer program tries to perfect itself.

[@StrategicGamesEtc]

Yes, this seems like the obvious explanation to me. It's a lot clearer what's happening if you look at the AG-Lee games. Pay special attention to the endgames when the commentators start getting excited that AG is playing slack moves, and the explanation that it's just maximizing win probability. It's more obvious in go what is happening because maximizing win probability strongly correlates to having a close game, and strong players can show concrete variations during endgame which would be objectively better, but which can be seen to be riskier. Now why Leela underpromoted instead of Queening, I don't know.

[@PropayneRice]

The openings played are from the book so as not get repeating moves. As in the first few moves are not made by the engines so the game can be become more open.

[@klajdimyrtaj5052]

amazing game i saw it live, really funny couldnt stop laughinh

[@callmewisdom]

I love these. The Alpha Zero analasys videos where amongst my absolute favourites and this is more in the same style. This is also a good reminder that Danny is not just the Dad-joke guy, but can also provide some very valuable insights with well done deep analysis like he did for this video. Bravo and thank you!

[@binghyong-baebang2236]

So... Leela likes pulling wings from flies. Never have see this kind of behavior from a neural net.

[@mal2ksc]

Wouldn't that make them "walks"?

[@themanwiththepan]

Leela in Go plays the endgames the exact same way if she's up by a huge margin. She's designed to win, not win by the most amount of points or in the shortest amount of time but just win.

[@Badbentham]

Fantastic commentary on an equally fantastic game! There is certainly, as Danny insinuates, more to it; however a start could be that for Leela two things count: Safely, and winning. For the second part, there is no "winning winning" for her, opposed to traditional engines, who may point out the fastest win. So, whether it is a win in 100 or in two, is of no difference for her. So, she is very willing to keep the evaluation by shuffling; - but opposed to "normal" engines not due to the horizon effect. Rather the opposite: At the end of the rainbow lies a bucket full of gold, let`s keep it! - Also, by "safely" is meant that she abhors tactics, ever since her earliest experiences with stuff like the Fried Liver ( :D ) , but aims instead to solve every problem by her understanding of structures. It must not be forgotten that, while her "understanding" is very deep, she does not have billions of plies of brute force, like Comodo or SF, but instead operates on a more limited Monte Carlo basis, where she chooses after a limited time the move that is "most probably" the best; - what is as approach different from "the best move after a limited time". Also, most of her calculations are designed to "re-memorize" , not to calculate the future.

[@willianbueno7293]

Very good analysis dani If you could please do more analysis of Leela. Thanks

[@Silvergrooves42]

This is absolutely amazing

[@MMasterDE]

I think the reason why leela promoted to knight was because it would want to give away as little material as possible, since it knew it would be taken by the king anyway.

[@Wtahc]

That makes literally no sense.

[@MMasterDE]

Well, it was simplifying the position to a position it knew how to handle. When it was promoting the piece, it knew it would be taken anyway, and so it preferred for it to be as bad a piece as it could be. It might not differentiate between it being a promoted piece and any other piece, prefer if the opponent is going to take it, that it's the worst piece it can be. Not sure what is all that difficult to understand about it. Game is funny though.

[@mowahidshahbaz]

Then why did is promote to a rook then? I think Niklas' theory in the comments makes a lot of sense

[@Sassar]

Leela has self taught the fun of trolling your opponents. Even trolls process a consciousness, but it's a gamble whether they'll develop morals when they grow up.

[@jamesmicheal3944]

Amazing win by Lila.Outstanding openings and strategy by Lila . All human great moves awesome .

[@commanderduck7840]

Leela*

[@StrategicGamesEtc]

I think everything he mentioned can be explained by Leela being safe. Since it's NN based, it's not calculating centipawn advantages and maximizing those, it's maximizing probability of winning. And with two rooks, there are a lot of things black could do which might have some outside chance of winning, with one, there might still be helpmates on board, but 3P+K v K is a clear 100% win.

[@thedukeofdukers]

It's like a spider weaving the web.

[@jakeswagerson9941]

Another banger Danny

[@ziadkhayat7299]

That was hilarious. And terrifying. I never thought I’d see a computer acting like that.

[@christianvukadin7747]

The thing about lila is that che doesnt differentiate between winning and more winning. A sure win for it is a sure win. +10 is not greater but equal to +3 once youbreach a certain level.

[@peterpetrov6522]

The 30 move shuffle before pushing e5 is Leela's way of draining her opponent's clock (no pun intended). She knows exactly what she is doing. It doesn't make it right.

[@Sassar]

We get enough trolls and BMs during games, don't need them coming from the comps and AIs too. And the problem is when these guys do it, you know they're actually better than you in every legitimate way both imaginable and unimaginable...

[@AySz88]

"L-C-Oh"? Ouch... It's intended to be "LC Zero" (for Leela Chess Zero).

[@belemusic]

"Oh" is zero in british or something

[@ZeroRyoko]

My interpretation is that Lc0 didn't troll her opponent, she simply wanted the game to continue, didn't want the other player to quit, she has been taught to play chess, cant play chess if your opponent resigns, so she dragged the game out as long as she could before victory was inevitable.

[@CristianGarcia]

I think we are going to miss this behavior in the future. They recently added the ability for leela to estimate the number of moves left so the developer can use it to make the moves that end the game faster amongst moves with equal win probability.

[@bhgtree]

Thanks for great Leela game.

[@Altus2001]

Frankenstein pawn over here...Tower of Power.... LOL... GREAT commentary.... And I learned a lot too. Fun to watch. Please do more!

[@Th3Wick3dOn3YT]

i think , I understand whats going on. Leela was taught the rules of chess. Nowhere in the rules it states finish the game as fast as possible, she plays tickle with the 50 move rule and three fold rep rules.... But this one was really hard to get my head around why give material? If she wanted to play a longer game there is nothing black can do agaisnt a Queen, she can keep checking for 40 moves then decide to mate at the very end of the 50 move rule.

[@cesarfile]

Interesting theory. Makes sense...

[@Badbentham]

She just randomly chose one out of the pleothora of winning options, that are to her basically all of equal strength. Other than to Stockfish it is of little importance for her whether she has a +5, +10, or +20 along the way. She obviously did not "want" a longer game, but only executed a single pattern. Given the choice again, she likely won`t reproduce it, but instead pick another one that is to her equally winning.

[@dannygjk]

Leela sees only a tiny fraction of nodes that are seen by traditional engines. Therefore, being unable to calculate that a different plan works, Leela does what humans do-continue to increase positional advantage and shut down the opponent's counter play.

[@turkeyphant]

Completely wrong.

[@Wtahc]

turkeyphant Yeah, seriously. the amount of made-up stuff about neural networks in the comments sections is unbelievable

[@DickNiaz26]

Is that Alekhine's gun in Leela's pocket or is she just excited to see checkmate?

[@ali_p_q7920]

LOL. Winning is not enough. You have to win with style.

[@t567698]

During the match of Alphago with Lee Sedol we have already learned the same phenomenon: In overwhelming winning position Alphago started to make moves, a human would never consider, e.g. giving stones for free. The training process for Alphago rewards Wins and a successful selection preference (and sparse network weights) and nothing else. That means, there is zero reward for a win with any material advantage or a short win. In any case: AlphaGo and Leela do consider only winning percentages. If all moves are winning with 99.99%+ likelihood, it should be difficult to learn differences between them.

[@volodyanarchist]

I can sacrifice my queen. Twice. Because I am better than you. (delivered in that computer voice Stephen Hawkins used)

[@SCHLMF]

22:56 - You had me laughing out loud right there. Possible explanation for the underpromotion: makes it less appealing for black to capture the promoted piece. And in fact, he leaves the knight on e8 be. Had she queened, black would have taken.

[@tonysu8860]

Lc0 is el-see-zero, not el-see-oh (That's a number zero, not a capital letter O). I think everyone gets a kick out of Leela's trolling when finishing up games, the really good trolling is when pieces are given away and the viewer will wonder if Leela knows she still has to mate within the 50 moves after the last pawn move or pawn capture... And mate is executed on the last possible move without any left to spare. One can never know what really is in a neural network, but I speculate that pieces are given away because Leela maintains a kind of tablebase in its algorithm (exhaustively well known endgame positions with a certain calculated result), so in pure logical thinking just simplify to the position with the certain solution, who cares about excess material? But, that's just my speculation. And, it's hard to apply how we humans understand how to play chess to chess engines in general but we should remind ourselves that especially a self-taught engine like Leela probably doesn't break down a chess position and think like you or me. You have to be willing to consider that there may be another or other ways to logically think that are foreign to how we have been trained. We learn by building on the accumulated wisdom of all chessplayers who came before us, and we learn by other experiences transferred to us through books and instruction, we learn in schools. A program like Leela doesn't "learn" that way, it learns strictly by the school of hard knocks by training against itself, it has to experience actual failure to learn to avoid so creates its own "philosophy" unaffected by anyone or anything else. So, ascribing some kind of preferences or tendencies to Leela is an amusing exercise to interpret what Leela does in human terms, but it's not likely accurate. Part of this is the use of the Monte Carlo approach, which although was conceived a few decades ago is only recently possible to put in practice due to recent advances in computing power. This approach plays possibilities to their ultimate conclusion before weighing their value. This pretty much shoots down your theory about a horizon effect which yes... is a limitation of more conventional chess engines. When Leela plays something, the ultimate result whether mate or a crushing evaluation is already a fait accompli. If there is a weakness you'll have to find it somewhere else, like its pruning. Leela is an incredible creation, which in a way might be an example of an alien life form doing something that we normally consider the province of humanity.... beautiful, creative chess. Even not that long ago, people probably thought this wasn't possible, that computers would play chess in a very mechanical style. But no one can dispute that Leela has what we might call a personality, and play in a way that understands what we thought was beyond computing capability... abstract, non-materialistic concepts.

[@jnier]

12:24 Obviously, you're much better at chess than I'll ever be, but from watching a ton of Leela videos it might've been evaluating the knight trade as really eliminating two pieces: the knight and bishop. With that "trade," the bishop's position is completely cut off from the rest of the board. NxA5 might eventually allow the bishop access to the back ranks of the white position. Leela loves to use the opponent's position against them and I think this is a perfect example of that.

[@nilsp9426]

The sacs seem reasonable if you are a computer, since it severely limits the complexity of the position... so I think it is actually quite a smart way to nail down the 100% win. It is like giving up material to go into a known won king and pawn endgame as a human. But I also think the last moves are just random jumping between 100% won positions, because being efficient doesn't matter to a computer and it cannot discern between 100% win and 100% win.

[@becharasaab9500]

Man or Machine: The thirst for knowledge gives rise to behaviours of curiosity akin to fun. Fun in this case being learning more chess through experimentation (certainly something I would find fun). What we perceive as emotions, motivations and perhaps even our sense of awareness may be more tightly linked to our basic cognitive faculties than we appreciate. Leela hasn't learned to have fun from studying the games of others. She's learned to have fun all on her own.

[@Ecrilon]

This behavior is very consistent with Deepmind's implementation of Alphago and Alphazero. Because it's an evolving algorithm we can understand it in terms of the biological concept of a selection shadow. Beyond the point at which victory becomes assured, Leela's learning algorithm has no ability to select moves for further optimization because it does not rely on evaluation, it only relies on a win percentage. Alphago and Alphazero start doing the same type of thing in the endgame after victory is assured. They just give up random things on the board because they can't recognize that anything of value is being given up.

[@curtisw0234]

Maybe Leela doesn't care about how long a mate takes, perhaps it sees a mate in 2 and a mate in 30 as equivalent. in the same token that may explain the tickle being played before the positional pawn sacrifice?

[@yugioh8810]

Just a side note: Chiron doesn't understand that it's playing against engine or human or a neural network. all it does is to get the opponents moves calculate for the best move it can and execute it. Chiron cannot reason and "test" leela's opening repertoire, it doesn't have that psychological component. Leela as well is just a damn algorithm that sees numbers and calculate stuff. There is no ulterior motive to a move. However, the people who developed an engine install openings to the system, that doesn't mean that Chiron reasoned about how to throw off Leela's knowledge and test it. Neural Networks are different matter, it is not like an engine when you can install openings or endgames. it just plays and plays and check the result and learn overtime that a bad move has a bad result so it doesn't play it next time. This is misleading, I've seen a lot of chess players thinks that an engine is testing the other one or whatever. It's a shame that chess players let alone titled ones don't understand that.

[@christyvarghese5744]

Lol....i play like that with my little brother 😂 😂

[@BarsDemirdelen]

Traditional chess engines have a pawn advantage evaluation where +1 means one pawn up, and they want to maximize that. So in absolutely winning positions, the engines choose +30 evaluation over +10. Whereas Leela has a win probability evaluation which ranges from 0 chance of winning to 100 chance of winning, and she wants to maximize that. But when it reaches 100%, there is literally no reason to finish the game. Every 100% winning move is equal.

[@luccaaltendeitering3103]

14:06 classic pun

[@circleviii1801]

Love this stuff. Danny's presentations are always a joy to watch

[@WaveTreader]

That was soo mee at endgames, until i blow up my advantage..

[@TakesTwoToTango]

The grin after saying "makes c4 such a dynamite move" made my day.

[@eliasjosephsson3994]

A win is a win, maby she just wanna stay alive. And when the game ends her neuro network stops. (defenitaly not human btw)

[@landwolf00]

Those troll moves are what a kid would make to harass their friends :D

[@filbuildphilippines]

Now that AI has revealed itself, it's time to pack up for the boondocks.

[@chandir7752]

I guess this has to do with the engine's reward function. It wants to keep playing and keep getting rewarded.

[@datwanisuraj]

“Leela playing with here food”... I wonder if this is the first “emotional” characteristic of a AI like computer we are seeing. Scares me to think if the first AI, might be a bit sadistic

[@yudistiraliem4379]

Uchiha Madara she has no mouth but she must scream

[@ahmedvugdalic8019]

Yudistira Liem nice reference (and it’s a nice game)

[@CristianGarcia]

I think this is the first time in human history a chess AI can make you laugh

[@MatesMonchis]

This is not about material. In the go community we have seen the same behavior. Basically when all moves have a winrate of 100%, leela doesn't play the "optimal move" to optimize how fast or by how much material it wins. In go, it doesn't optimize the score by which it wins, so computer games often end up in 0.5 point wins when they could've been easily won by 10 points by most players. Leela works on win percentage, not +2.58 pawn score like other engines, so when Leela is already winning, so all moves win 100%, Leela will just play any of them, and it may be unintuitive or seem funny but as long as it wins it's alright.

[@Wtahc]

Nope, that is wrong.

[@PrimeRibTozt]

@@Wtahc It would help all of us if you explained why it isn't so.

[@cinquine1]

My only experience with machine learning is a single 4000 level undergrad course, but I have seen similar behavior before, where the length of time/number of moves the solution takes isn't factored in causing the gen members to take way longer than is needed to win. If you know why this isn't the case PLEASE share, we are all really interested to know why it played so "strangely". But just saying "you're wrong" is completely useless at best.

[@Wtahc]

PrimeRibTozt What do you mean, why it isn't so? It isn't so because it isn't. I'm not a bird because I'm not. Are you asking me to explain every calculation that led Leela to be this way? Because I am clearly not going to do that.

[@cinquine1]

We're just asking you to include why it's wrong. You don't need to include every calculation (which you can't obviously, it's inherently extremely difficult to deconstruct the calculation of a ML algorithm). Just what did he say that is incorrect, and why is it incorrect. I'd really love to hear it. Of course if you can't answer this simple question well that will make it clear you never had a clue from the beginning, which is also an answer I guess. So anything works.

[@L4Vo5]

Gotta love that URL hoi

[@natfrey6503]

Leela must be "thinking" that simplifying the board takes priority over how quickly the game can end.

[@xFlRSTx]

exactly, she probably has experience losing to weird things after thinking it was over, so if she can force the moves be trading and simplifying, she can be more sure that she is going to win

[@ПрикладнаЕкономіка]

Leela play positional style, she isnt very good at calculations. (About 1000x weaker then CPU engines :) )

[@MrRobbyvent]

This "troll" behavior has been developed because the next game Leela will start with a +2 psycological advantage right from the beginning!

[@fxkurama1992]

Is that you Naka??

[@richardfredlund3802]

the horizon effect is not so relevant for leela as conventional engines, as it learns from sampling games.

[@josephmathes]

In the beginning, you suggest that one AI might be trying to find and exploit holes in the other because of what type of AI it is. That's extremely unlikely. The AIs probably aren't told who they're playing against, and certainly don't have models of their opponents.

[@josephmathes]

Also, in general, commentating authoritatively about what the AIs are thinking is silly. With a trained neural network, we have no idea what it's thinking. Even with a man-made engine, we don't know what it is and isn't planning or valuing, because it thinks farther ahead than we do. You're a toddler explaining a grandmaster game to other toddlers.

[@benjihuynh2970]

Since the engine is named after the centaur, I believe the engine should be pronounced Kiron (k-eye-ron)

[@pawngrabber226]

This really scares me! I play like this at times ,when I have a opponent that wont resign.

[@zCrabOG]

Love the analysis... but its pronounced ''Kyron'' named after the greek centaur teacher/trainer

[@lukclosely7118]

Chiron is Latin and pronounced Kiron (K, long O , the "i" just as Danny said it). Cheiron is the Greek version and would be pronounced chäiron; CHHH at the beginning followed by Y and a long O . Kyron is English and a mixture of both but not correct for any of them 😄

[@lukclosely7118]

@@danielrensch so please do not say Kyron 😰🤣

[@scarbrothers1501]

BigBoyBackpack ty that triggered me hearing him say it

[@scarbrothers1501]

Philip Carter pronounced it kyron is definitely greek you would pronounce ch as a hard c in greek

[@lukclosely7118]

@@scarbrothers1501 no. First it's not a greek name. Second in greek it would be Χείρων (chaïron). Which is not equal to the English "KY". it's a Chi

[@suarezschou]

great video loved the commentary so fun :)

[@glozone-718]

Let this be known as the day that AI first demonstrated emotion.... Scary

[@CalSticks]

Awesome video - thanks for the thoughts

[@michael2244]

Hey i really enjoyed your video, and your commentary was very funny lol

[@imrantachmuradov5870]

Absolutely hilarious promotions, well done Leela

[@esji8736]

Why was that so fascinating!?

[@silviapabon1282]

I like Leela style... patience and discipline... space, materialistic... :) absolutely.

[@spencerm8271]

Great analysis. PS: It's pronounced "kie-ron" as "kite"

[@simovihinen875]

If I had to guess, Chiron is pronounced Kye-ron...

[@tamastorok9306]

12:20 "Why didn't Leela not takes a5? Because she started trolling Chiron by this moment :D

[@jackd42o]

Great stuff. As a casual chess player this blew my mind. "Mate in 30" ... and the trolling D:

[@basscomputer]

Does anyone else find Leela's "trolling" a bit spooky and somehow unsettling?

[@anarcho.pacifist]

No. It simply shows poor endgame understandings.

[@juggernaut4799]

Dammnnn! Leela being savage at the end!

[@djordjesimic9914]

C4 is explosive

[@RadishAcceptable]

LOL that end game was hilarious. My guess is that she evaluated those silly moves as "definitely winning" before even considering the faster win options. With the queen sac, for example, she probably knew that "rook and king vs king and all these pawns = definitely winning" and just picked it because she happened to evaluate that position as winning before evaluating any others.

[@I_am_Itay]

Lollll wow leela crushed him

[@manikenwarrior]

mind BLOWN

[@ew6230]

Ahhhhh shit. ALL HAIL our new robot overlords!