Leela crushes World No.1 Stockfish Engine giving pawn odds like Paul Morphy - End of Era approaches?

[@kingscrusher]

My Paul Morphy Course: kingscrusher.tv/paulmorphy My other chess courses: kingscrusher.tv/chesscourses Replayable game with indented variations: www.chessworld.net/chessclubs/ltpgnviewer32/ltpgnboard.asp?GameID=5023857&v=1z8yR2xveJw

[@bodeamir4484]

I guess Im kinda off topic but does anyone know of a good site to watch newly released movies online ?

[@huxleylian8853]

@Bode Amir Flixportal

[@jbthepianist]

Poor show from Leela. Should have underpromoted to a bishop at the end there for maximum lolz.

[@adrianzahariuc2784]

Ironically, black got mated by the f pawn.

[@fujiapple9675]

6:08 In this wild variation with Qe4+, the queen ends up in Siberia and it brings back fond memories of when if first got drawn to this channel's content. It was in another Leela Chess game against Komodo. Leela had the black pieces and played the French defense in the advance variation, and after a series of captures, Leela gives up the rook on a8 and puts Komodo's queen in Siberia, (which was eventually given up for a bishop), all the while Leela's dynamic tactics on the kingside left Komodo helpless.

[@thom1218]

Per SF google dev form in comment below - there's some initial speculation that the huge advantage SF sees due to the pawn odds setup makes it think its in a middle game with a large advantage, i.e. beyond the opening stage, and in those cases it uses a different type of "lazier" evaluation that doesn't catch the Queen trap (but has other strengths) Because normal games begin with almost no advantage for either side, this different style eval never causes problems in "normal" games because the small advantage difference is always small enough to cause SF to use the slower but more thorough evaluation.

[@joseraulcapablanca8564]

A fantastic odds win for Leela, some of the off the board mates were indeed Morphy like. Thanks KC keep up the good work

[@u.v.s.5583]

Just the usual business of crushing kings and eating fish and chips. Sure, great analysis as always, makes one appreciate this strange and stunning game even more!

[@joseraulcapablanca8564]

U.V. S. Too right, the analysis always gives new insight and ideas. He is also charming with his talk of fish and chips etc.

[@jessehammer123]

I think Leela is just playing opportunistically- making the best of whatever happens. No F-pawn? No problem, attack down the F-file! Can’t castle? Activate your king in the middlegame! No use trying to paper over weaknesses, just use them to your advantage. Brilliant.

[@ilyakalinin2660]

Leela also won another game against Stockfish in this opening (but with black)in CCCC

[@dtracers]

Leela knows pawns better than any other engine

[@ogezpb3927]

so, i was wondering what's the best reply to 1. e4, giving SF unlimited time, I came back to it after awhile and e6 was the answer. i've tried this several times and it's always 1. e4 / e6. so i'm curious/confused as to why this reply, even with the pawn odds, is questionable to you.

[@thom1218]

From SF google dev form: groups.google.com/forum/?fromgroups=#!topic/fishcooking/5fj7O1pe6lQ Nickolas: Houdini also made the same mistake in game 21. Interestingly, every version of Stockfish I tested instantly recognized Qh4?? as a blunder (and g3! as an equalizing defense) once the move had been made. It definitely seems like there was some overzealous pruning here. Crocogoat: SF probably makes weird moves early because it uses lazy evaluation. SF thinks it has enough of an advantage so it stops using normal eval. This completely compromises its opening. In normal games this would never be a problem. It's possible Houdini uses the same technique.

[@mytwitzworld]

Leela, got me on my knees

[@RedArremer]

Really the whole game went sour because of SF's blind spot issue with thinking 8...Qh4+ holds the advantage. It doesn't work and actually puts black in a bad position after the proper response of 9. g3! When you backslide the analysis after what happened, SF realizes Qh4 was a blunder, and finds that 8...c5 still holds 'some' advantage (and is a proper move human-wise when attacking the center). The blind spot on this move alone is so severe, that even the current dev builds of SF to this day STILL don't realize it's a blunder until around depth 37, which on my octal took almost 9 minutes to reach. Indeed the previous move of 7...Bb4 is also tied into this blind spot, but would take considerably longer for SF to realize the line doesn't work and should play Nf6 instead. I've asked about this issue and the counter I got was that it might be detrimental to try and patch this blind spot as it might make SF weaker in many other positions. I think it's the problem of tuning these engines to perform well in blitz, and the vast majority of the time, pruning to reach a deeper ply is beneficial. It just allows for these blind spots to get overlooked in rare cases.

[@bhgtree]

A bit off-topic but here is a interesting comparison of the world champions since Steinitz using SF: en.wikipedia.org/wiki/Comparison_of_top_chess_players_throughout_history#Markovian_Model (below the table is links to the researchers website).

[@jackmuller5478]

fyi its not "overcalculation" its overpruning, meaning it calculated less then necessary because it excluded or "pruned" the g3 move while leela did not.

[@kingscrusher]

Surely the "source issue" is over calculation though ?! - because trying to see further ahead increases the risks of missing something. The other engines didn't calculate as deeply and never really considered e6. They just played e5.

[@jackmuller5478]

how deep you calculate or not doesnt have anything to do with this. sf just pruned it out, so it didnt see g3. youre confusing causation with consequence i believe. pruning caused sf not to consider g3, and as a consequence, it was able to calculate deeper. its a matter of semantics, really, not a matter of fact. (this is based on my very limited knowledge concerning sf, so if im wrong, pls correct me.)

[@zlac]

@Jack Müller Maybe it is semantics? Chicken or the egg? I still agree with KC. Here's an example. Being broke and wanting money is a causation. Stealing stuff is a consequence of not having money. By your logic, Stealing money is a causation and having money is a consequence of stealing. While it's "technically true", it really is the other way around. :D In this case, the causation is that programmers want to make Stockfish calculate as deep as possible to "see further ahead". The only way to achieve this is by pruning more and as a consequence of this, SF pruned g3.

[@jackmuller5478]

@zlac the way you put it, i wouldnt disagree. because then, the desire of the devs to make stockfish be able to look deeper is the "source issue" as it caused the pruning which caused the missed g3 move, as you said. however, saying the amount of calculation itself is the issue, i think, is something i cant stand behind, as it didnt directly cause anything. it was just the result of the desire to make sf better.

[@u.v.s.5583]

Leela should use this opening more often against SF. What a gambit, what a plan, she knew it all the way!

[@Th3Wick3dOn3YT]

I think leela was putting her king to do some ballet dance in the endgame maybe as a show of how happy she was feeling.

[@rakso472]

32:03 Over the board you could use a bishop instead :D

[@morsmortis4114]

Thanks! Great Video!

[@modolief]

"The business of crushing kings" ... Leela chess style 😎

[@zebreuuu]

there is a character in the anime no game no life (never loses any game LOL) whom says the only way to win a chess engine is to play imperfect moves 'cause the engine will always play the perfect move, here is just that exemplified.. :D

[@sovaz1997]

All engines doesn't made best moves

[@zebreuuu]

sure dude, the character says it, not me.

[@firstlast1092]

I love that anime its pretty well done

[@zebreuuu]

well, in fact what i'm saying is that the case here reminds me the statement of the character, once stockfish assumes best play and reply with best play every move.

[@andremendoncabastos5103]

I almost sure that even right now, humans(GM's) make sometimes a better moves than any engine! The problem as like Nakamura said is that "engine plays a like perfect chess". I think him doesn't say that is perfect in a sense of no better possibilities, but because for us(humans), to play like a machine is very heavy all the time, take too much energy, time, psycologial pressure and we can't always do the best move and plan. But I think it was good idea if good chess player can think a move and a engine elect the best move! Perhaps, I believe that, some moves from humans will be better that any engine, but just some moves...

[@smashu2]

If you let Stockfish think much longer maybe hours it will reject this e6 Qh4+ variation unless there is a pruning problem. Or if you show SF the line than you take back it start to reject the line because it store it's analysis.

[@callmewisdom]

what does he call the pawn on f6? Foam pawn? Fawn pawn? Form pawn?

[@morsmortis4114]

Thorn pawn. it's an Accent " fin' "

[@RadicalCaveman]

It's a Cockney-esque pronunciation of "thorn pawn."

[@EGarrett01]

Porn Thawn.

[@megaman4686]

9/10 didn’t promote to bishop

[@RadicalCaveman]

I would have totally promoted to a Bishop! But then, I'm a mere hooman.

[@-ace-52]

Great analysis as usual and great game by Leela, quieting the haters

[@GnuReligion]

You cannot use and upside-down rook as a 2nd queen? Well, now I know! Seems a matter of very short time before optimized GPU / FPGA hardware is made specifically for Lc0-like engines.

[@brendanward2991]

GnuReligion Leela should have promoted to a bishop. The dark-square bishop was available.

[@RadicalCaveman]

If Stockfish had calculated still more, it would have seen the Queen-trap and played e5. Less or much more is more than more!

[@EGarrett01]

We do have to do something about that "retirement community" background that Tryfon uses.

[@n8style]

Anyone know why Leela has to be "lead" to the starting position?

[@Envengerx]

She uses last 8 moves for calculations

[@philipstevenson5166]

It will come down to how many man hours go into upgrading stockfish. Chess is quite limited, but it seems still that Leela will win because it can be improved with less human effort. In more complicated fields, humans and their representatives are likely to lose more quickly. With better hardware, even more so. Get out of the stock market while you can.

[@sanjaygoyal5395]

One day i will marry leela

[@joelperlstein1255]

Sanjay Goyal ..fook no ill wrek u mate i swear on me mum ;p

[@mustafaunal1834]

Go to the back, I'm in line! ;)

[@bhgtree]

Leela will kiss Komodo during one of their games and he'll turn into a handsome prince.....

[@ucmanhchessonline288]

Why you not live stream ?

[@kingscrusher]

I do stream - later tonight and also usually on Sundays at minimum

[@guepardiez]

No f8B#, I'm disappointed.

[@megagagnon1]

I had the same thought. Why not checkmate by underpromotion?

[@saunadive7594]

gotta love Siberia

[@dr.mikeybee]

I'm continually amazed by your analyses.

[@herzwatithink9289]

KC is always pleasingly erudite

[@j.hanleysmith8333]

How did stockfish not find all of the bad qe4 variations? That's weird

[@TheGreatSnoozer]

it did! just one ply too late

[@miladibrahim1068]

Are you sure Stockfish had no bug and it was full strength??

[@kingscrusher]

Stockfish is doing well overall so far - leading but with now two losses to Leela

[@rakso472]

I guess you need to calculate far enough to see the downsides of an idea, or not be able to calculate far enough to find the idea in the first place. Stockfish found the idea but not the downsides.

[@kingscrusher]

Yes that is what I was trying to say with "over-calculation" - sometimes "simple chess" is better - more principled, and if a calculation is inaccurate, it doesn't destroy one's position as easily.

[@balthazarbeutelwolf9097]

Perhaps in this case not over-calculation, but over-pruning the search tree. To be able to go deeper in analysis SF cuts out the computation of lines that are "clearly losing anyway". Unfortunately every now and then such a "clearly losing" line has some venom, as in this case the move g3.

[@axelbarrientos8906]

Leela > Bowsette

[@Stockfish1511]

Srsly th hype is getting of the chart. Stockfish dominated the field with like 200 games unbeaten. Anihilated houdini that came second, beat leela in classical and lost none. Leela beats sf in some handicap bonus game twice, in which one was atleast questionable as stockfish was making blunders which is unheard of and everyone lose their minds. As for now sf is surely better

[@dannygjk]

"stockfish was making blunders which is unheard of" Please clarify that for me.

[@Stockfish1511]

stockfish does not make terrible blunders like qh3+ Besides stockfish beaten leela couple of mins ago. You leela fangirls need to calm down.

[@RobBCactive]

Morphy would have sac-ed a load more pieces! ;)

[@victorvincent1247]

Take that Stockfish!!!

[@azem9338]

Leela is still weaker than Stockfish 9

[@580player]

Something is fishy about this stockfish. Stockfish plays better than this.

[@sonaruo]

Well positional playing trash material playing. All the games leela traps pieces of opponent and even with 1 piece down the opponent can not use that advantage. Put the opponent in positional disaster and then goes for the kill and everything fell apart.

[@rageagainstthemachineragea2497]

😃👍

[@matthewviramontes3131]

I'm not impressed with these wins where some of the moves are forced, or when there's pieces missing, or anything like that. In my mind, the only games that count are when the engines start from initial positions and play which moves that they choose.

[@kingscrusher]

Engines usually use human opening books

[@matthewviramontes3131]

@@kingscrusher Yes, but even with an opening book, it has a plethora of options to choose from, not just one forced opening move. I do prefer though the games that don't allow opening book, so the engine is truly playing it's own game.

[@Envengerx]

Opening and ending books are not true ai according to me.

[@THEBOYDALEK]

I agree here why just not let them play what they choose, we don't do this with GM's lol. Alternatively we have to play it with colours reversed??

[@edwardsolomon1951]

Because then every game would be the exact same between two distinct engines. All chess engines are Turing Machines. Two Turing Machines playing chess against each other, Machine One having set of instructions X and Machine Two having set of Instructions Y, will produce one unique game for X vs Y (X = white, Y = black) and one unique game for Y vs X (Y = white, X = black). Unless of course, human intervention alters the openings in a manner that doesn't give an advantage to either side.