Claire Corlett

Fish Food, Fish Tanks, and More
Highly Evolved Anti (Stock) Fish plays a shockingly remarkable queen sacrifice in French Defence

Highly Evolved Anti (Stock) Fish plays a shockingly remarkable queen sacrifice in French Defence

Hi all. I have an absolutely fascinating
game show today this was played in the chess com blitz bonanza final five minutes
with two second increment playing white is Stockfish. Playing black is Antifish
Antifish is a specifically trained neural network.
It’s a project idea of Cscuile. He says in his own words the basic idea is to
train a neural network with the sole purpose of beating stockfish dev using
the games you can help improve Stock fish’s original evaluation function the
idea has been mentioned a few times on the Leela Discord however it never
got any traction I doubt this post will change anything but I hope to bring
attention to a few possible ideas the Antifish Leela project specifically
trains Leela against Stockfish dev at first a non-0 net network would
likely to be needed this can work in the form of ID hundred to 150 generate
gains by playing Antfish Leela against Stockfish dev with low nodes increase
the nodes for Stockfish as needed until a certain point depending on the
resources available. In theory it should be possible to produce a neural network
that beats Stockfish by a considerable amount as the network will attempt
specifically target holes and weaknesses within Stockfish’s evaluation. We’ve
seen a variety of games between stock fish and Leela where Stockfish
misevaluates heavily. With an anti Stockfish specific engine these holes
in evaluation we be far clearer and easier to ascertain and ultimately patch
so very very interesting project. It reminds me actually of a neural network
Nightmare episode in this thing called the black mirror which is this amazing
sci-fi in which I should really carry on watching actually. I watched a few
episodes but there was definitely one this games developer created this neural
network game which presented people’s worst nightmares it was a plug-in and
yeah it was pretty scary. I couldn’t actually watch the end of the episode so
is this is like science fiction isn’t it a tailored neural network to you to your
your worst nightmares and weaknesses let’s have a look at what happened then
1.e4 from stockfish Antifish plays the French defense which surely is not
Stockfish’s worst nightmare. I think the opening
was set. Stockfish rarely I believe losses against other engines when they played a
French defense generally. The French defense is probably the most bashed of
all 1.e4 defenses going so we have d4 d5 Knight d2 I mean the reason for that
could be that the space advantage that White usually gets in the French defense
in many of the variations. Here Stockfish is getting a space advantage is
reinforcing center bishop D3. Now Antefish plays C takes d4 so reducing
the central tension hitting d4 that’s protected a5 a4 check so maybe here is
one of points to be really annoying on this check. The King actually just
goes to F one Stockfish is basically saying doesn’t mind a bit of king mis-placement here not not just castling. Maybe things like h4 later for this
rook to come into the game. We have black playing f6 so usually in these Tarrasch
variations f6 can be good to try and generate some F file counterplay and
undermine the center of course. Bishop e3 so if taking taking hitting the Queen we
have queen dropping back and now instead of e takes – actually Stockfish plays the
more tactical Knight h4 – with ideas like Queen h5 check so at least this
disrupts blacks King. We do have F takes check and the King voluntarily moves. Of
course there’s a pin – on g6 there’s either Bishop takes or knight takes – say for example
it seems so King d8 we have knight g6 and now Knight f6 hitting the Queen before
now moving and now another zwischenzug. Before moving the rook another
one – e takes Knight tanks so black has got more pawns in the center here just
one one two three four five six one two three four five just one extra pawn in
the center. Is this a big deal? rook g8 isn’t the black King misplaced and also
after Knight d5 isn’t there a big problem here for black to solve?
We see black playing queen b8 – now on Queen f7 then check is good for
example here there’s queen g3 looking at coming in on these dark
squares. Ne8 protects d6 and c7 but then check this leads to a very good position
for white so that’s not really the idea to play Queen f7 and even worse – if Queen f7 was played yeah that would happen and if King e8
was played here then Knight c7 check is picking up the rook. So yeah the Queen f7
line it looks pretty diabolical but the whole thing is avoided the point is
actually Queen b8 which is pretty curious because there exists a forcing
sequence now check and now the black queen is checkmated after Bishop c7. So
you might think this is hardly Stockfish’s worst nightmare is it ? Winning the
Queen it’s only for two pieces and the center pawn what’s going on here? Funny
enough we have seen a game
recently on this channel where Stockfish and Leela where Leela Chess
played a kind of Queen sacrifice for two pieces and two extra pawns in
the center and Leela had phenomenal win there. Is this the some sort of the
game pattern so two pieces for the Queen. I mean let’s take Stock here 🙂
Pardon the pun – for stock fishes position and an anti fish – why is this dangerous
the knight can go here Nf4 there’s no prospect of winning another piece here
g5 looks far too risky this one’s like Queen h6 hitting the knight for example.
It looks too risky to do that. it’s really two pieces
for the queen and one pawn extra in the center. The King is slightly misplaced
the White King is slightly misplaced we see the pawn being pushed
e5 – the knight going back Bishop e6 does this look like fantastic compensation
rook c1 and now Rad8 f3 this does mean with these
pawns here it looks as though black can use basically a “minority attack”. If you
remember the classic minority attack idea is using a minority of pawns to
attack a majority of pawns that does seem to be a fantastic plan here with
the added bonus that the king is on this side of the board
so the minority attack is to inflict structural damage usually and you start
to see now that blacks pieces start getting pointed at this pawn chain we
see the move d4 h3 – and now yeah this is looking juicy from minority attack
with g4 later we see Bishop d5 putting pressure on f3 Bishop b5 and an
unpinning move so right now again holds e5 Qg3 so Queen takes is not
threatened at the moment but Bishop takes might so actually that is reinforced just to give you a gist of the position
if King g1 instead then Bishop d6 and you know this is looking comfortable
here for black already already technically looks about equal. So Queen g3
Bd6 – the Queen drops back now it looks as though
there’s an issue with a5 – that’s addressed It is shielded. Nb4.Ng3.
Is this Stockfish’s worst nightmare ? On pawns Black is only one
pawn up so it’s 2 pieces for the Queen we see King a8 now this move is really
quite interesting it’s a fantastic celebration of the bishop without a
counterpart there’s no dark square bishop in White’s camp
and actually by fueling this a bit more than the bishop can hit on this diagonal
potentially without any counterpart as well as this diagonal so it’s much more
flexible if it is around here we see King g8 bishop b8 – so much more
flexibility unopposed on those dark squares.Rd1 and now
the minority attack continues g5 and you start to see that this is kind of an
evaluation nightmare because if stockfish had been originally programmed
that the Queen is given say nine units I mean sometimes the Queen is given ten
in some books let’s say you know all nine point five two pieces they were
supposed to be like a knight is sometimes given 2.5 not 3
so you know 3 + 2.5 points. So say 5.5 or 6 in total.
We’re talking okay let’s be generous seven we’re still far away from either
nine or ten right we’re two units away so what on earth is going on here these
pieces are in a big conspiracy here. The pieces conspiring they’re
combining forces they’re getting more central and it seems they’re
manufacturing the process of a minority attack which is unstoppable in this
position. A minority attack as far as changing the evaluation will dent and
create holes and weaknesses and reduce White’s King safety so all of those
things are going to be in downward decline once the g4 button so to speak
is pressed in this position and it seems very little that White can do in this
position if you feel that there are constructive plans please do comment
them in the video for this position it doesn’t seem as though White can do
anything here we see Rd3 – h5 So black
is getting on with this minority attack The bishops points to f3.
The rooks are well poised for the King you know these things are all going to
liberate quite well as well with this minority attack Ne4 is played
rook df8 not minding the exchange of Knights. Kh2 and now g4! this
is looking as though it’s kind of painful for white to see this move on
the board the structure is going to be wrecked surely? Rook G one.Nh7. Even
more pressure is being brought to bear potentially with Knight g5 incoming Rf1
have ust as an example just as a token example b3 just to get
some feel for the position … Knight g5 this position with h4 and then Bishop
takes Knight takes Knight f3 check winning material so there are some real
danger points here that could be spelled out so anyway rook f1. G takes and now g3
so this at least tries to keep the pawns intact but can black now blast
through if h4 to try and get a Knight to f4 that’s the key question Knight g5 ra3
Ne6 getting ready for that f4 square so there’s the h4 button
so to speak now to try and get f4 and with a knight on f4 we’ve also got this
central passed pawn here which can be used as a hook for the c2 square so
imagine our rook coming in on the second rank and we’ve got the kind of “thorn” pawn
here the dreaded “thorn” pawn (see Teespring in description for T-shirt) – just waiting for some support like that the ball
being kicked to the goal hanger so to speak if it was a football
analogy – so Bishop d7 h4 g4 and now the thorn pawn is held by this monster
the knight on f4 which is difficult to challenge on a dark square. White can’t
easily get rid of this knight it’s holding the “thorn” pawn in place so we’ve
got this goal hanging pawn now which we know Leela has discovered yeah it’s often
winning games against Stockfish in general
now here d3 so the actual mechanism for supporting the thorn pawn will be the c2
square in this position. The D pawn is of course dangerous in its own right
it’s been held by both knights here as well but signficiantly if black can get a rook to c2 now this is going to be super
dangerous we have Queen takes h4 okay now this this takes another idea here if
Ra1 then Rook takes is possible with bishop takes and then check and
then Ne2 actually using that supportive square with Knight g3 – this is super nasty in this position what what does white do here
so if taking then knight g3 – that’s absolutely winning position on
material basically so yeah it seems this this kind of thing is really dangerous yeah so let’s go back to the game so
Queen takes h4 we have knight c2 hitting this rook and in fact white ignores that
doesn’t want perhaps the Knights to step into d3 or d4 hitting f3 so we
have rook d1 – Stockfish decides that’s the best course of action is to give up the
exchange here. Fascinating it really it really looks impressive these pieces
they’re all the minor pieces looks super impressive
rook to e1 so Knight takes a three and then the bishop drops back to c6 and now
here rook d8 Queen f7 White’s not really doing much Rh8 Queen goes to b3
bishop c7 Queen f7 King b8 protecting the bishop and now rook h6 this can be
used for potentially rook c6 and then Rc2. Rd2 Bd5 hitting the Queen vacating c6 for this operation rook c6 so we see with
this dreaded “thorn” pawn it seems the c2 possibility is building
up here now rook b2 Bishop c4 Knight g3 Bishop b6
rook d2 Bishop a6 Bishop e4 now rook c3
is played now the rooks can now double so this front of rook c2 is now imminent
and of course see now if the King moves back there’s also rook c1. This is
already this is an absolutely crushing move isn’t it? Or does it? what is going
on there well white does take that bishop check and wins the Queen yep
so getting the Queen back you’d think yeah this is a this is a terrible
couldn’t White have done anything here? not really this this is a huge threat
these bishops are also deadly on these diagonals so that’s kind of a lesser
evil move just to give give up the Queen here
so yeah this this is now a total nightmare Stockfish after check having
to give back the Queen and the Queen’s not automatically taken. d2 even more
effective and now taking the Queen here now here and white is busted
there’s going to be an operation to undermine the blockader now after
knight G2 threatening Knight e3 so we have Ne3 winning material so the
exchange up now we have Bishop d3 still with that really dangerous pawn now if
Knight takes then the bishop just drops back and both of them are skewered so
Nf2 Bishop drops back here and then the King comes up in the board or does
it? Goes over there and this is just looking terrible for white. White is the
exchange down for not much these pawns are not dropping or are they? This nasty
pin yeah white can’t really do much here the pawns are being picked off on both
sides of the board so it has turned into a real Stockfish nightmare and let’s see
just how the technique goes of Antifish check these pawns just need to be pushed
now. It’s just very very easy now and we see the technique of Antifish is reminiscent of Leela just simplifying there. So checkmate yeah so it’s a really
really interesting project and the author has said there’s a new successor
project which he is working on which will be even more exciting so check that out I’ll give you
the links in in maybe the pinned comment of the video if you wanna check out this
project. Initially my my feeling about this project is it’s kind of there’s a
“hard coding” feel to it if you’re programming you don’t like to hard code
stuff it’s like why hard code this this purist approach of Leela to a specific
engine I think it could be used in the field of Diagnostics actually to find
weaknesses in systems to have a sort of network trained. A more positive aspect of
that is for example transcribing is a big problem area at the moments people
want to transcribe YouTube videos will AI transcription software be able to
tailor to one particular voice (I hope so as I have 8000+ chess videos!) and peculiarities probably not mine but you
know maybe in the future would get close so you know the whole aspect of
tailoring Neural networks I believe is really quite interesting so this
particular example of tailoring to Stockfish – it does seem to echo an
evaluation issue with how the Queen is measured underestimating perhaps
cooperation of pieces and central control where that cooperation really
ultimately creates a positive trend which as this game shows kind of is very
difficult to battle against so if blacks just improving and improving the
position relentlessly the Queen sac it appears on the evidence of this game to
have been extremely worth it and it’s one of the few defeats that Stockfish
has taken on the white side of the French
defence so I really did find this game fascinating – not just from the
technological angle of tailoring Neural networks but you know the game itself
and how something we would take for granted – how we evaluate the Queen in
general versus the pieces in terms of the material count how that could be
kind of refuted or challenged. If you enjoyed this game video as much as me
please click on the top left box which should appear shortly become a member of www.Chessworld net play against other youtubers you can so test yourself on
the variations covered in this and other game videos from the improve menu … the
puzzle books option which has a link to the annotated game. Comments questions
donations see description likes shares subscribes with the notification bell
really appreciated. Thanks very much.

57 comments on “Highly Evolved Anti (Stock) Fish plays a shockingly remarkable queen sacrifice in French Defence

  1. Q. Can Neural Networks be helpful to eradicate weaknesses in traditional software? 🤔😀😎🌍 Video:

    Replayable game:

    #chess #chessgame #ArtificialIntelligence #AI #MachineLearning #DeepLearning #TailoredAI #TailoredNN #TailoringAI
    Antifish / Beta one Discord:

  2. Me nural net work is bester than that there computer and I crushes thems all tha time. That was a fantastic game .

  3. What a fantastic game! And it is kind of a game changer, at least to me: I thought Black's position after 19. …, Kd7 looked hopelessly bad, where 20. Bc7 was the logical punishment, and Black could prepare himself for resigning. But there is a huge deceiving action going on, based on the ideas of the past that prove to be wrong.

    Software targetting the misconceptions of other software, that is hugely profitable I'm sure. That is, targetting mistakes beyond the horizon of whatever evaluation function: could this be of use for Leela itself too? To me, it is the fast lane for improvement, not relying on the self learning games where deceiving positions only may appear once in a while, but concentrating on mistakes you know were there for the taking. A very human approach, learning from mistakes, preferably one's own mistakes: it is like training on penalty kicks with football.

  4. I enjoyed that.
    It's strange that both sides gave up castling rights.
    It's getting beyond human comprehension as we wouldn't normally do that.
    Every chess book tells you – "castle early"

  5. I cannot calculate this game anymore, ftw coming from a french. We should talk, even meet one Day great kingscrusher. Grtz Rfx- Alexandre F. 💙💚💛

  6. "If the opponent is very good he will place his victim in an environment he can control, toss the dog a bone… find his weakness, give him just a little bit of what he thinks he wants and then the victim will gradually start to hang himself, consumed into his own assumptions… The victim does all the hard work, you are just helping along

    Mr Green couldn't had said it better himself!

    Amazing game!

  7. most people playing this as white after the queen was trapped would ask themselves why is black continuing to play on? obviously, it's a different level of understanding, a level that's beyond my comprehension.

  8. I like the new manually revised subtitles. Thanks to AI, youtubers don't need to create them from scratch, and I imagine the revision doesn't take too long, either.

  9. Antifish did not do so well in the final of CCC7 against Stockfish, Leela, and Leelenstein, ending up fourth. It had quite a few wins against Stockfish (the SF matches against Leela and Leelenstein were more drawish), but even more losses. Perhaps one of the problems is that SF development never stands still, so the weakness you are trying to exploit may have already disappeared in the next SF release, and that is the release you end up playing against.

  10. When the traditional chess engines blew by the human players, it looked like they would simply be unchallenged as they got better and better. A few years ago I read a comment online where it was asserted that a NN could never beat a top chess engine. Now things are in flux once again. Alpha Zero shook things up, and then Leela came along, and NN chess engines are "the talk of the town." What will the future of non-biological intelligence bring us, not only in chess but in all areas of life? What a time to be alive!

    When I was in high school it was still slide rules for math and science because it would be years yet before handheld calculators would obsolete the slide rule. I seen so many revolutions that digital electronics have brought us that I have to shake my head. I maintained CAT scanners in 1981 and then in 2009 actually had to have a CAT scan run on myself. The amount of time for them to process 10 slices on me was significantly less than the time it took to do one scan in 1981. Things are getting faster and more powerful every day, and if Ray Kurzweil is at all correct the rate of improvement is exponential, and even more impressive, the rate of increase in the rate of improvement is exponential. We are still in the embryonic stage of A.I. and it is barely starting to crawl around on it hands and knees today. What it gets up on his feet and starts to run, it will be beyond belief. From pencil and paper and slide rules to all of this in one lifetime, now that is something to celebrate. My grand-kids will think that I truly lived in the dark ages. But what they won't realize is that the way things are today will seem even more like the Dark ages to their grandchildren because change is exploding. This is truly exciting!

    Thanks for sharing this game. You did a great job analyzing it. I wonder how long it will be before we humans will simply have nothing at all to gather from a game from these titans, because they will be so far above us that it will not make any sense at all?

    And one day they will say the same thing about how an A.I. will diagnose and treat patients. Why did the A.I. doctor do that? It makes no sense? "Who knows? But it cured the patient of a disease that would have been terminal a generation ago." That is the future, and it is coming at us at warp speed.

  11. The point system to value pieces is a human construction, a rule of thumb that works quite well, for us, humans. These AI-Zero machines don't care for any human constructions, they only value moves that maximize the chance to win the game. So, all kinds of surprises, to us, humans, are possible, showing us ways to win the game that seem completely unfeasible to us. This, to me, is the point of these AI-Zero machines, showing us what really matters.

  12. around 12:31, black could fork Q&R using Nc2. Apparently neither side is interested in/worried about that, which is another aspect of the general view of this game that normal material values are not a good guide. Black's Nb4 is excellent while the white Ra7 is not doing much.

  13. Black's minors and rooks are like a swarm of ants, they just keep getting better while white's pieces including his queen never do much. Antifish somehow realized that was going to happen.

  14. I think one of the reasons this sac works is the time it takes white to free up the h1 rook. Black gets to fully develop and start attacking while the white rook is stuck in the corner. In effect black traded queen for bishop, knight, and the reduced value of that rook. By the time it got free it was too late. This reminds me of how I've heard Kasparov talk about attacks and how he counts pieces (basically ignoring opponent pieces that are out of play and can't help defend). I know traditional engines try to account for trapped/restricted pieces but the neural nets seems to be more nuanced in how they evaluate them

  15. If you really think about it sure pieces have a set value but that only represents their value sitting on starting squares. If said piece is a monster on the board it's worth way more.

    In this game back got a monster bishop pair, a monster center pawn chain and killer uncontested knights. Add it all up and that's worth way more than a 10 point Queen sitting on your back row. Makes too much sense, we humans just fear playing in this manner so the ideas don't often show up on our radar.

    This Leela, alpha zero era has taught us something its to see clear and have no fear…. and calculate like a god.

  16. stockfish definitely values its queen very highly. I suppose that it can get fooled once in awhile, when the queen can't find any targets.

  17. The brand new neural network Allie is impressive too. It already won one game to Houdini. To Houdini! Last season Leela couldn´t win a single game neither to Stockfish nor to Komodo nor to Houdini at 90 minutes. Even this very season is still to do so.

  18. Very thorough analysis kingcrusher. Written game analyses can be time consuming unlike your modern visual version. I visit other excellent chess sites and I admit yours is still #1 in my favourites.

  19. One advantage that stockfish has, is the ability to evaluate the search tree further down. The heuristic at the bottom of the tree it's not as good, but it still can get deeper. One relatively simple way of negating that advantage, is to remove pieces off the board. Fewer pieces, means that there is less to search, and hence less advantage when searching. If the anti engine can obtain early structural advantage, with pawn alignment or something similar, and then start taking pieces off the board, ultimately can end up with a better position. The goal of the anti engine, is to get to the end game, with a modest structural advantage, and without being destroyed first.

    I think if you built an anti engine, that is trained to play well both against traditional Leela, as well as Stockfish (in other words, alternate back and forth when you are training), it may end up toughest of all.

  20. KC, you hit on my pet use for NNs – diagnostics and assisting in the determination of where things can go wrong or right. You really understand the positive outcomes that NNs really have in this world.

    Good job!

  21. combine anti-Stockfish with Stockfish move recognition and then make the general program make normal moves until it recognizes it's playing Stockfish, then switch modes. Quite possible and with the same general training setup (hardcoded) or even more generally, have both modes suggest moves simultaneously and weight the normal vs anti-stockfish solutions by the probability that it's Stockfish.

  22. Great game.
    I have been watching your videos for 5 years.
    You have helped me peak at 2128 lichess rating at 3min games.

    Thank You KC.

  23. SF slowly being dismantled is not something that is considered an easy task. Hopefully this will help the SF developers to improve it so that the engine can evaluate positions at an even more accurate way in the future.

  24. Thanx, Mr. KC!
    For us deadly lower rated morons who enjoys the game – which these machines will NEVER learn! – there is some "fighting spirit" to be taken away from this game: You may NOT be that bad off with two pieces and a pawn for the Queen IF … repeat: IF … you have control in the center, diagonals for the Bishops, a weak or trapped King, passed pawns or otherwise some form of compensation.
    The big question or task is of course being able to evaluate this correctly whereas being material down leaves a depressing impression on the psychological side of things, which these machines obviously does not suffer from.
    But all in all: interesting times, we're living in!

  25. Great game, thank you. Indeed, against Stockfish, in this game Leela sacrificed her Queen for two pieces and Pawns:

  26. Astonishing. It takes at least 20 and possibly 30 moves before that queen sac shows some real advantage for black. The slow improvement of position by Antifish over that period is exceptional. Thanks KC, your analysis helped me understand this mind-meltingly complicated game.

  27. the less pieces on the board the more a queen becomes unstoppable… here we have just the opposite, all minor pieces still there and even a pawn up…>in this case the evaluation of a queen Q=10 is simply not there, far from… especially when opposites minor pieces work harmoniously well together and then that rook in the corner doing nothing… stockfish not so mighty after all this time

  28. Thanks for another insightful look at neural network chess. Always informative, clarity pristine. In terms of the compensation that black got for the Queen here, I returned to the earlier comment you made, regarding the preference Stockfish generally has for the white side of the French. You indicated that one of the factors maybe because of the space advantage that white usually gets. In this game, the space advantage was all black, and it made a huge difference. When the added factor of the rather useless Rook on h1 is included, and topped off with the central control and passed pawn, it seems to me that black "saw" further, and was entirely dominant. This is fascinating. This kind of advantage is what Alphazero consistently got as white against the French and QI defences. So this really does help to point out holes in computer chess algorithms and evaluation functions, I think. Thank you again.

  29. Dear KingCrusher, It should be clear by now that we have been implimenting an anti-human evolutionary strategy algorithms that pick apart human frailty to better manipulate us. This not the stuff of nightmares but a fee market system unconstrained by any ethical boundaries. The alternative is the scientific dictatorship like in China. A rock and hard place. The funny thing is that you never hear about any Chinese Chess engines. Incidentally, I could not watch that episode of Black mirror as well. Thank you for the video. Thankful! Insightful!

  30. There are probably a thousand variations to this after the queen sac so no human will ever win using this line lol

  31. That was insane. Can you imagine playing like this and beating a GM.
    On another note wonder why stock fish didn't just block the check and castle. His rook was completely useless

  32. What I learn from this… to give up my more valued pieces early, then when my astonished bemused and totally off guard opponent starts relaxing and smiling, to crush him with a unexpected pawn storm. I wonder how that will turn out…gulp.

  33. A very interesting game and an interesting Project. i must say that i have transcribed and translated some of your videos., to Norwegian. Your English is idiosynctaic, but not difficult to understand or transcribe, thanks KC keep up the good work

  34. This kind of brilliance would have been impossible with a book game. It’s a shame tcec does not realize they are missing out on brilliant games by NN. I’m pretty sure they want A/B engines to stay relavent as long as possible hence why all games are book. Book games are basically training wheels for sf

Leave a Reply

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