Это видео недоступно.
Сожалеем об этом.
Stockfish Can't Evaluate This Position But YOU CAN
HTML-код
- Опубликовано: 3 июл 2022
- FEN:
q7/8/2p5/B2p2pp/5pp1/2N3k1/6P1/7K w - - 0 1
The chess set behind me:
The Pieces: bit.ly/3zBJzlO
The Board: bit.ly/3b3B8pb
✅ My Courses! ✅
chessvibescour...
👕 New Shirts! 👕
Chessvibes.shop
☑️ Play Chess Here:
chess.com?ref_id=5885046
☑️ Support Nelson Here:
paypal.me/Ches...
Links are affiliate links and help support the Chess Vibes channel via a commission.
You should point out that if the queen ever takes the bishop on g1, the white king can run out and take both of the passed pawns, at which point _Black_ will be stalemated.
Thats the finny way of ending this game
Oh thank you now I understand
Was surprised he didn’t mention that, it seems a logical line to explore.
What if the queen makes it to f2 when the bishop is on g1?
@@jasonhalverson1180 should still be a stalemate...white would be forced to take with the bishop, which means either black takes with the pawn and white goes g3, or they ignore the take-back and let the white king out to capture the passed pawns to stalemate the other way.
I just played against Stockfish 8 on Lichess with white in this position and apparently I blundered 15 times and Stockfish as well blundered 17 times. Lol
Did you consider the option of not blundering that much?
@@C0urne is that some kind of unobvious sarcasm or are you real? Lol
@@chesscrush I'll take the sarcasm route. Well played!
@@C0urne In case you were serious, the position in the videos is a draw and white just has to move the bishop back and forth in the end position so I did that with the 50 move rule and because the Engine thinks it's winning for white it showed that black was +20 a few times which isn't true thus it thinks me and Stockfish blundered even though both played perfectly ;)
@@chesscrush C0urne was very obviously being sarcastic.
This is the well-known horizon effect that has plagued chess engines since the very beginning. When they reach their search depth, they use their static evaluation function, and all position dynamics are lost. That explains Fritz’s score very well. NNUE’s (efficiently updatable neural network) AI is clearly a cut above however!
There is no difficulty to write a SW that will be able to make the same considerations human do, and conclude this one as a draw. The point is that such SW will lose to NN in 99.99% of "normal" cases, and it will be much more expensive to develop.
Yep. As soon as an engine search deepth would go beyond 50 moves it can see the 0.0 (guess that will take a decade or two). Or when the table bases will include such special positions (probably even longer time frame).
What a computer cannot see is, that there are too many possible positions for Queen and Bishop to not get in a threefold repetition (time to add that to engines, eh).
What I'm curios about is, what would be the rating of older engines like Fritz 6.
@@dtfhhn Seeing just beyond 50 moves is not enough because the black has two pawns it can sacrifice to delay the inevitable for another 100 moves.
@@yigalzemach9083 Absolutely true. Therein lies one of the current grand challenges in AI: combining human-like experiential learning and pattern-based heuristics, and human-like logical deduction.
Stockfish actually wasn't using NNUE in the analysis. If you look closely, you can see it is Stockfish 15, which actually has no NNUE! If you go into the analysis settings and click on the engine name, you can use Stockfish 15 NNUE, which actually uses NNUE. This means that even without NNUE, Stockfish 15 was able to find the draw moderately quickly. If you toggle on NNUE, it finds the draw much much quicker.
Not only that, but check out the depth value. It caps at 20 on every analysis. That means that not only was Stockfish 15 not using NNUE, it was capped to a max depth of 20. And yet still it was able to find the draw in a matter of minutes. Stockfish 15 is just too good!!
I think this stuff happens often because if it can't go to higher depth, it will starts counting material.
True
I wonder if it’s kinda true to an extent, cause white can always throw and black can only throw by putting their queen on the bishops diagonal. So like in game it’d be favorable to be black in that position imo. Not saying that’s how the engine is analyzing the position, but it’s a thought.
@@jamesmoniz5263 The engine repeats the same Bishop move as white but still says it’s losing. For noobs I guess that black has a better chance to win but other than that idk why it says losing.
Some 40 years ago I pointed out that if you are *purely objective*, there is nothing like "better position" or "better move" (almost). Nobody understood my claim (then), but now with the engines I believe most of playes would agree. Objectively, purely objectively, there are exactly 3 kinds of position (white wins, draw, black wins) - and nothing in between. All you need is a perfect machine to analyze the position brute-force (all possible routes) to decide. Mate-in-1 position and mate-in-200 postion are objectively the same (you are not getting any extra for winning faster). As a consequence, there are, being purely-objective, only 3 kinds of a move: (regular - one that turn a win to win or draw to draw; bad - turns win to draw or draw to lose; and disaster - turns winnig position into a losing one. Purely-objective, nothing more than that.
So what does it say about our usual judegement of "slightly better position"? What does +0.17 evaluation by the machine mean? We are telling something when we say that, don't we? Yes, we do. But that "something" is not purely objective. It has some subjective parts in it. It says, if you are playing agaist a human, and you consider yourself as imperfect (and that what we all are), there are greater chances that your opponent will miss in posiotion A than in position B. Fine, but surprising (chess is subjective?! yes it is).
It also says what good players already knew intuitively a long time ago: It does matter who your opponent is. A good move against one GM may be cosidered less good against another. Surely when your opponent is a SW. And you can't say this is untrue because an objective analysis says something else, since there is no objective distinction as long as the 2 moves under consideration are "regular" as defined above.
To understand my point, consider a position of white having Q + 5 pawns, black 4 pawns (and nothing else). Now white makes a blunder that loses a full Q, but he is still winnig. Objectively, this is just a regular move. But everyone will call it a blunder, of course. They are right! However, the evaluation of this move as a blunder is not purely objective, it has a subjective element tightly coupled to it. Yes, very funny indeed.
No, they won't. In older and more amateur ai, sure. In more modern and stronger AI, NNUE is used instead.
This kind of fortress positions is a common problem for chess engines. There's a huge material advantage for one side, but it can't search deep enough to figure out that pointless shuffling can only lead to a 3-fold repetition. Looks like the latest Stockfish 15 handles it really well. I just confimed it with a standalone Stockfish 14 and it can't find a draw even after few minutes of running, while Stockfish 15 finds it just in one second
My guess guess would be that Stockfish 15 detects that only 2 pieces ever move, and just assumes that it's a draw.
Could be fun to build a problem to test this: Can someone find a setup, where 2 pieces have to travel 15+ moves, before the position "unlocks"? And if so, what is Stockfish 15's evaluation of the initial position?
@@tokeivo Pretty sure Stockfish 15 has learned how to tackle with this position. This puzzle was made a year or two ago which is why Stockfish 14 and under struggle to evaluate it. It might of been automatically been installed in Stockfish 15 that this was a draw.
I'm pretty sure this Stockfish knows what its on about now. Once its realised that the position is a draw there is almost nothing you can do about it as Stockfish can see 10s of moves into the future.
50 move rule would come into play before 3 fold répétition
Layton Jr the 50-move rule can be reset when black sacrifices both pawns, but then it'd actually be a max 150 moves to draw
the engine switch at 6:38 from cloud engine depth 84 to local engine depth 31, there is your change in evaluation. 50 move rule is beyond engine depth
I don't think stockfish 15 was confused at the start, I think it saw the draw but still evaluated the small advantage for black. I have noticed through using it that when the game is a draw with perfect play, but one side is "playing for a win" and the other has no winning chances whatsoever, that it will evaluate a small advantage for the side with winning chances to account for less than perfect play.
yeah ive seen that too, for example the rook v bishop endgame i think
No. Engines do not account for less than perfect play when they calculate the lines they also don't consider anything like "playing for a win". Engines simply find the best moves with no bias for either side. If an engine sees a forced draw it just flat out gives a 0 eval.
@@HeaDStrongTerran you should really look at engines more. I have put known draw positions and been given a .1-.2 advantage for one side because it has the pawn.
@@anthonychrane560 for the past half year I have been looking at them a lot since I am writing my own.
Just because you know it's a draw doesn't mean an engine knows that. Engines don't see plans they see concrete lines. A good endgame player can instantly see a position is drawn because there is no plan available and also it can determine an endgame to be winning long before he sees a concrete plan.
Engines rely on pre calculated table bases or they have to find a concrete winning line.
So in a drawn position an engine might still look for a win and give some advantage to one side until it sees the concrete best line leading to a draw and then it gives a flat 0.
@@HeaDStrongTerran I am talking about the king vs king and 1 pawn endgame where the king is on the space directly in front of the pawn and the opposing king has opposition. If it is the turn of the king with the pawn its a known draw, and the engine will recommend a line that ends in a draw, but still evaluate the position as slightly winning for the side with a pawn.
An older version of stockfish evaluates it as 0.00.
Which means either the new version has somehow gotten worse at endgame, and also can't recognize its own line that ends in a draw, or it's intentional.
7:30 what if we capture white's pawn instead of their bishop? Sure, you lose the queen, but you still have connected passed pawns and... oh. King on g2, bishop takes g3, mate. Nevermind. That just throws the game.
thank you
Better yet, back at 5:10, why would white not take the queen?
@@reubenmanzo2054 Because it's still a draw
@@knownas2017 Let's test this. Set up your board as it is at that time in the video, I'll do the same. White to move, I'll play white.
Bxb2
Your only legal move is c4.
Kg1
Your only legal move is c3.
Bc1
Your only legal move is c2.
Bxg5+
Your only legal move is Kxg5.
Kf1
Ok, your next move.
Actually, I realised that with the bishop sacrifice, white loses. Never mind.
Lichess Stockfish shows 0.0 (CLOUD) at the beginning probably because someone went very deep into the variation, like few dozens of moves and dropping at least one pawn, until Stockfish finally understood the position, and then he backtracked up, one move at a time, letting Stockfish recalculate the previous moves with what he found lower on the games tree (and with whatever it does).
This way, after going back to the initial position, it thought it is a draw, and then it got CLOUDED.
Then, when you try to play some moves, you probably go into different variation that got clouded and he stops understanding it again, thinking it's losing for white and trying to figure it out.
IN SHORT: it indeed doesn't understand what's going on and it's showing a draw at the beginning only because someone played deep enough and backtracked the moves from the point it finally understood that it's a draw :-)
Actually, it looks like people already evaluated the position to a high search depth on lichess, and then it saved to the Cloud. When he was making the moves on the board, Lichess was getting the positions from the Cloud that had already been calculated to a high depth. That happened until he made Qa5. When Qa5 was played, the position was not longer in the Cloud, so Stockfish 14 started calculating the evaluation on its own on his computer. Since Stockfish 14 doesn't understand the position until very high depths, it went from 0.0 to -6. If Stockfish 14 continues to calculate to the depths shown in the Cloud, the evaluation would most likely be 0.0.
You should have gone over what happens if the queen sacrifices itself for the bishop
King takes queen then black can only move those 2 pawns. The black king and other 4 black pawns are all trapped so the white king will get those 2 black pawns up by force then stalemate cause black has no legal moves. Position is bad cause the king and 4 pawns are blocked in maybe if the king was on the opposite side of the pawn wall he could take that 1 pawn that is blockading the entire board but
Is there a hope if queen captures the white pawn?
Suggestion: When showing a position, put "White to move" or "Black to move" in the graphic.
It's pretty pointless, since over 99% of the time is the turn of the pieces on the lower side of the board, basically your "point of view" that is about to act in any puzzle like this. Probably even higher than 99%. I think I have never seen a puzzle where if it's white's turn to move, but you see it from black's perspective.
In 99.9% of cases, you just have to see which way pawns go to (with the line numbers) and the player whose pawns move upwards has to move
I support that!
Leela told me all around it was between -4.09 and -2.40. Never found the draw after even 20 mins
EDIT: SF14 was the same as Leela and SF15 found the draw from the beginning in exactly one second. Interesting how some find it and others don't.
Leela is run on a supercomputer, so I highly doubt you have the full-power engine.
The board: Normal game and a draw
The engine: *BROKEN AND CONFUSED FOR 69 YEARS*
Hey Nelson, I just found your channel 2 days ago and I absolutely love your videos. I love the topics you pick up and I love that you do in depth analyses of chess positions like these. Keep up the good work.
Let’s ask AlphaZero.
Stockfish 14 is saying that it’s not a draw because the equation wasn’t updated. In stockfish 15, stockfish team added 8 features that drastically improve the evaluation NNUE system. That’s why stockfish 15 is now better than AlphaZero and he don’t get baited for material.
The old stockfish was litteraly obsessed with material gain, the new one (15) has a different vision.
Stockfish has been better than Alpha Zero for many many years now. Stockfish 15 in paticular has been a cooperation from the stockfish team and the Leela team to make it even surpass SF 14 in efficiency and evaluation. Alpha Zero is completely outdated nowadays and has even lost to older stockfish versions. I know the way A0 revolutionised the game is legendary but it has been some years since then and developing never stops. However we are quite close to the ceiling of chess already with Stockfish 15.
@@Jartran72 AlphaZero isn't even like a fully serious project I think, they just did some basic AI work and training in a supercomputer, and it just happened to beat old stockfish, the best engine of it's time. But yeah Stockfish 14 and + is better than AlphaZero.
Stockfish 14 was already marginally better than AlphaZero. There's so much misinformation about the AI that it's hilarious
@@GMPranav It beat Stockfish11 because Stockfish was running on an old laptop deliberately meant to skew the results - whilst AlphaZero was using a massive supercomputer endorsed by Google. If you plug the same moves into Stockfish11, Stockfish only gives a favorable evaluation to the unfavorable moves played in the game below approx depth 18 which is not enough for Stockfish to run its full capabilities.
@@scarcedude3353 the laptop theory is debunked. It was a one off match. Stockfish ran on a proper pc for 1000 matches and was demolished by Alpha.
Used Rybka, eval started out around -14 but made progress until staying at -5.09, never realizing it's a draw.
Rybka doesnt need to realise that its a draw, because im pretty sure that realising is not something that Rybka can do, what it can do is playing chess moves, so did you manage to mate Rybka yet from this position? because my bet is that it will hold you to a draw every single time.
@@esrasor You're right about a chess engine "not realizing it's a draw". Of course, as of yet (at least as far as we know) chess engines (or any AI) don't "realize" anything. The words are metaphors for human intelligence and consciousness. This video demonstrates that the position is a draw. There's no disputing that fact, so no I didn't manage to mate Rybka. The question asked in the video was "Why doesn't Stockfish evaluate the position as a draw"? We were asked to comment on what other engines did with this position. I reported what Rybka did, which was better than the first Stockfish evaluation, but still failed at (forgive my metaphor!) recognizing the position as a draw. I hope this explanation was better than my original comment and clears up any confusion.
It's simply the horizon effect in combination with the 50-move-rule. The Fritz-part shows it best. It reaches a depth of 40, which is shy 10 of what it needed to reach to conclude it's a draw.
You'd need a separate algorithm to detect these situations without thinking 50 moves ahead. It doesn't actually reduce playing-strength and really only is important when people use the engines for analysis.
Indeed. I installed Stockfish 15 in Arena 3.5.1 on my hex-core AMD Ryzen 5 CPU, and Stockfish immediately evaluated this position as draw. Earlier version of Stockfish did not evaluate this position properly (they shown black's advantage). It is a first time that chess engine can detect a fortress (I constantly provoked Stockfish to capture my bishop resulting a stalemate).
For all comments saying black can just sac the queen for the bishop, the king is close enough to stop both pawns before they advance, and white's pawn on g2 prevents all of black's other pieces from moving. ITS A DRAW.
yup, don't see why other people aren't seeing that
What if you sac the queen for the pawn
@@brianmontgomery580 then black gets mated.
I tried this out, as Black, on the hardest setting on a Chess app. it is possible for black to win if you get your queen to F3. if your opponent takes with the white pawn you take with the pawn on G4. then push a check with G3 pawn. this opens it up. If they don't take. then push one of your far pawns out. this will lure the bishop out then move queen to F2. take the bishop or let it take your Queen then you take with your F3 Pawn. I am a complete Noob when it comes to chess. so I might be wrong.
What if you put your queen on f2? Given where the bishop is, the bishop can take the queen or move to h2. If the bishop moves to he, blacks pawn can take the bishop and then be taken by the white king. Blacks queen can then take wife’s pawn and then be taken by whites king. Black has too many pawns for the white king to stop them all. So bishop to h2 does not work. Only other legal move is bishop takes queen. Blacks pawn takes whites bishop. If whites king goes to h2, threatening checkmate next move with its pawn, black can push its pawn, preventing whites pawn from advancing and blacks king can escape. White has moves so it is not a stalemate.
What if @ 7:08 black moves Queen to F2? Looks to me like it Bishop takes queen, pawn will get promoted. If bishop moves the other way, pawn can capture leaving white a move to avoid a stale-mate. I might be missing something though.
pawn wont be promoted cause the white pawn will move, check the king so the king has to move, either way he moves, makes it a draw.
edit:
the black pawn is preventing the white king movement so he has to move the pawn and then well draws the game
I also think that Qf2 is a move worth noticing. It won't change the fact that by playing correctly white can make a draw, but there are actually some moves that can force a mate for black, if white don't take the queen
@@vladzador white will have to take the Queen as that would be the only legal move
I remember this, this is a puzzle by the brilliant composer Frédéric Lazard. He has a lot of good compositions, I remember this one in particular was one of ChessNetwork's cool chess puzzle videos. If you have the time or the inclination you should check out some more of his work!
Lichess initially thinks it's a draw because of it's Cloud, when you put in an unkown move it starts freaking out
for these fortress type positions, just use Crystal, a modified stockfish especially made for this
Yeah I have tried it before. Stockfish 15 actually understands most of these kinds of puzzles. The only puzzle that I have never had Stockfish 15 solve is 3B4/1r2p3/r2p1p2/bkp1P1p1/1p1P1PPp/p1P4P/PPB1K3/8 w - - 0 1. Other chess puzzles will cause it to evaluate the position incorrectly, but that doesn't stop it from playing the best moves, similarly in this puzzle. For that FEN though, I have never had Stockfish find Ba4+ to seal the draw. Crystal was able to find it after about a minute of thinking or so.
Queen capturing that white pawn is a possible line that frees the black king and possible win for black
I had a thought.
What about having the Queen take White's Pawn?
This frees the White King. Black can now escort his barricade of pawns down to promote into new Queens, and White could only capture one before King takes Bishop.
If you drop the queen on F2 so that the bishop is forced to take it, pawn takes bishop and the board opens some. That would allow black to advance the pawns and get a new queen and run white down.
As a simple pleb 1200 i didn't understand why can't white sac the queen for the bishop and push forward.
It took me 10 seconds of play with the engine to understand why...
Cue the "it's always sunny" meme of Charlie with the conspiracy board for stockfish's "plan". 😆 🤣 😂
For what it's worth, Fritz 18 came to a point where draw is basically guaranteed via the following line:
Nc3-e4+ Kg3-h4
Ne4-g3 Qa8-c8
Ba5-e1 f4xg3
Be1-f2= (evaluated as -0.01, though that's probably because black has field control and white has to respond correctly to all black moves to cause a draw)
At this point, from human examination, capturing the bishop by g3xf2 leads to forced stalemate by g3+ (Either black King Move results in no valid moves for white, and even if the black queen could capture the checking pawn, that would be a stalemate as well), so that line can be dismissed in its entirety. The Queen is more or less in the same scenario as from the line examined in this video, which means there's no way for black to force a white play that sets up the possibility of white getting mated.
Probably the most curious part about this setup is that so long as white does not misplay, Black cannot capture the Bishop. If there is a scenario where Black can force white into a position where black can capture the bishop and NOT get forced into stalemate, I don't see it. Capturing the bishop along the diagonal from e3 to a7 is a stalemate (no legal moves for white). Capturing the bishop at g1 results in a loss of queen via Kh1xg1, and with that, eventual stalemate (the two "free" black pawns will get captured via white king). Black queen capturing bishop at f2 is a stalemate. Black pawn capturing bishop at f2 results in the above described forced stalemate.
There is such scenario in the queen protecting the f5 square variation. He says that it is making the same position, but it is not, as now white can't go to f5-g1 diagonal with the bishop backing up the knight on e1.
As has already been mentioned, this is the horizon effect of original chess engines vs NNUE for SF15. I'm no expert whatsoever, I believe 'originally' a computer would just endlessly calculate tenths of moves of the queen and bishop and at the end of the depth tree conclude Black's up 11 points in material -4 for a paralyzed king: there's no such thing as "this is a draw because no progress". In NNUE this kind of evaluation exists. Not exactly sure how this works, but I believe it 'acknowledges' that White is not worse with any particular bishop move, meaning there's no progress, thus technically the game must be drawn. I'd be interested to know where the difference in SF14's evaluation comes from, though.
Stockfish 15 saw the draw at depth 20 while Fritz couldn't see shit at depth 37
Almost certainly the horizon effect in action. en.wikipedia.org/wiki/Horizon_effect The game isn't drawn as far as the engine is concerned if there is still sufficient strength to win on the board and threefold repetition of position or fifty moves without a capture can be avoided. The exponential explosion of the game tree nodes guarantees that the engine can't look that far ahead. So the engine will keep rattling the queen around with no real plan but still seeing sufficient strength to win on the board.
Nice to see somebody else on this channel who understands chess engines 😄
lichess initially says 0.0 for me. Then I follow the draw line and it jumps to -6 lol
For anyone curios after Ne4+ Kh4, Ng3 any queen move Be1 seems to be the only move that holds for white because your threatening Nf5 double check and mate so black has to play fxg3 which you then can play Bf2 which you control the diagonal and black can not capture due to a stalemate. Going back now to Ng3 if black plays Qxa5 then Nf5# wins for white.
What if you sacrifice you're queen against the bishop on g1. Couldn't you bring one of you're pawns trough?
That looks like a plan until you realise the black king is stuck, so the white king is also just in time to beat the pawns.
If the bishop takes it's an advantage for for white maybe even win, better than the guaranteed draw for white
@@AgaresOaks You're right. Thank you
@@AgaresOaks At 6:56 Put the queen to F2. Force the bishop to take queen. Pawn takes bishop. Then you can move in your other pawn for a queen, while white king can move back and forth between G1 and G2.
@@my3dviews what if bishop to h2 and not take the queen?
Maybe you said it (I had no sound), but for completeness you could also show what happens when queen takes the bishop on f1 or the pawn on g2
The bishop can’t get to f1. It’s a dark-squared bishop.
But if queen takes bishop on g1, then the white king runs out and takes the c and d pawns and black gets stalemated.
I played around with this position before viewing any of this, & evaluated it as a draw. I found exactly the same sequences of moves as the presenter & Stockfish! I did try desperately like Stockfish to force a win for Black, but no dice. Even sacrificing the Q for B doesn't work because White reaches the c & d pawns no trouble at all. So, draw!
SF15 on a 2020 budget laptop went to 0.0 within a fraction of a second in all variations. SF14 on the same hardware gives big advantage for black although it does suggest optimal play.
Why cant, when the bishop and queen are on the bottom row, the queen take the bishop, obv the king takes the queen, then black make an attack with his pawns?
King can run to the c1, d1, c2 and d2 squares before any promotion, the pawns are way too slow
@@Wltrwllyngaeiou it would be a slow grind to move all pawns within two spaces of the other side of the board and then that's game.
The king can't be everywhere at once.
@@hariman7727 nope, it's an easy draw for the king to block the two adjacent pawns. If they both move to the second rank, then one can be captured while still controlling the queening square
@@hariman7727 Try it yourself first. 🙄
All that i see in this video is stockfish finding the correct moves instantly, whats happening is humans being distracted by stockfish evaluation number where we should be looking at the lines that it suggests instead. conclusion stockfish found the drawing line instantly.
But Stockfish 14 thought Black had a big advantage. It didn't know White could hold the draw even if Black plays one of these lines. It matters because, if an engine had White and a human had Black, the engine might resign.
@@rosiefay7283the engine does not resign when analysing, but now you mention it im not sure if engines resign in engine vs engine matches, i assume not, and in the man vs machine matches the resigning was done by the human operator.
@@esrasor engines absolutely resign in engine versus engine matches
@@esrasor they do resign in engine v. engine matches when the eval hits a certain point, like 10 points in favor of the other side
@@kemcolian2001 UCI literally doesn't support resigning. The GUI can adjudicate a game according to eval, but it's not the engine's choice.
I have been working on my own engine for over half a year and have studied how engines work. Searching moves at depth 20 and beyond is already an insane task. To do it naively is just impossible. So chess engines employ a lot of very sophisticated methods to simplify such momentous tasks. Those methods are very complicated and sometimes they are even counter productive.
To create a good engine not only these methods need to be implemented they also need to be fine tuned for best results. No matter how you tune them there is almost always guaranteed to be certain patterns or positions where they fail by either killing the performance or by being unable to make sense of the position.
Doing research for my development I have seen posts of other developers talking about 'problem' positions that trip up their engines. But a dev in stockfish team could probably put this position into the engine in sone debug or analytics mode and quickly find what's tripping it up exactly.
i think its because technically this game could go on almost forever without ever reaching a draw by repeating moves
There is also a funny variant in the position of 5:52. Suppose that White plays Ba6 (any move on the long diagonal works). If Black plays Qxg2+ after Kxg2 it's stalemate in two... for White! Black has two legal moves left (c5 and d4) but the bishop can take these pawns and after that Black is stalemated.
Actually white can checkmate black with Bb8 then after black moves a pawn Bxg3#.
You can also mate with Bb8 followed by Bxg3#
I immediately saw that the diagonal for the QB needed to be opened, but it needed to also be a forcing move (ie. a check), Unfortunately I found the check at K2 and not at K4
4:12 at this position is check mate (amazing an engine can't see it)... Queen goes to b2, only move is bishop anywhere.. queen c1, bishop block... queen d2, e1, e2...etc.. until you go Queen f1... bishop blocks...then Queen f2, sacrifice your queen...only move bishop takes queen...pawn takes bishop...king moves to h2 (only move), you give check with pawn G3...king return to h1 (only move), you promote pawn to a rook/queen in f1. Check Mate
i find it funny how many people are trying to point out flaws in this that doesn't actually exist... if one side makes a blunder, then sure, its a checkmate. but the game is a draw otherwise. after black sacrifices the queen on f2, and black takes the bishop, white isn't forced to move the king at all. white can play g3+, forcing the king to take, which is stalemate.
I've just found that stockfish doesn't understand forced stalemate traps. It typically comes up in unavoidable rook chase stalemate traps.
I think it's counting material at the end. Since black's pieces are net worth more, they have a slight advantage over white.
If black playes Qe4 then you have to be careful because Bf2 is a blunder, which is followed by gxf2 takes, Kh2 then g3+, Kh1 and lastly f1=Q#.
Also if Qf4 then Bf2 is another blunder, because gxf2, g3+ is forced, Kh4 then you have to take then queen and then black will promote and win. If king takes pawn on g3, then it is a draw by stalemate
How about qween takes bishop on g1?
What if you put queen on f2, the bishop can only do one move- take it, then u can take bishop with pawn, promote pawn to queen.......
There is such scenario in the queen protecting the f5 square variation. He says that it is making the same position, but it is not, as now white can't go to f5-g1 diagonal with the bishop backing up the knight on e1.
At 9:36 Q takes Bishop. W King must take Queen... Now the WK has exits, but only 1 peice... a stuck pawn. K must leave its holding pattern... So let's see if you covered that move... Ok. If Bq takes wb the WK takes bq Black promotes pawn... but black can only move it's two pawns and the moment the wk takes the last remaining BP that can move it is a stalemate...
Basically there must be some workaround for the evaluation functions of stockfish and whatever engines you can find, to evaluate positions like this, where it really isn't obvious whether is it a draw or not, even though a human player can work this out, the computer kinda have to see indefinitely into the future in order to figure this one out, which is obviously impossible, we cannot have infinite computing power, so the algorithm has some bias for certain positions, that happens to make the evaluation of this positions, and i'm sure many others, completely wrong. It is sort of like when computers have to do math with real / floating point numbers. As there is no limit to how long a decimal can be, the computers have to sacrifice some of the precision in order to be able to work with a large spectrum of numbers, which is why sometimes things like 1/10 = 0.0999999999994382 happen.
There also was a line where the black takes the bishop on g1 with the queen and still has connected two. But if my calculations are right the white king can block them and draw the game anyway.
Yes if black sacs the queen, the white king recaptures and has time to stop both the pawns, meanwhile, black is out of moves and the stalemate will be forced sooner or later.
Stuckfish then.
😂
,,We will look at Queen F8 Later“
Never looked at Queen F8
At 3:23, there's an annotation in the bottom right: 3. …fxg3 4. Bf2! gxf2 5. g3+ and stalemate next move. (Note that …fxg3 is forced to stop Nf5#, and if not …gxf2, we end up with the bishop on the diagonal as in the video)
One of the rare situations in which the superior side is the side being stalemated(if the black queen captures the white bishop on g1)
Isn't Stockfish telling "the likely result" with its index? If so, that -0.17 seems pretty correct, as white has losing moves it can make. As white keeps making the right moves to force a draw, Stockfish determines that its going to end in draw by repetition at the very least. Remember that white can lose in that position by moving the bishop away from that diagonal (giving the 2 pawns room to advance), and holding the bishop in that diagonal is not forced, but just optimal for white. So seems that Stockfish 15 is correct at all times.
7:30 Stockfish thinks it's winning, but not by as much as it was half a dozen moves ago.
8:49 According to the notes for Stockfish 15.1, "This release also introduces a new convention for the evaluation that is reported by search. An evaluation of +1 is now no longer tied to the value of one pawn, but to the likelihood of winning the game." IOW, it appears that Stockfish 15 is a new algorithm, hence the discrepancy with 15.
I think the engines gets confused because if you make a mistake it's not a draw.
My teacher taught his students. Even when you can see that you've lost, my example will be you're down to your king the opponents got too rooks and a king. Keep trying to force the drow at least until it's clear that they know what they're doing
Suppose black captures the bishop and sacrifices the queen. I haven't played it out, but is it possible to promote one or both pawns? The worst that can happen is a draw which is the likely outcome already.
Ok so I tried this out, as Black, on the hardest setting on the Chess app. it is possible for black to win if you get your queen to F3. if your opponent takes with the white pawn you take with the pawn on G4. then push a check with G3 pawn. this opens it up.
If they don't take. then push one of your far pawns out. this will lure the bishop out then move queen to F2. take the bishop or let it take your then you take with your F3 Pawn.
Engines are optimized to find the best move, not the exact evaluation. If one move is at least -10 and all other moves are at most -11, they don't have to work out the precise evaluation in order to know that that one move is the best.
clearly shows that fritz and sf14 dont have a good way to decide "there is no progress possible here", they just calculate to their given depth and then use their algorithms to judge material and position, and black will always have superior material and position throughout the entire calculation, white cant force stalemate or draw by repetition within the calculated depth...you need to set the depth to well above 50 to make sure that after the knight gets captured there is enough depth left for the engine to calculate the 50 moves rule and decide for draw by that...but it might be that even that fails as black can sacrifice two pawns to reset the counter, so probably fritz and sf14 need like 160 depth to judge this correctly...
Once pawn takes knight, queen sacrifices and takes bishop. King takes queen, and pawns advance. King can't take pawns if they advance together. The advance has to be timed to prevent king from getting in front of queen pawn. Force the king to move back. King can't go around due to other pawns. First pawn to reach last file mates. Unless there is a time- or move- limit, black wins. I presume that's what you are looking for.
King cant take but black is in a zugzwang position...
@@lastorderx20001 I figured out my mistake after posting the video. I am too rusty at playing chess.
Alternative Title:Human Roasting Engines
If a given position has CLOUD in green, it means the position or a line of that position are memorized, like an opening but for puzzles.
I just put this on Stockfish 15 and it finds the drawing line very quickly
At 5:32 what if the Queen just takes the bishop, that would prevent the stalemate no? Or I guess the white king would take the pawns after which black is in stalemate?
it's 5am and I haven't slept yet
why can't the queen f-1 check, move to f-2 and force the bishop to be captured?
it either trades for the queen and leaves a black pawn blocking g-1, thus letting the two connected pawns through, or moves to h-2, letting the queen go free anyway
question: im no chess expert but lets say the white bishop is taken at d4. like at 4:39, wont it end up like: king to g1 , queen to d1 resulting in a checkmate? the king cant go: f2 because black pawn at g3, likewise king cant go h2 because of g3 pawn. cant go f1 or h1 because queen at d1?
am i missing something?
....oh... wait... so let me get this straight: because kings cant move into a position which causes them to go into check and because if blacks move would cause a king to be unable to move due to it entering check if it did but not putting it under check before hand means that it isnt check mate means that black cant make the move...
.... white to move: knight to e4, king to h4, knight to g3, king to g3 taking knight, bishop to b6, queen to a2, bishop to f2: check, g4 pawn goes to g3 to block check, bishop must either move or take pawn, if takes pawn king takes bishop, white king g1 queen a1: checkmate if pawn takes bishop same result basically.
alternatively after knight goes e4: king f3 knight must move to avoid capture from king or d5 pawn. if knight goes e2 basically similar result but not from pawn. but that opens a whole nother can of worms for black but may give it an oppurturnity to counter whites movements...
it's all about the speed of the engines. If they can evaluate 50 depth, they would recognize that no process can be made and it's a draw. However, when they are at lower depths, the position always look better for black due to the material advantage. At that point, it does not know if that material advantage can be solidified within 50 moves before the game is drawn automatically.
Stockfish and Fritz are great at tactical sequences and all sorts of computation and combination. However, they lack what we might call "pattern recognition" as animals do. If you have seen AlphaZero vs Stockfish games, Alpha won many games with very simple 'ideas'. For example, in one of the games it "locked" two pieces of Stockfish. Stockfish failed to understand "long-term" effects of that lock.
What if you scraficed the Queen for the Bishop. Maybe that's why the engines are confused ablout.
The white king is in time to stop both pawns, you can try it yourself, the black king is stuck
I'm guessing why the engine gets confused is because black has a wide variety of moves open to them, so despite the vast majority of the moves being functionally indistinguishable, there are too many possible board arrangements to reach a repeated position draw if black wants to avoid it (which they would if they have the advantage) and black can also push their pawns to avoid the 50 move rule, so the engine would have to analyse well over 100 moves deep to reach a draw, unless the engine recognises that black cannot win with white's bishop on that diagonal and short circuits the analysis.
Why can't we play Qxg1+ when white blocks
What if the queen sacrifices itself for white's pawn?
black still couldn’t win. White captures the pawn on g2 and shuttles the bishop along the diagonal, capturing pawns whenever necessary. At that point, black can’t move so it’s stalemate.
If Queen sacs itself for White's pawn, then Black loses. King will take on g2 and the Bishop will come back to deliver mate with Bxg3. Its a Mate in 3 with Kxg2 followed by Bf2 and Bxg3#
Because there is no limit of moves (because putting on check reset the move counts for the draw) and the position is never twiced the same the rules doesn't apply to "force" the draw.... therefore the engine continue playing considering he has a huge advantage.....
Now this is why the engine fails but I wonder if there is an extra rule o enforce the draw ? because otherwise in tournaments the clock might make black win.... I am really curious to know if there is an extra rule for tournaments that is not fed into the engine.
I'm not a pro, and would appreciate insight on why black can't force a checkmate with:
Queen F8
White Bishop (anything)
Queen F1 check
Bishop blocks
Queen F2
Bishop forced to take Queen
Black pawn counter-take.
White can then:
Push G2 pawn (black queens and has mates)
Or
King H2
Black pawn G3 check
King H1
Black Queens on F1
So, what am I missing?
White's g3 is check
I think what make stockfish confuse because black can do queen sacs f3, and if white blunder to take the queen with pawn, its black win i think, since black king can move now.
But if white dont take queen sac its draw
Since a chess engine is mostly built for winning matches, recognizing a draw reliably just has lower priority. This effect just doesn’t make any engine lose any game, so it’s not considered a true error.
We had several such positions in a computer chess course I attended in the 90s. One was a solid fence of all 16 pawns blocking the way through completely. Ever side only have pieces only on their side. One side just had a king, the other had lots of other pieces, but none of them could hit a pawn of the other color. So obviously this was a draw. But chess engines at the time could not evaluate the position properly, they didn’t understand that no pawn move was possible and that the no pawn could be hit.
r3kb1r/8/1p3p2/pPp1pPp1/P1PpP1Pp/3P3P/8/4K3 w kq - 0 1
Lichess still today considers this a -13.3 for black.
Even worse is this situation: r2qkb1r/8/1p3p2/pPp1pPp1/P1PpP1Pp/3P3P/8/3QK3 w kq - 0 1
Because here Stockfish doesn’t understand that the only winning move is to take a pawn with the queen (to open up the fence). So Stockfish will run into a draw while I could win the game: Just take a pawn, sacrifice the black queen, and the remaining pieces are powerful enough to overcome the white queen.
What happens if black sacs their queen taking on g2? I don't think that line was ever explored in this video
What about n to e2? Mate in two?
nice
fyi: I used 365chess and it said -12.xx all the time lol. One could assume because technically white could do a mistake and black has better winning chances, (for instance black queen to e4,f4 and you can't sacrifice your bishop as explained contrary) while the only bad thing black can do is sacrifice the queen - but it's not how the evaluation works, it's indeed flawed and should show 0-0 if it knows.
what happen if you sac the queen . force the bishop to block the bishop then sac the queen for the bishop king must take the queen so its now free to move and you got the two pass pawn to get a new queen
well black's king and pawns are totally paralized and the white king is able to gobble the two passed pawns.
and engines look for winning moves, even if a queen sac is possible and leads to a win for black, the engine wouldn't allow that for white.
The king can catch the pawns
White doesn't have to take the queen, unless black puts the queen on f2, in which case in the end white has g3+ which is stalemate again.
If you sac the queen for the pawn you take back with the king then continue to move the bishop back and forth same as before. The kingside is completely frozen as long as white's king stays on g2 and the queenside pawns will be taken by the bishop when they're forced to advance. It's a draw.
If black puts their queen on f3 hoping you take it with the pawn.... just ignore it. Position remains as is.
If black sacs for the bishop then the pawn on g2 locks up the kingside while the king has enough time to catch the queenside pawns. Still a draw.
@@cmck362 How about at 7:04 Black moves queen to F2. Force the bishop to take queen, since it is the only move. Pawn then takes bishop. Then black can move in the other pawn or even both pawns for queens, while white king can move back and forth between G1 and G2. So, black can win.
At 5:46 cant black move the queen to f2, white takes queen, black takes bishop, king forced move to h2, queen promote, black wins?
Edit: can someone help me understand why that wouldn’t work?
White can move g3 rather than Kh2
@@billcook4768 ahh thanks. didn't see that at first.
The title is wrong, stockfish very quickly knew it was a draw
3:48 what if queen to f8 to F3 or F2 to taked by pawn and bishop woth sacrificed the queen?(the first idea is eliminated the pawn G2?)
I noticed that Stockfish 14+ was searching out to well over 40 moves (each side). this was probably deep enough for it to exhaust all the possibilities to determine the position was a draw. Also, do you use end game table base files on your PC? that can help the engine determine the outcome faster. When you made the queen move, Stockfish hadn't analyzed that line (the depth shows 31 (15 moves for each side), so it doesn't yet know it was an actual draw so the score just indicates the material difference. If you let it run long enough, it would again figure out that was an actual draw and report a score of 0.00. MOST of the positions like this (that "confuse" engines) usually involve deep search that might be beyond the limits of the engine. but notice, the correct move was made, so it doesn't matter! It wasn't like the engine was choosing the wrong move from the get go, though that can happen sometimes.
it actually is winning for black.
you take a bishop with the queen, king takes and its not a steelmate anymore. and black just moves the pawns and wins the game.
White king can stop both black pawns so it is a draw
@@user-de1gi8wh4s not all of them at once. he cant split and stop either right ones and left ones at the same time.
@@user-de1gi8wh4s if white king goes left, black pawns on the right simply proceed forward. white king wont be able to stop both sides. black king doesnt need to do anything.
@@96annihilator black pawns on the right are stucked, they can't move, black king is also stucked there
You can sacrifice the queen by taking the pawn on g2 leaving an escape for the white king when it takes the queen. Thus ending the stalemate, unless i am missing something?
Minute 4:10 what if black sacrifices the queen for the bishop and then tries to promote is two powns on the left side of the board?
Complete chess novice here but…what happens if black sacrifices the queen on f3? Wouldn’t this give white room to move (at least with the bishop) and maybe a passing chance at black maneuvering a pawn to the other side of the board? Maybe I can’t see far enough ahead but it seems like something I’d try in my games that have no consequence.
I came to that same conclusion. I am also a complete novice at chess.
what if you do Qc7 and push your c poan to promote? ...or also exchange queen for the bishop and then a poan to promote, something like that? thank you.
Instead of crediting a random internet person, credit the author of this tricky study: Frédéric Lazard (1946)
What happens if queen takes bishop when it’s pinned? Doesn’t black have a way to win the pawn endgame after its queen is taken?
When it's just queen and bishop left, can you not go Queen A1, forcing bishop to block, then Queen A2, any move the bishop wants to make, then Queen G2, King's forced to take the queen, you can then move the pawns C or D down, forcing the king to have to move out on white's side, and you either get a C, D or G queen? (genuinely asking this, because it looks to me - a chees noob - that you could do that?)
White still has the bishop which picks off the c and d pawns.
The suicidal knight draw pattern lol
2:30
Be1 is checkmate
The black king is on h4
The h5, g5, and g4 squares are blocked by black pawns
The g3 square is controlled by the white knight.
The h3 square is blocked by the white pawn.
The black pawns are too far away from the white bishop to capture it.
The white bishop is not in the line of sight of the black queen.
I’m suprised you didn’t mention that in the video.
Edit:
Sry I didn’t see pawn to g3
Taking the bishop is draw coz the black king is stuck, What if the queen sacrifices for the pawn or goes on f3 so that white has a move , then the pawns can march forward??
if the queen goes to f3 u don't have to take it, just ignore it
@@zeninguem1937 If queen goes to F2, then bishop either has to take it, or move to H2. That would allow one of the pawns to advance for another queen.
@@my3dviews queen to F2 doesn’t work. Bishop takes queen, pawn takes bishop, then white pushes his pawn forward for check which forces the black king to either g3 or h3 either of which end in stalemate
@@Robert-mx1io Okay. I'll look that over.
amazing ... even if the queen takes the bishop at 4:08 its still a stalemate as the white king can kill the free black pawns and then black cannot move anymore.