Wow. I'm curious what part of the problem it's losing the thread on. That would be the reason it needs depth 96. There is some position where it veers down a rabbit hole and eliminates the winning idea from candidate moves too early. This early candidate pruning is the reason engines on very rare occasion cannot see some winning moves at any reasonable depth. It's also why you get a funny scenario where engine can't something until you put the first move of it on the board, then it instantly figures out the rest.
@@stoutlager6325 I don't think the depth number is that important in this position. I let my local computer (allowed 4 threads, 1 GB hash table, no table base) think on it for 5 min it was at depth 100+ and saw the M12. Depth 100 is impressive if it's near the opening or there are still many pieces to the board. Fun fact: it had a much harder (longer) time solving the Tal puzzle... So the Tal puzzle is still GOATed
@@fahimp3 You can solve is much faster if you set threads>1 and MultiPV=10. For me it takes 3 seconds to find M+12 (threads=16, multipv=10, hash=4096, modern 7950x3d cpu)
The thresholds used for pruning and depth reduction, which allow searching incredible depths very accurately, ignores this line because it looks too crap to be worth checking. You can easily tune the engine to consider these positions, and it will solve puzzles like this almost instantly... but its overall playing strength will suffer tremendously.
Yeah I was wondering if the mate-in-n engines that try to find the forced checkmates in the endgame would handle this easily or not. I'm surprised Stockfish's NNUE didn't help out here though.
@@CoalOres NNUE's (and even larger NN's) aren't able ot properly evaluate extremely complex positions like this - exhaustive tree search is the only way to find a solution
Chess: Why do Jewish, Hindu & Christian players do better than Muslim players? So far the top players of Muslim / Arab countries I've seen either are ethnically Jewish or have changed to federations of Christian/Jewish/Hindu countries.
@@1992jamo Yay Leela!!! Stockfish is a TALENTLESS patzer who lost to Leela 3x classical in a row in the 2019 WFRCC finals lol Good guys vs Bad guys - Axis of Evil (Good guys?) wesley so, hans niemann, bobby fischer, sergey karjakin, vlad kramnik, ian nepomniachtchi, alireza firouzja, fabiano caruana, levon aronian, viktor korchnoi, pal benko, mikhail tal, amin tabatabaei, gukesh dommaraju, maxime vachier-lagrave, nodirbek abdusattorov, juan manuel bellón (anna cramling's dad), vidit gujrathi, andrii baryshpolets, vishy anand, jan-krzysztof duda, vladimir fedoseev, lichess, walter tevis, max pomeranc, chesscube, Sergey's hypothetical pure 9LX parallel, PCAP, 9LX, christopher yoo, gata kamsky, sam sevian, michael adams, leela, ATLA netflix fire nation & sokka in original, ferdinand marcos sr, ben johnson of perpetual chess podcast(?), elon musk(?), nikhil kamath, bibi netanyahu, both vladimir putin & volodymyr zelenskyy (because of sergey karjakin), ben shapiro(?)/alan dershowitz(?), mustafa barghouti, barack obama, donald trump, tristan & andrew tate 🇺🇦🇮🇱🇹🇼🇮🇳🇺🇲🇬🇧🇦🇺🇯🇵 - Allies of Evil (Bad guys?) magnus carlsen, hikaru nakamura, garry kasparov, daniil dubov, veselin topalov, liren ding, aryan tari, abhimanyu mishra, tigran l petrosian, tigran v petrosian, paul keres, efim geller, parham maghsoodloo, pranav venkatesh, sébastien feller, vincent keymer, paco vallejo, pentala harikrishna, arkady dvorkovich, peter heine nielsen, richárd rapport, alexey sarana, ches***m, fred & josh waitzkin, PCA, FIDE, NCFP, chess1, andrew tang, alejandro ramirez, sam shankland, nigel short, stockfish, ATLA original fire nation & sokka in netflix, prospero pichay jr, lex fridman(?), mark zuckerberg(?), sumner redstone, ariel sharon, ali khamenei & mahmoud abbas (because of alireza firouzja & randa seder), alan dershowitz(?)/ben shapiro(?), bassem youssef, joe biden, george bush jr & sr, piers morgan 🇷🇺🇵🇸🇨🇳🇵🇰🇳🇴🇮🇷🇮🇶🇰🇵 - axis of evil (female) (good girls?) tingjie lei (or paikidze), janelle frayna, anna cramling, dina belenkaya, rahma zein, randa seder, lotis key, carissa yip (or sara khadem), irina krush, heather walker, robbi jade lew - allies of evil (female) (bad girls?) wenjun ju, bella khotenashvili, alexandra & andrea botez, abby martin (so much for XMCr-bQk_DM ft mike prysner), turnabouttess, leny so, atousa pourkashiyan - nakamura, jennifer yu, dani chambers, nemo zhou - maybe slightly axis: agadmator but idk antonio radić kinda likes magnus. also roderick nava who replied to me. - maybe slightly allies: gothamchess but idk levy rozman kinda likes wesley. also deniel causo who didn't reply to me. Note: They correspond. Eg Wesley Vs Magnus Bobby Vs Garry Hans Vs Hikaru Christopher Yoo Vs Andrew Tang Dina Vs Andrea etc More info: FlNsFKeq-iE
9:05 in case anyone didn't see why the h pawn can't move, it's because it gives up the g6 square. And Kg6# covers the newly vacant h7 escape square. Edit: I spoke too soon! At 10:28 agadmator covers this.
Seems so - not only do you have to calculate advantageous moves for all of your candidate pieces, but you also have to calculate disadvantageous moves for all of your opponents pieces. Lots of calculating! 😵💫
@@zooniba his Brother in law does show a lot of short videos with similar chess puzzles but also videos of chess engine games Really good Channel I definitely recommend him Jozarov Chess
This was actually beautiful to watch. It reminds me why I love chess. It is never a straight forward game regardless of how much theory and computers advance. One new move and it opens a whole world of uncertain possibilities! Thank you for the video
I saw the idea fairly quickly and thought I had solved it after 4. Ba1, but I missed all the other nuances, and I had no idea how deep the position was until I played through your full explanation! Thanks for sharing this.
Its a great puzzle because it has all the elements: The obvious move isn't the first move. There is a queen sacrifice. The obvious move only gets played at the right moment. A zugzwang becomes evident. A pointless pawn push ensues just to use up moves to complicate. An inevitable discovered check forces black to sacrifice material until he's mated.
Beautiful puzzle. Seeing the Queen sacrifice & understanding zugzwang idea is very hard indeed. FWIW, Crystal sees Bc7 instantly, even on my phone. For those who don't know, Crystal is an engine derived from Stockfish and is really good at solving studies and compositions.
If you let your engine analyse just a single line (PV) it wont find Bc7 cause of heavy pruning at the start of the search tree which is done mainly for speed. If you use the Multi-PV option Stockfish will find it. Ive played around with it a bit and SF wont find it up to Multi PV =5, but when you use Multi-PV = 6 then it finds Bc7 in about 20 seconds.
@@lockaltube Multipv = 10 is just forcing it to try about every legal move. I tried 10 first too and then tested what the lowest was to find it... which is 6
This is just Stockfish's "Late Move Reductions" at work. It's not solved by using higher depth by itself because Stockfish just prunes the branches where it sacrifices a queen without compensation in the next 10 moves and only calculates the 'non-terrible at first sight' moves to the full depth.. Purely depth-wise (without pruning), it would find the mate at depth 23 (since it's a mate in 12). The reason it does find it eventually, is that at extremely high depth, as a "side effect", it's also going to less aggressively prune the first few moves options. The proper way to computer-solve puzzles like this , is to run Stockfish with specific parameters that turn off the aggressive pruning (which is worse for general play, but much better for solving 'counter intuitive' puzzles)
I gave this to my own copy of Stockfish 16.1 running on a Linux computer. With my default setting of 4 cores on a 4GHz machine, the evaluation went to mate in 12 on ply 62, consuming a little over 17 seconds.
What I find interesting is that if you put in Bc7 the engine suddenly sees the win. When I first put the position in an engine, I forgot to add the a5 pawn, but it still thought that the position is a draw, until depth 36, where the solution is also Bc7. The a5 Pawn makes the position so much harder for the engine.
It's because the engine doesn't even check the move because it appears to lose the queen. The moment it does check the move it finds this immediately. And if you use brute force full depth analysis then it also finds it immediately.
I'm not sure if I've got this right (and I usually do get these things wrong!), but at the position at 9:10 , can't black play Qf8? The white king can only check black by moving away from the defence of his pawn, but then black can block check with Qg7. If BxQg7 and then KxBg7, the white pawn can be picked up by the black king, which can then defend his own pawn on the next move - and so white loses. If, after Qf8, white plays a waiting move, say Bb2, black has Ne8+. Although the white king can then move to a square defending his pawn (e6) while checking black, black then has Ng7+ blocking the check with a check of its own. Again, white loses, because the black pawn can then move after white's next move and so there's no more zugzwang, and white can't stop the pawn marching towards queening without losing both bishop and white pawn.
Probably because engines are build to discard possible moves using a zero window search and depth reductions. This allows them to look down the important paths at a far greater depth in most common positions, but apparently this queen sacrifice is just too deep to for the engine to search deeper. Very interesting :) This position also relys on zug zwang a lot, so maybe the engine discards these lines after a null move search early on...
What are you talking about. On my phone I have SF_NN_12 and it quickly finds 1. Bc7 as the best move (depth 30 it sees that it is mate in 12 and it took less than 1 minute. And from then it switches eval between 0.00 and mate in 12 a few times, no idea why that happens. If you set up the board with the Bishop already at c7 and black to play, funny enough it takes only 1 second to realize that it is mate in 11 with best defense.
Stockfish 16.1 in Arena can find the mate in 12 in just 7 seconds with depth 20/21. Just turn on the Multi PV mode to 10. Edit: LC0 and Sugar NN found it too in 15 seconds
One big question that didn't get a solid answer: Why can't Black play h6 or h5 at any point? 🤔 Answer: Tempo turned Positional. The Black Queen needs to be on the 6th rank for h6/h5 to work, but after Be5: It was impossible... in 3 different way!
I had increased the number of variations to 8, and stockfish 15 at home was able to find the solution in about 40 seconds. I think it just never considers the queen sacrifice in the first instance so it just calculates the drawn line deeper and deeper --- it truly misses the creative spark that humans have!
White has only 25 possible moves, so you have a 4% chance of finding the winning move by blind guesswork. There are 3 for the pawn, 2 for the king, 4 for the bishop, and 16 for the queen (7 vertical, 2 horizontal, and 7 diagonal). *Edit:* This is not to say that this is a simple puzzle to solve, or see the final solution clearly at once. Rather, it shows that even under considerable constraints, there are still myriad possibilities. I was just saying that it's possible to get the first move of the sequence right by sheer luck, and the odds aren't as long as they are in some other positions.
@@malsawmzela609 Very true. I wasn't trying to suggest the whole sequence is easy. I should have made that clearer. I find one thing especially interesting, and it just shows how complex and free-wheeling chess is. I said there are only 25 possible moves for White in this situation. Yet, even though White is down to a quarter of its starting material, and its king is sharply constrained by Black's pieces, 25 moves is still 5 moves *more* (or 25% more) than are possible at the start of a game.
I analyzed the position on my (modest) laptop (2.6 GHz base speed, 8 GB RAM); it took about ten minutes to reach depth 72 and then after ten minutes it jumped at once from 72 to depth 84 with the solution. Overall, it took about 22 minutes and 1.4 B nodes to find the solution. The Stockfish version was Stockfish dev-20240723-b55217fd. I would say it isn't a big thing. It takes time, but the engine does find it.
One of the greatest puzzles I’ve ever seen. I was thinking pawn g7 then queen b8 able to line up with the dark square bishop by going queen d6 and go from there
Dear @Agadmator #Agadmator #submission #correction You say that one Dutch composer's first name is "Istvan" in this video, but his first name is actually "Gijs." I think you are mixing him up with Istvan Bilek, the Hungarian. Source: ua-cam.com/video/8wCJalNkTEI/v-deo.html Agadmator's other video, haha
Stockfish 16 plays gxh7 and rates it 0.00, but as I play Bc7 it starts calling it -M11. It doesn't find the first move, but if started from the second move it does perfectly fine i.e. mate in 11.
Oldie, but goodie. I love games where engines fail to solve it. I showed this puzzle to a friend who argued that Stockfish today can solve anything in seconds. Whoops!
I think the reason why engines can't find it is because they don't actually consider every possible move. They reject 'bad' moves and don't even consider them anymore. I've sat here and watched Fritz 19 analysis the position for the past 2 hours. There are 25 possible moves for white and Fritz only spends time calculating 2 of them, gxh7+ and gxf7+, the rest it skips over. So even though it's up to a depth of 42, which is deep enough to see a mate in 12, it's only considering 2 moves at that depth. Unless it goes back and looks at the other 23 moves, it will never find the mate. I'll let it run for a while just to see what happens, may even let it go overnight.
@@BREAKocean Tournament/serious 6 man board 1 team play was around 2350. Online more than 10 years ago in 10 minute was 2650. I gave up trying to be serious in my teens when I played with people who I could not hope to match without devoting my life to the game.
Bc7 Qxc8, gxf7 Kh8, Be5 as if white king moves it threatens checkmate, and if h6, the white king moves to g6 mate, that's my guess. Plus after Be5 Qc5, white should not moving king but Move Bishop somewhere else, Like a1, as the immediate king moves allows white to trade the bishop and allowing blacks Kg7.
If you remove the black pawn from A5 (thereby not having the pawn moves at all) the engine solves it.(Mate in 8). Crazy. Similar set of moves. Starts with the same bishop move.
I'm more surprised that engines cannot solve this. What are they getting stuck with? Makes me question engines 'way of think'. Are they not considering what looks like a suboptimal move, unless they see the victory, so they discard the first move as they may not be able to calculate the victory for 10 moves ahead. Engines don't fully calculate moves to forced victory, but rather moves that increase chances to win. Please someone correct me if that's not what happens. Very interesting video.
@@wilkinsune Agadmator’s Brother in law Jozarov has a great channel on you tube he showcases a lot of engine games Stockfish Leela C0 and other engines Great Channel
Stockfish 16.1 (version 2.3.7) came up with answer in seconds using infinite analysis on my intel iMac Pro 2017 3.2 GHz 8-Core Intel Xeon W 64GB 2666MHz. Stockfish says Depth=86/24 8169 kN/s
Believe me, i did find the correct idea of mating the king on corner with that bishop. I did figure out the bishop move and then capturing the f pawn. But didnt really get upto the end
This such a beautiful puzzle, I stopped thinking after setting up the idea, but didn't calculate deeply enough and fell for the same trap as Antonio😂. The pawn on a5 is the true mvp of this puzzle, without calculating forcing moves out of instinct I wouldn't have seen white's resource, but without it black would have lost sooner. I'll be sure to brag about solving half of this to my friends at the bar and the library
It's not about the depth of the calculation or the power of the computer. The engine has already decided between 2 moves (overlooking another move that is the winning move). From this point on, the engine won't go back in the decision. Ironically, if you use a old brute force version of Fritz, you probably solve this using a Penthium II
I also asked Stockfisch. After about 3 minutes he had the solution (AMD Ryzen 3950x with the setting to use 8 cores out of 16). I then tried it with 16 cores and was very surprised! Stockfish didn't need half the time but twice as much (6 minutes). Obviously a bug in Stockfish...
Not a bug, stockfish distributes tree branches in a way to be _the best chess engine,_ not for solving puzzles. Imagine you have a tree, so that in the start it splits in 2 branches, then these branches split in 5 or 100 branches and so on, and you never can see the end of the tree, you can maybe see the huge number of branches nearby. And you have N workers, and your task is to schedule workers to saw this tree for logs (and potentially find a golden branch). Of course you will process the tree differently depending on the number of workers (each worker will process every Nth branch, N changes as avalanche)
Well, as I understand it, the skill is to estimate which branches could be promising and to investigate them more intensively. (i.e. to search in depth and not in breadth). Anyway - I did a few more tests with different chess engines. Some do not find the solution at all (like LC0, Plentychess, Komodo and Dragon v1). Stockfish did not find the solution in version 14.1, but in version 16.1 after 40 seconds and in the latest development version after 143 seconds. Very astonishing in this test: ShashChess -> whether in the old version 14.1 (from 2021) or in the current version 35.3, the solution is found in less than one second!
That's a curious quirk of the engine that it can't evaluate the position. That said, as soon as you move bc7, the engine figures out the rest immediately.
The Microsoft outage was due to someone using a supercomputer to try to analyse this to depth 100
😂
*Crowdstrike
@@aryan519 fr bro
On Lichess the position is now analysed to depth 96 and Stockfish can see the mate in 12.
Wow. I'm curious what part of the problem it's losing the thread on. That would be the reason it needs depth 96. There is some position where it veers down a rabbit hole and eliminates the winning idea from candidate moves too early. This early candidate pruning is the reason engines on very rare occasion cannot see some winning moves at any reasonable depth. It's also why you get a funny scenario where engine can't something until you put the first move of it on the board, then it instantly figures out the rest.
@@stoutlager6325 I don't think the depth number is that important in this position. I let my local computer (allowed 4 threads, 1 GB hash table, no table base) think on it for 5 min it was at depth 100+ and saw the M12. Depth 100 is impressive if it's near the opening or there are still many pieces to the board.
Fun fact: it had a much harder (longer) time solving the Tal puzzle... So the Tal puzzle is still GOATed
@@fahimp3What program did you use?
@@pioussutherland En Croissant GUI + Stockfish 16.1
@@fahimp3 You can solve is much faster if you set threads>1 and MultiPV=10. For me it takes 3 seconds to find M+12 (threads=16, multipv=10, hash=4096, modern 7950x3d cpu)
The thresholds used for pruning and depth reduction, which allow searching incredible depths very accurately, ignores this line because it looks too crap to be worth checking. You can easily tune the engine to consider these positions, and it will solve puzzles like this almost instantly... but its overall playing strength will suffer tremendously.
Yeah I was wondering if the mate-in-n engines that try to find the forced checkmates in the endgame would handle this easily or not. I'm surprised Stockfish's NNUE didn't help out here though.
@@CoalOres NNUE's (and even larger NN's) aren't able ot properly evaluate extremely complex positions like this - exhaustive tree search is the only way to find a solution
Why not automatically retune to look for easy wins?
lichess on 74 depth couldn't solve it (the move wasn't in top 5), but it immediately sees #11 after the first move
Interestingly enough LC0 sees the solution almost instantly.
@@1992jamo does it have the position in its database, or does it figure out the solution?
Chess: Why do Jewish, Hindu & Christian players do better than Muslim players?
So far the top players of Muslim / Arab countries I've seen either are ethnically Jewish or have changed to federations of Christian/Jewish/Hindu countries.
@@1992jamo
Yay Leela!!! Stockfish is a TALENTLESS patzer who lost to Leela 3x classical in a row in the 2019 WFRCC finals lol
Good guys vs Bad guys
- Axis of Evil (Good guys?)
wesley so, hans niemann, bobby fischer, sergey karjakin, vlad kramnik, ian nepomniachtchi, alireza firouzja, fabiano caruana, levon aronian, viktor korchnoi, pal benko, mikhail tal, amin tabatabaei, gukesh dommaraju, maxime vachier-lagrave, nodirbek abdusattorov, juan manuel bellón (anna cramling's dad), vidit gujrathi, andrii baryshpolets, vishy anand, jan-krzysztof duda, vladimir fedoseev, lichess, walter tevis, max pomeranc, chesscube, Sergey's hypothetical pure 9LX parallel, PCAP, 9LX, christopher yoo, gata kamsky, sam sevian, michael adams, leela, ATLA netflix fire nation & sokka in original, ferdinand marcos sr, ben johnson of perpetual chess podcast(?), elon musk(?), nikhil kamath, bibi netanyahu, both vladimir putin & volodymyr zelenskyy (because of sergey karjakin), ben shapiro(?)/alan dershowitz(?), mustafa barghouti, barack obama, donald trump, tristan & andrew tate
🇺🇦🇮🇱🇹🇼🇮🇳🇺🇲🇬🇧🇦🇺🇯🇵
- Allies of Evil (Bad guys?)
magnus carlsen, hikaru nakamura, garry kasparov, daniil dubov, veselin topalov, liren ding, aryan tari, abhimanyu mishra, tigran l petrosian, tigran v petrosian, paul keres, efim geller, parham maghsoodloo, pranav venkatesh, sébastien feller, vincent keymer, paco vallejo, pentala harikrishna, arkady dvorkovich, peter heine nielsen, richárd rapport, alexey sarana, ches***m, fred & josh waitzkin, PCA, FIDE, NCFP, chess1, andrew tang, alejandro ramirez, sam shankland, nigel short, stockfish, ATLA original fire nation & sokka in netflix, prospero pichay jr, lex fridman(?), mark zuckerberg(?), sumner redstone, ariel sharon, ali khamenei & mahmoud abbas (because of alireza firouzja & randa seder), alan dershowitz(?)/ben shapiro(?), bassem youssef, joe biden, george bush jr & sr, piers morgan
🇷🇺🇵🇸🇨🇳🇵🇰🇳🇴🇮🇷🇮🇶🇰🇵
- axis of evil (female) (good girls?)
tingjie lei (or paikidze), janelle frayna, anna cramling, dina belenkaya, rahma zein, randa seder, lotis key, carissa yip (or sara khadem), irina krush, heather walker, robbi jade lew
- allies of evil (female) (bad girls?)
wenjun ju, bella khotenashvili, alexandra & andrea botez, abby martin (so much for XMCr-bQk_DM ft mike prysner), turnabouttess, leny so, atousa pourkashiyan - nakamura, jennifer yu, dani chambers, nemo zhou
- maybe slightly axis:
agadmator but idk antonio radić kinda likes magnus. also roderick nava who replied to me.
- maybe slightly allies:
gothamchess but idk levy rozman kinda likes wesley. also deniel causo who didn't reply to me.
Note: They correspond.
Eg
Wesley Vs Magnus
Bobby Vs Garry
Hans Vs Hikaru
Christopher Yoo Vs Andrew Tang
Dina Vs Andrea
etc
More info:
FlNsFKeq-iE
@@nicbentulanconsider changing your marij**na deliver person. Quite sure they are providing you from their long expired stock 😂
So Lichess at depth 70 couldn't find bishop to c7, and you gave me 5 seconds to do it??? Well played Antonio... well played.
Hahaha it's a joke probably but I hope u have a pause button. 😅
come on, don't be so weak. I found the 12 moves solution in 3 sec
Confound it all, if only it weren't for that distracting picture of Topalov I'd have foud it at once!
😂
Cool. Story.
9:05 in case anyone didn't see why the h pawn can't move, it's because it gives up the g6 square. And Kg6# covers the newly vacant h7 escape square. Edit: I spoke too soon! At 10:28 agadmator covers this.
Moves provoking zugzwang are the hardest to find.
Seems so - not only do you have to calculate advantageous moves for all of your candidate pieces, but you also have to calculate disadvantageous moves for all of your opponents pieces. Lots of calculating! 😵💫
Not sure if Agadmator has done this type of video before but already looking forward to more of these puzzle videos 🙏🏽
@@zooniba his Brother in law does show a lot of short videos with similar chess puzzles but also videos of chess engine games Really good Channel I definitely recommend him Jozarov Chess
he has some puzzles on the channel. He doesn't do it too often but when it does, you know is gonna be a very cool one. Look it up
@@michaelmassaro4375 my super engine cant solve this.
Bless Agad for thinking I can find a mate in 12 when I didn’t find mate in 1 yesterday
This was actually beautiful to watch. It reminds me why I love chess. It is never a straight forward game regardless of how much theory and computers advance. One new move and it opens a whole world of uncertain possibilities! Thank you for the video
I saw the idea fairly quickly and thought I had solved it after 4. Ba1, but I missed all the other nuances, and I had no idea how deep the position was until I played through your full explanation! Thanks for sharing this.
Its a great puzzle because it has all the elements: The obvious move isn't the first move. There is a queen sacrifice. The obvious move only gets played at the right moment. A zugzwang becomes evident. A pointless pawn push ensues just to use up moves to complicate. An inevitable discovered check forces black to sacrifice material until he's mated.
Thanks, Agad's brother in law!
I ran the puzzle on Stockfish 16.1 and once you do the first bishop move, it then suddenly goes from "draw" to mate in 11. Crazy!!!
Beautiful puzzle. Seeing the Queen sacrifice & understanding zugzwang idea is very hard indeed.
FWIW, Crystal sees Bc7 instantly, even on my phone. For those who don't know, Crystal is an engine derived from Stockfish and is really good at solving studies and compositions.
If you let your engine analyse just a single line (PV) it wont find Bc7 cause of heavy pruning at the start of the search tree which is done mainly for speed. If you use the Multi-PV option Stockfish will find it. Ive played around with it a bit and SF wont find it up to Multi PV =5, but when you use Multi-PV = 6 then it finds Bc7 in about 20 seconds.
Set threads=16, multipv=10, hash=4096 and it finds the solution in 3 seconds.
@@lockaltube Multipv = 10 is just forcing it to try about every legal move. I tried 10 first too and then tested what the lowest was to find it... which is 6
This is just Stockfish's "Late Move Reductions" at work. It's not solved by using higher depth by itself because Stockfish just prunes the branches where it sacrifices a queen without compensation in the next 10 moves and only calculates the 'non-terrible at first sight' moves to the full depth.. Purely depth-wise (without pruning), it would find the mate at depth 23 (since it's a mate in 12). The reason it does find it eventually, is that at extremely high depth, as a "side effect", it's also going to less aggressively prune the first few moves options. The proper way to computer-solve puzzles like this , is to run Stockfish with specific parameters that turn off the aggressive pruning (which is worse for general play, but much better for solving 'counter intuitive' puzzles)
I gave this to my own copy of Stockfish 16.1 running on a Linux computer. With my default setting of 4 cores on a 4GHz machine, the evaluation went to mate in 12 on ply 62, consuming a little over 17 seconds.
Excellent puzzle. Just to put it out there though, stockfish cannot solve it at depth 64, but was able to get it right at depth 65.
Excellent puzzle, thank you
Your Brother -in law Jozarov showcases a lot of Great puzzles as well as being the Best Engine Channel on you tube Enjoy your day
@@michaelmassaro4375 I’ll have to check that out. I’m an absolute puzzle whore
What I find interesting is that if you put in Bc7 the engine suddenly sees the win. When I first put the position in an engine, I forgot to add the a5 pawn, but it still thought that the position is a draw, until depth 36, where the solution is also Bc7. The a5 Pawn makes the position so much harder for the engine.
yeah because without the a5 pawn it is not mate in 12 (was it?) but mate in 8
It's because the engine doesn't even check the move because it appears to lose the queen. The moment it does check the move it finds this immediately. And if you use brute force full depth analysis then it also finds it immediately.
Shout out to Gauri, he posted this puzzle many days ago . Thanks Agad for testing my memory.
Only Tal could find such moves over the board
I'm not sure if I've got this right (and I usually do get these things wrong!), but at the position at 9:10 , can't black play Qf8? The white king can only check black by moving away from the defence of his pawn, but then black can block check with Qg7. If BxQg7 and then KxBg7, the white pawn can be picked up by the black king, which can then defend his own pawn on the next move - and so white loses. If, after Qf8, white plays a waiting move, say Bb2, black has Ne8+. Although the white king can then move to a square defending his pawn (e6) while checking black, black then has Ng7+ blocking the check with a check of its own. Again, white loses, because the black pawn can then move after white's next move and so there's no more zugzwang, and white can't stop the pawn marching towards queening without losing both bishop and white pawn.
after Qg7 white promotes pawn to queen.
@@sandsand2515 ah - you're right. I knew there had to be a mistake in my analysis somewhere. Thanks!
Will show this to my uncle at the bar... or library.
depth65: mate in 12: 1. Bc7 Qxc8 2.gxf7+ Kh8 3.Be5 Qc5 4.Bb2 Nc7 5.Ba1 a4 6.Bb2 a3 7.Ba1 a2 8.Bb2 a1=Q 9.Bxa1 Nd5+ 10.Ke6+ Nc3 11.Bxc3+ Qxc3 12.f8=Q#
8:37 i was thinking Nd5+ Kf5/g5+ Nc3+ forces white's king away from its pawn but white can instead do Ke6+
have you tried feeding the position to the same engines after x moves have been played? just to see from where they pick up the solution
Lichess gets it after you put in the first Bishop move.
oh okay. ty for the info
Probably because engines are build to discard possible moves using a zero window search and depth reductions. This allows them to look down the important paths at a far greater depth in most common positions, but apparently this queen sacrifice is just too deep to for the engine to search deeper. Very interesting :)
This position also relys on zug zwang a lot, so maybe the engine discards these lines after a null move search early on...
Saw this same puzzle on Gaurichess, beautiful
same. a couple of months ago
It Proves computers can't reach human madness. Also Topalov's photo reminds me of his game against Shirov where he moved Bb3 to Ba2 after bxa4
Computers can, with the right settings.
@@thuroria7631 but the settings need to be set by humans
Wow. What an elegant puzzle.
I solved the solution up to the same point Agadmator did, but it wasn’t until I was told Ba1 didn’t work that I found Bb2. Great puzzle!
What are you talking about. On my phone I have SF_NN_12 and it quickly finds 1. Bc7 as the best move (depth 30 it sees that it is mate in 12 and it took less than 1 minute. And from then it switches eval between 0.00 and mate in 12 a few times, no idea why that happens. If you set up the board with the Bishop already at c7 and black to play, funny enough it takes only 1 second to realize that it is mate in 11 with best defense.
Wow, incredible puzzle. Left speechless.
Stockfish 16 on depth 83 couldn't find it too. Amazing puzzle, I'm impressed.
Stockfish 16.1 in Arena can find the mate in 12 in just 7 seconds with depth 20/21. Just turn on the Multi PV mode to 10.
Edit: LC0 and Sugar NN found it too in 15 seconds
After setting Multi PV mode to 10. Stockfish 16.1 found the solution in less than 1 second on my end XD.
@@kuei4604 we should swap computers 😅
One big question that didn't get a solid answer: Why can't Black play h6 or h5 at any point? 🤔
Answer: Tempo turned Positional. The Black Queen needs to be on the 6th rank for h6/h5 to work, but after Be5: It was impossible... in 3 different way!
After bishop is in the long diagonal if you play h5/h6, you give one square to the white king which results in M1..
@@odysseasv7138 Replying to bump this up, I'm sure many people didn't notice that
@@odysseasv7138 Exactly: h5/h6 Kg6++
I had this solved at 2:30
I had increased the number of variations to 8, and stockfish 15 at home was able to find the solution in about 40 seconds. I think it just never considers the queen sacrifice in the first instance so it just calculates the drawn line deeper and deeper --- it truly misses the creative spark that humans have!
White has only 25 possible moves, so you have a 4% chance of finding the winning move by blind guesswork. There are 3 for the pawn, 2 for the king, 4 for the bishop, and 16 for the queen (7 vertical, 2 horizontal, and 7 diagonal). *Edit:* This is not to say that this is a simple puzzle to solve, or see the final solution clearly at once. Rather, it shows that even under considerable constraints, there are still myriad possibilities. I was just saying that it's possible to get the first move of the sequence right by sheer luck, and the odds aren't as long as they are in some other positions.
But each of those 25 possible moves have other endless variations once the move is made.
@@malsawmzela609 Very true. I wasn't trying to suggest the whole sequence is easy. I should have made that clearer. I find one thing especially interesting, and it just shows how complex and free-wheeling chess is. I said there are only 25 possible moves for White in this situation. Yet, even though White is down to a quarter of its starting material, and its king is sharply constrained by Black's pieces, 25 moves is still 5 moves *more* (or 25% more) than are possible at the start of a game.
I analyzed the position on my (modest) laptop (2.6 GHz base speed, 8 GB RAM); it took about ten minutes to reach depth 72 and then after ten minutes it jumped at once from 72 to depth 84 with the solution. Overall, it took about 22 minutes and 1.4 B nodes to find the solution. The Stockfish version was Stockfish dev-20240723-b55217fd. I would say it isn't a big thing. It takes time, but the engine does find it.
One of the greatest puzzles I’ve ever seen. I was thinking pawn g7 then queen b8 able to line up with the dark square bishop by going queen d6 and go from there
LC0 finds the solution pretty much instantly within 100k nodes.
You should have mentioned in the notes the name of the composer of this endgame study.
Great puzzle bro
Dear @Agadmator #Agadmator #submission #correction You say that one Dutch composer's first name is "Istvan" in this video, but his first name is actually "Gijs." I think you are mixing him up with Istvan Bilek, the Hungarian.
Source: ua-cam.com/video/8wCJalNkTEI/v-deo.html Agadmator's other video, haha
Why not bishop to e7?
Beautiful! ,stuff
Stockfish 16 plays gxh7 and rates it 0.00, but as I play Bc7 it starts calling it -M11. It doesn't find the first move, but if started from the second move it does perfectly fine i.e. mate in 11.
Oldie, but goodie. I love games where engines fail to solve it. I showed this puzzle to a friend who argued that Stockfish today can solve anything in seconds. Whoops!
I think the reason why engines can't find it is because they don't actually consider every possible move. They reject 'bad' moves and don't even consider them anymore. I've sat here and watched Fritz 19 analysis the position for the past 2 hours. There are 25 possible moves for white and Fritz only spends time calculating 2 of them, gxh7+ and gxf7+, the rest it skips over. So even though it's up to a depth of 42, which is deep enough to see a mate in 12, it's only considering 2 moves at that depth. Unless it goes back and looks at the other 23 moves, it will never find the mate.
I'll let it run for a while just to see what happens, may even let it go overnight.
Do engines go back to less likely moves in earlier move numbers, when all their other lines yield nothing?
Not me always eyeing my queen for a sac on every pause the video moment that the queen is still on the board
Glad I didn't even give that one a try, that would have been hopeless
I stared at this position for 20 minutes and gave up. Then, just before the solution was given I spotted the move.
Thats impressive whats your elo?
@@BREAKocean Tournament/serious 6 man board 1 team play was around 2350. Online more than 10 years ago in 10 minute was 2650.
I gave up trying to be serious in my teens when I played with people who I could not hope to match without devoting my life to the game.
The glory of chess is one can explore exciting new worlds from the comfort of one's living room.
Bc7 Qxc8, gxf7 Kh8, Be5 as if white king moves it threatens checkmate, and if h6, the white king moves to g6 mate, that's my guess. Plus after Be5 Qc5, white should not moving king but Move Bishop somewhere else, Like a1, as the immediate king moves allows white to trade the bishop and allowing blacks Kg7.
I made it Ba8, I hot lazy a didn't consider Nc7, but only took about 5 minutes to get that far, plus releasing gxh7 doesn't work also made it easier.
just needed "couple of seconds" to beat depth 70 lichess on this one.
If you remove the black pawn from A5 (thereby not having the pawn moves at all) the engine solves it.(Mate in 8). Crazy. Similar set of moves. Starts with the same bishop move.
on depth 60 stockfish15.1 unable to solve it my mobile was so heated😅, then i played 1st move and suddenly mate in 11 was recognised by engine😊
#suggestion Firouzja vs. Caruana Crunchlabs Masters match #2 LoL Egregious opening blunder by Firouzja
I'm more surprised that engines cannot solve this. What are they getting stuck with?
Makes me question engines 'way of think'.
Are they not considering what looks like a suboptimal move, unless they see the victory, so they discard the first move as they may not be able to calculate the victory for 10 moves ahead.
Engines don't fully calculate moves to forced victory, but rather moves that increase chances to win.
Please someone correct me if that's not what happens. Very interesting video.
@@wilkinsune Agadmator’s Brother in law Jozarov has a great channel on you tube he showcases a lot of engine games Stockfish Leela C0 and other engines Great Channel
Selective search.
Stockfish is extremely complicated these days, for whatever reason, the branch that lead to victory was pruned too early in the search tree.
Stockfish 16.1 (version 2.3.7) came up with answer in seconds using infinite analysis on my intel iMac Pro 2017 3.2 GHz 8-Core Intel Xeon W 64GB 2666MHz. Stockfish says Depth=86/24 8169 kN/s
I'm using Stockfish 16.1 and it found M12 at depth 48
Excellent puzzle ..I tried Stockfish 16 and Lela 0.... Lela 0 solves it quickly
Stockfish took 2 minutes
@@hisdudeness690 Stockfish 16.1 onwards solves it fast, but stockfish 16 which is also what Agad used can't find it.
This was just epic
Stockfish 16.1 could not find the move, Komodo took like 10 seconds and found Bc7 +M12
16.1 can solve it at depth 65.
That's beautiful ❤️, tks
Anyone knows how can actulally think of such moves i was thinking of Be7,Bc7 didnt even appear to me
Stockfish 16 just solved it on depth 79.
Believe me, i did find the correct idea of mating the king on corner with that bishop. I did figure out the bishop move and then capturing the f pawn. But didnt really get upto the end
👏🏾👏🏾👏🏾👏🏾👏🏾 I couldn’t.
The reason Stockfish can't find it is simply it's pruning process, no?
that was insane
That's the best puzzle I've ever seen in my life.
This such a beautiful puzzle, I stopped thinking after setting up the idea, but didn't calculate deeply enough and fell for the same trap as Antonio😂. The pawn on a5 is the true mvp of this puzzle, without calculating forcing moves out of instinct I wouldn't have seen white's resource, but without it black would have lost sooner. I'll be sure to brag about solving half of this to my friends at the bar and the library
Being a fan of Danya Naroditsky.
He solved this puzzle and being a agadmator subscriber I can solve this eyes closed
I think i saw this in daniel naroditsky speedrun
Yeah, me too. I had to scroll for a while to find someone pointing this out. Thought I was crazy for a second.
Awesome⚡️‼️
Normal Stockfish 6.1 with MultiPV=10 (yes, it is a common idea to increase MultiPV for puzzles) on 16-core CPU after 3 seconds sees mate in 12
It's not about the depth of the calculation or the power of the computer. The engine has already decided between 2 moves (overlooking another move that is the winning move). From this point on, the engine won't go back in the decision. Ironically, if you use a old brute force version of Fritz, you probably solve this using a Penthium II
Lichess on depth like 95+ was able to find it
I didn't know who Topalov was, and I thought you had a picture there of your brother-in-law in honor of his contribution lol
Very, very nice!
The closest counter I've got for black at 8:55 is Qg1
I also asked Stockfisch. After about 3 minutes he had the solution (AMD Ryzen 3950x with the setting to use 8 cores out of 16). I then tried it with 16 cores and was very surprised! Stockfish didn't need half the time but twice as much (6 minutes). Obviously a bug in Stockfish...
Not a bug, stockfish distributes tree branches in a way to be _the best chess engine,_ not for solving puzzles. Imagine you have a tree, so that in the start it splits in 2 branches, then these branches split in 5 or 100 branches and so on, and you never can see the end of the tree, you can maybe see the huge number of branches nearby. And you have N workers, and your task is to schedule workers to saw this tree for logs (and potentially find a golden branch). Of course you will process the tree differently depending on the number of workers (each worker will process every Nth branch, N changes as avalanche)
Well, as I understand it, the skill is to estimate which branches could be promising and to investigate them more intensively. (i.e. to search in depth and not in breadth). Anyway - I did a few more tests with different chess engines. Some do not find the solution at all (like LC0, Plentychess, Komodo and Dragon v1). Stockfish did not find the solution in version 14.1, but in version 16.1 after 40 seconds and in the latest development version after 143 seconds. Very astonishing in this test: ShashChess -> whether in the old version 14.1 (from 2021) or in the current version 35.3, the solution is found in less than one second!
Tal: "I'm going to have a cigarette in the park. Be right back."
My first thought was to capture the pawn at A5 by Bishop
Wow what a puzzle
So its all about a 1 move tempi which is impossible to see 5 moves ahead.
This shows that man is still better than supercomputer 😅
No it was not supercomputer
Analyze with Leela immediately spots it
Chess is so hard but congrats Agadmator that he did job of Super computer i think only Alpha zero can find it within seconds 😊
@@ehglt I think Stockfish is actually stronger than alphazero
Stockfish is much stronger than alpha zero now actually
That's a curious quirk of the engine that it can't evaluate the position. That said, as soon as you move bc7, the engine figures out the rest immediately.
I find these puzzles.. but fsil to solve mate in 1 😂
Οf course, we all found the 12 moves solution 😁
Tal would've found it for sure, it has a queen sac after all
Agadmator for president
As soon as I saw the puzzle my first thought was Bxa5