What You Get Wrong About Chess Engines

[@wasp7969]

this is the most underrated chess channel

[@ChessDojo]

Thanks! 😊

[@turul9392]

Absolutely great channel. I use the engine only after spending at least an hour on the game myself and taking notes.

[@ericwagner7747]

another absolute banger kostya - nice job

[@mcronrn]

I use 3 engines to learn my openings - they regularly give 3 different answers at move 4-10, at which point I pick the one I like the best 🤷🏻‍♂️

[@ajarnray4115]

Eye opening video I had an engine tell me a fortress was a sure win for the other party. When I played it against the highest Lichess SF engine it couldn't win at all and that was the first time I noticed our silicon buddy also has senior moments :-)

[@ChessDojo]

Great point as well! Engines definitely struggle with fortresses

[@laurentsaltoflife9267]

Nice job. Would you consider a follow-up on "unplayable opening lines at amateur level" (because SF gives +1) ? :-)

[@patrickdaly1088]

I have an 80% win rate over ~100 games with white in the Pirc Defense: Chinese variation, which I actually reach by transposition from 1.d4 Nf6 2. Nc3 g6. People who want to play the king's indian defense, don't feel that Nc3 challenges the setup, and the Pirc Chinese Variation is considered to be inaccurate play for white by stockfish, giving up the entire advantage of having the white pieces, for an equal position. But KID players don't want to play d6-d5, they're playing the KID, they want to play e5 so their LSB is good. "Bad moves" = 80% winrate. Best moves = my opponents actually know the position. I actually had an IM go back and forth with me on the position being not so difficult to play, open up SF to refute me, then miss my ideas completely and be surprised at what happened, because I didn't play the "top moves" and got a better position somehow lmao. Stockfish's taint is pervasive.

[@40inchvintage]

@@patrickdaly1088 Which Stockfish version did the IM use?

[@patrickdaly1088]

@@40inchvintage If you think that matters, I feel like you've missed the point of this video and my comment completely.

[@ishanr8697]

What do you do if they go 2...d5?

[@patrickdaly1088]

@@ishanr8697 If I'm feeling spicy in 1 minute or 3 minute, I premove e4 and we transpose into blackmar-diemer gambit. If not, maybe a jobava london, maybe I move my knight and play c3, or c4, play for e4 later, maybe I do something else. d5 is the "refutation" but it's just a normal chess position where many plans exist.

[@pedrodaffunchio675]

This was very useful. I would love some video showing the proper way to use an engine. Thanks for the content

[@patrickdaly1088]

Thank you so much for this!!! This will actually be the video I link to people who worship the engine too much for the forseeable future, bookmarked. I'd also like to second the request for "Why Stockfish's best move is losing you games" AKA 'practical chess.' Mamedyarov's video series on his best games on c.c, literally every game he says "Engine says to play this but this is losing for humans."

[@ChessDojo]

Awesome, thanks!

[@MultiMarcd]

Great job Kostya. The worst thing with that, it's the fact that it's so easy to always believes in the engine. Luckily for me i still play sometimes the Bronstein variation in the Caro-Kann even though the machine don't like that. And in the endgame, the evaluations sometimes are so far from the reality.

[@Life-Sky]

Many times I analyze with engine on lichess for example, and see why a move is innacurate/mistake, I ask myself: 1- Is it too advanced? Many tactics or positional things are just too advanced for me to focus right now, I would be beating myself for something that I would have never found and barely understand with the engine on. 2- Is it something I'm working on? Maybe I say: "Oh nice idea" or "Oh i missed this tactic", but you can't improve everything so I select one or two things to take from the analysis and go from there. 3- How can I use this to help me think next? Trying to generate a concept or something to look for (I didn't look for checks or I was playing too passively) so that I can think about it from a general perspective and not just "Oh i didn't play X which is the best move"

[@chriscoski3233]

Such a timely post! I was just talking with the younger players in my chess club about this at our weekly meeting last night. I'll definitely be directing their attention to this in the hopes it'll carry more weight coming from you!

[@ChessDojo]

Awesome!

[@vishnufafat]

Can you also cover engine andys who don't play an opening line because its only +0.2 instead of +0.5? :D

[@zenchess]

Kind of hilarious really when low rated players think they know more than famous chess authors because their depth 18 engine told them so....:) Then again, with a lot of the classic books there are actual analysis errors, I guess it just takes a stronger player who can properly understand how an engine works and is able to judge when the engine is doing a worse-than-human defense in order to survive as many moves as possible.

[@epictetusofhierapolis4461]

I'd add further that even those cases where there were actual analysis errors, even from a human standpoint, the analyses still have instructional value and should be taken and internalized as such.

[@Life-Sky]

@@epictetusofhierapolis4461 GM Daniel Narodisky did some book puzzles on youtube and was able to see ideas that the author didn't consider but that the engine considered best move, also the engine would see a very instructive plan that neither the author or the GM considered but was still a very simple straight forward idea.

[@epictetusofhierapolis4461]

@@Life-Sky That's fine Ivan. But I don't know if you're trying to make a point regarding my previous post, or just stating that engines can be helpful.

[@letsmakeit110]

yeah almost as funny as when 2500 elo gms think they know more than 3200 elo SF at depth 22 If anything I find most of the material on chess goes the other way. Stop using your engines. Stockfish is no good it can't be trusted look at this handful of fantasy positions it doesnt understand. But if you actually sit down and spar with your engine, your understanding will grow exponentially. You'll get whopped over and over, but you'll eventually become convinced that Ra2 is better than Rd1, AND you'll understand why. That's the silicon road to chess improvement. If you don't trust me, maybe you'd trust GM Matthew Sadler who wrote the book, or GM Larry Kaufman who wrote Komodo.

[@letsmakeit110]

I wrote a program to scrape 60,000,000 unique analyed positions from lichess games. I ask my database "is a knight on the rim really grim?" ofc it says yes average -0.40 every coach will tell you that. But I ask it plenty of other things too. Like if I'm castled with a rook on f1 or f8, and I have to make a recapture on g3/g6, is it better to take towards the center with h pawn, or better to open the rook with f-pawn? Maybe it depends on how many other pawns there are. Maybe how many center pawns specifically? Point is I can ask my silicon coach any positional question I want, if I know how to code for it.

[@slowslav6653]

The last example is actiually pretty funny :)

[@danielguel1897]

Wow, great video. Lots of people need to hear this ;)

[@SLAMBANGO]

Stockfish on my phone is not that strong enough to beat you every single game at 5 seconds per move regardless of how much time you get.

[@dsrguru]

Great video! It validates my recent thinking that all sound moves (moves that preserve either a decisive position or a holdable one) are objectively equal, and it's just a question of what moves introduce paths through the tree of possible continuations that pose your opponent problems that they might not solve.

[@TuringMachine001]

Yes! Mathematically either the position is winning for white, winning for black or drawn. Everything else, including so-called "objective" engine evaluations, is just heuristics. At the end of the day, practicality for humans is what should guide our choices between mathematically equally-evaluated moves.

[@firerook]

If you play a weaker engine and get a winning position there always a place when they play a dumb move (objectively better but not the critical move). Give a free queen in a non-critical position to avoid a crazy hard to find mate or give free pieces to get a mate in 7 instead of a mate in 6.

[@macnolds4145]

I must disagree with some of what Kostya is saying. The issue is not that engines "sometimes don't understand" a position or that they "sometimes get things wrong". The issue is that the goal of chess is simply to win, and a search for "understanding" is how humans cope with complexity. It is not an objective truth about the nature of the game. (For example, it is not objectively "true" that one should try to control the center squares; that bit of advice is just how humans must think due to the alarming complexity of chess). In short and simple games, humans can robotically memorize the best sequence of moves to play and reliably achieve the appropriate result. However, when a game has too many possibilities for such a tactic, humans develop heuristics, theories, and strategies in an effort to not only play reliably and consistently- but to also have a means of explaining why some things work and why other things fail. The only "truth" in chess (or any game with similar or greater complexity) is that the outcome of sequences of moves is what makes a move bad/good. In fact, often times one should disobey ideas such as "don't move the same piece twice in the opening", or "castle early", and so on. No entity- man or machine- knows the "truth" of chess, but I contend that AI computers are actually closer to this unattainable goal than us mere mortals, as they lack any foundational biases.

[@SLAMBANGO]

In the short term you are correct that computers can find more accurate moves quickly. But you put a grandmaster and a computer together to analyze over time and and the GM can come much closer to the truth than the computer or either one alone. Computers still have blind spots, mate in 12 they cannot see it at all and humans see it instantly... Especially when the sea coins begins with a Queen sacrifice

[@macnolds4145]

@@SLAMBANGO Humans will never find the kinds of lines that involve a series of moves that individually betray the collection of human wisdom about chess, and that only work because the end of a long, annoying sequence leads to "advantage". Our powers of calculation are limited and we can't play sequences that don't have some kind of clear narrative (or logical explanation for each individual move). We always want a "why" that goes beyond "it works", but sometimes the only answer is simply, "because it works".

[@Alex-fu6ip]

Good callout, I’m working through Krammnik’s Chessable course, thinking in chess right now and he regularly says much the same thing, citing practical motifs where an engine line is less relevant. I recently started messing with Lc0 and had it play itself several times from the same position. Though the %value of success is more intuitive one cool thing is have it play itself from the same position several times and it will get different results, sometimes even losing. There is also one more thing to consider: having engines evaluate endgames is very dubious from a practical perspective. They run it down to the last legal move before resigning. IRL someone likely resigns or agrees to a draw.

[@jonathanhenderson9422]

The day I realized that chess engines (especially on browsers) were, let's just say, "limited," is when it gave me a double-exclam "brilliant move" on literally the easiest move to find in a bishop-sac checkmate sequence rather than the move that it actually took me time to find to JUSTIFY the bishop-sac!

[@erickt22991]

chess players need to learn out this channel. These guys are great.

[@marcofrey2903]

Great points here. That’s just the tip of the iceberg. There’s also a lot of nuance around “practicality”. If you play an engine line, you also have to understand it on a deep positional or tactical level. Otherwise you’d be better off playing a decent move you actually understand. And then theres the issue of practically based on one’s own style as well as the rating range they’re up against. Heck, even the Elephant gambit could be crushing in the 800-1200 range. That’s what aspiring players should understand. To be fair…many of them do 😂Gambits are more popular than ever.

[@babstra55]

this is something that's bothered me for 20 years as both a computer scientist and chess player. giving something a number doesn't make it true. losing to engine doesn't mean it's right. there's a huge difference between strength and being right. ask any dictator or a bully. a great way to cure yourself from engine religion is to watch some jonathan schrantz gambit videos where he draws and even wins stockfish with questionable openings. in some of them engine thinks it's winning by up to +10 pawns, yet draws or even loses the game. and the question you should be asking yourself is "how is it possible the engine evaluated it was a queen up or more, but 5-6 moves later was lost?" and the answer is NOT that it didn't look deep enough, the engine will get 5-6 moves deep instantly. also the fact that it'll completely flip its mind later isn't any guarantee its correct then. it should make you MORE suspicious, not less. people confuse the truth of the position with engine's ability to crush the best players in the world like they were little children. that just means it statistically makes less fatal mistakes, it doesn't mean the engine understands the position. it doesn't even mean it makes no fatal blunders. then there's ofc the other side of it, that it makes no difference what the engine says when you're playing a human. even if we give it the benefit of a doubt and assume it's absolute truth, the advantageous position it gives you often requires 4000+ rating play to handle. can you do it? obviously not. can magnus do it? obviously not. nobody can. and if nobody can do it, what does the evaluation mean even, if in some academic sense the win is there? and these are all very crude multiple piece blunder evaluation errors. it's of course exponentially worse when the engine gives you +0.4 vs +0.2 to choose from. it means nothing. and even if it does, there's no way to find out except competent human analysis over years and even decades. and still the +0.2 can in practice produce 20% better winning chances. use them for pointing out things. then if it makes sense TO YOU, okay it matters. but if you can't deduce a CONCRETE justification for the engine line, it doesn't mean anything. you will not be able to find the moves if you don't understand why. be especially conscious for not letting it go with nebulous abstractions like 'the position looks good for white'. if that were enough we wouldn't need the engine in the first place. the strenghts of the position must be concrete, or you're just fooling yourself. and it's SO easy to fool yourself when you know the engine can crush any human with piece odds. but exactly because of that crushing strength, you should also understand that less than piece differences in the evaluation can't be trusted. the truth of the position is obscured by the crushing strength.

[@jamesknapp64]

I'll say similiar to what I said in puzzels comments. Engines are good at identifing blunders and why its a blunder. If youre under 2300 if the eval difference between moves is less than 1 point they are often of equal quality. Like I said in the puzzle comment there was a puzzle I got wrong where the correct sequence was give a check with the Bishop and the computer line gives up the Queen rather than moving out of the check. Thus the computer line says play a position where its a side is down a Queen for a Bishop, a situation that is basically resignation for humans. The reason being there was a 4 move checkmate if the King moves out of the way (taking advantage of a pin I failed to see). Computer sees the 4 move checkmate pattern. I doubt most sub 1000s see that checkmate pattern. IMO it be a better puzzle if it was the same position but started after the King moved out of the way forcing you to find that 4 move checkmate , since no human in slow time controls is giving up the Queen like than, they either move the King or resign. Overall grreat video.

[@chessforfunonly1586]

When I misinterpretated what Stockfish depth 20 tells me, it's still way more accurate than the 3 ply blunders I come up with! LOL

[@Extirpo]

Great video! Btw in Tal/Spassky I found Kf5. Am I a chess machine? 🤔

[@NidusFormicarum]

In general I agree. However, these computer moves that just throw pieces away to avoid mate are only irrelevant from the attacker's point of view, not from defender's! What do I mean by that? What I mean is that it is still important to be able to caculate these stupid looking give away moves to avoid mate - to have the ability. Would I continute in this case? Of course not, I would be a queen down for nothing. But just change things a little bit so that I had another dangerous past pawn and another rook and the defensive idea might had workde there. In that case I woudl play on even if I saw a forced mate for my opponent in a couple of moves, since in that case I would have somehting to play for and to hope for. So... What I want to say is that it is a matter of interpretation and that "irrelevant" moves and lines might be relevant depending on your point of view. But that seems to be one of the major point of your video anyway. :)

[@jamesehrhart9364]

From a logical standpoint the engines will never be perfect. Because chess can be considered a complete logical system we know that there are lines, moves etc that are winning that can never be calculated.

[@samuecchi]

13:22 posted on my birthday :')

[@Chess_Improvement]

great video, we can't play like engines why rely on it

[@berndmayer3984]

In some positions, the best move is not recommended. It could be that it only prolongs your agony, while the second-best move only loses clearly if the opponent finds the difficult refutation, otherwise you stand to win. The engine does not know human champagne or seltzer.

[@JohnSmith-ck5qk]

fun fact the true fried liver (Nxf7 sacrifice) is considered equal, and stockfish and leela draw each as black in that position, but that doesn't mean humans playing the black pieces should actually play the fried liver though

[@ishanr8697]

Engines think Fried Liver is better for White after Nxf7 (despite being a piece down) but say that d4 is objectively better. This seems correct to me. Leela, for example, says 1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5 Nxd5 6.Nxf7 Kxf7 7.Qf3+ is 55% expected score for White. r1bq1b1r/ppp2kpp/2n5/3np3/2B5/5Q2/PPPP1PPP/RNB1K2R b KQ - 1 7 After 6. d4! r1bqkb1r/ppp2ppp/2n5/3np1N1/2BP4/8/PPP2PPP/RNBQK2R b KQkq - 0 6 Leela claims 61% for White if 6...Nxd4 and >67% for all other continuations. Seems reasonable to me. I would try 5...Nxd5?! as Black if it weren't for 6.d4! It's a great way to learn to defend. 5...b5!? 5...Nd4!? and 5...Na5! are all great moves to help you learn chess too.

[@spacerocks9740]

I have a question whenever I'm the opening I pin the knight with a bishop and the opponent try to play h3 to run my bishop away from the pin the best move computer shows (most of the time)is to take the knight and lose the bishop pair and I do not understand why ,the reason I do not understand it beacuse after taking the knight nither his pawn structure weakens nor create any threat but I still wonder why engine suggest to lose the pair of bishop for no reason

[@Heroball299]

Don't just automatically pin the knight with the bishop. Unless you plan to swap it off. It's sometimes offside after h3 etc

[@nikamumladze8220]

great video

[@westsidebilling]

Thank You! Letting a machine do your thinking for you is not wise. The engine is a guide, NOT a gospel.

[@mitchellfabian7694]

Hallelujah

[@kwhd559]

Yeah it's annoying when casual players use an engine, and then criticise a stronger player's judgment based solely on some numeric evaluation they don't fully understand to begin with - For example in the course I just created for the Chessable platform, I provide practical solutions to a whole host of popular internet Gambits. In my chapter on the Halloween gambit, a user informed me that I gave a sub-par antidote and his engine says that xyz is the refutation. Such ill considered comments not backed by any chess understanding are just infuriating, makes you want to swat the user over the head - Here is what got his undies in a bunch. I gave a carefully considered line aimed at returning the piece to takeover the initiative : [1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Nxe5 Nxe5 5.d4 Nc6 6.d5 Ne5 7.f4 Ng6 and the line continues] However my critic informs me that 5...Nc6 is already a mistake, and that the engine considers 5...Ng6 as the refutation [1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.Nxe5 Nxe5 5.d4 Ng6 6.e5 Ng8]. The thing is I had already considered the 'engine approved' 5...Ng6 and determined that for your average player it granted too much initiative for White. It would seem that many average players have little notion of what is practical in an actual game setting, and how important it is for human players to limit the opponent's active counterplay and avoid passive positions. You would think that after so many of them keep losing to nonsense like The Stafford Gambit that they would be able to surmise this, but sadly few do. Instead they keep trying to reproduce optimal engine refutations, that often require sophisticated, and awkward counter intuitive play in order to prove a winning advantage. Go figure!

[@morphykg1503]

We need to have a D.A.R.E program except for amateur chess players to not use engines

[@Evilanious]

This is partially just a failure of communication/interpretation. Clearly the author of 'how to beat your dad' thought the other continuations were obviously losing for white (and they are, if you are a strong enough player) and not interesting to consider, whereas finding Rh8 is hard. But that might not have been obvious to Jim Jones. I don't know how good he is but it might be that calculating those other lines is just hard enough for him to throw him off in that regard, especially if he only looked at the engine lines without trying to understand why the engine would recommend them.

[@pushingyourownagenda9889]

this is soo funny

[@RobBCactive]

Having done many training problems with the benefit of engine checking, I have found many real errors in older material. These would include defences found during analysis, which forced critical attacking moves, where the original author relied on cooperative play. This occurs even in material sold for beginning/improver players who for example are likely to consider blocking discovered check, as well as capturing the checking Q, so spend a very long time looking for better moves only to find the answer given suggests mate in two when the accurate defence maintains a rough material balance in a mid game position.

[@ChessDojo]

Yep, sometimes books have mistakes!

[@Erroll21Oscar25]

@ChessDojo Hi, I am new here, and applaud your video and the purpose of it, very well done! However, the "10 out of 10" endgame may be trickier than you (and I) thought.... After 1.gxf3 ...Ra2 2.Rb1 b4 Black has some serious threats that are difficult to meet... eg after 3.Bd1 both ...Nc5 as well as ...b3 are very good options for Black... White ought to be winning this but with Black activating all his pieces and the initiative with the pawn advances it is very easy to go wrong for White since there are very tricky tactics involved. Say 3...b3 foll. by Nb6 and Ba6 and you can see where this is going...

[@ChessDojo]

Did you use the engine for this analysis? :D

[@Erroll21Oscar25]

Let me start by asking: do you have a plan on how to continue after 1. gxf3 Ra8-a2 ... I was more thinking of how the position after gxf3 should be a relatively easy win according to you in your excellent vid, but after Bl playing ...Ra2 I had a hard time bring this home for White AT ALL, let alone "easily" and was over 5 minutes in thought, decided on Rb1, turned on the old Rybka and let it play the Black pieces.... I had a hard time to put it mildly, and being an endgame afficionado I started digging deep, using the engine as a tool. Wanna continue? My response with white was , forced , 2.Rb1, obviously you can't allow the pawn to be taken with tempo, foll by 3 Bd1 and sliding the King over, Black is calling the initial shots there... If white can steer clear through multiple tactical difficulties he should have the edge converting with the extra pawn, but that is more than 20 moves down the line, that is how complex this endgame "can be" --

[@mikem668]

Not that strong? So can Kostya beat Stockfish at 20 ply on a phone? To me this argument is like saying a Ferrari isn't faster than a Camry because most of us would drive a Ferrari into a wall. *** The question is why Kasparov worried about engine analysis and why so many world class players play like engines. (This is true in Go by the way, which is a much harder game than chess). Finally, how do you think younger players have gotten so good so fast? Engines. And even databases are filled with engine analysis. *** No one is denying that the human factor is important. Magnus is the best example. He's well aware of when you need to calculate and when it's not necessary to find the optimal move. He also plays the clock and the opponent. *** I agree that engines are a mixed bag. But I find it impossible to believe that great players of the past wouldn't care what the best engines thought. Didn't Fischer quote Lasker on lies and truth?

[@ChessDojo]

Main point from that segment was that 20 ply stockfish is hardly going to give you the "truth" of the position

[@mikem668]

@@ChessDojo Agree. But AlphaZero has made me believe that the truth or any truth other than simple tactics or tablebases is beyond us for the foreseeable future. It has suggested that win probability is probably a better measure. Knowing that the win probably for you guys would be different than most of us. * I had a friend, state champion, beat Kudrin twice in the blitz tournament for IIRC the US Championship spot. His chess improvement was in neutral. I kept arguing he should start using an engine and pay attention to the small eval differences. He eventually did, and improved almost 100 pts. I don't think the engine was that much better than 20 ply. And now a 20 ply engine combined with a neural net computed off line, called NNUE IIRC, is another improvement. It ran on my phone. Couldn't tell you the rating. *** My general point is that chess players, especially those who view the game as science or art, tend to seek the truth. I like the Dojo because it raises interesting questions. I didn't mean to come off as doubting your advice, knowledge or ability. These are difficult questions, and at this point in my life I'm more interested in the theoretical aspect than the practical aspect. The three of you do a great job blending the two. Thanks.

[@aiGeis]

@@mikem668 Dude, why are you even talking about alphazero when it's dubious that its even much stronger than stockfish 8, or at all stronger than 9 (when the match took place 9 was already out) It's the most overhyped chess engine of all time.

[@mikem668]

@@aiGeis "AlphaZero’s strength is impressive, but its method is far more important. AlphaZero isn’t just applying human knowledge and plowing through billions of positions to generate moves - it’s creating its own knowledge first. And, based on its results and my observations, the knowledge it generates for itself is unique and superior. We aren’t just getting faster results the way we do from a calculator. Instead of a postcard from a far-off land, it’s a telescope that has the potential to let us see for ourselves." Garry Kasparov Forward to Game Changer * But of course you know more than Kasparov, dude.

[@mikem668]

@@BeFourCM According to the USCF in 2004, my rating is around the 80th percentile. Not very good, I admit. But better than 75% of tournament chess players in the US. So far from average. I suspect that I've been playing chess longer than you've been alive. You might want to think twice before accusing someone you don't know of cheating. It's bad manners.