Stockfish 19 | Hacker News Reader

Stockfish 19

stockfishchess.org

279 pointsatiedebee169 comments

Interesting bug or regression.

Let’s take the very famous game Kasparov - Topalov, Wijk aan Zee 1999 and look at the position just after Topalov accepts Kasparov’s Rook Sacrifice:

https://samboy.github.io/blog/Kasparov.html#RookSacAccepted - https://lichess.org/lwiPq9wB#48 - White to move

FEN: b2r3r/k4p1p/p2q1np1/Np1P4/3p1Q2/P4PPB/1PP4P/1K2R3 w - - 0 25

Run the above position in Stockfish 18 with a depth of 25: [1] It shows that White is 3.53 pawns ahead—i.e. White is clearly winning.

Run the above position in Stockfish 19, however, and Stockfish 19 at depth 25 says that, while White is ahead, it’s only a 1.2 pawn edge—Black supposedly still has good drawing chances.

At a depth of 35 ply, Stockfish 19 sees that it’s a clear White victory (6.79 pawns ahead), but it takes far deeper search for Stockfish 19 to see the win compared to Stockfish 18.

[1] To run the position, I type the following commands:

  position fen b2r3r/k4p1p/p2q1np1/Np1P4/3p1Q2/P4PPB/1PP4P/1K2R3 w - - 0 25
  d
  go depth 25

I also see the issue on Lichess’s GUI by going to this URL:

https://lichess.org/lwiPq9wB#48

Evaluation doesn't matter as far as I understand. If 19 beats 18 in a match on billion playouts then it's a feature.

Any usability of stockfish evaluation for human analysis is collateral.

Unfortunately there seem to be no serious fork of stockfish that would care about usability for human analysis.

I was recently doing some larger study and I stumbled upon a bug/feature of stockfish where it accepts small number of invalid cache reads for a lot of speed.

It would result in infrequent blunders which however affected my analysis, so I removed this optimization in my fork.

Usability for human analysis will likely be more successful going a similar direction Maia is going with the lc0 like net trained to make human moves. SF is inherently so far removed from the human approach to chess (relatively speaking), that it would likely require a lot more work than just tuning it to work well on some positions.

There would be value in prep if one had an engine that automatically understood what positions are challenging for humans, as right now a big part of high level prep is finding opening lines that look like they lose by a bit in stockfish vs stockfish, but are full of landmines for a top GM to step on. That's where a lot of time goes nowadays, past the pure memorization. The issue is that value drops a lot when other players have access to the same tools, because then they will also pay attention to those lines. The value is in the information imbalance, but everyone has the same tools.

It might help young players learning without a coach, but my understanding is that most are just looking at engine evaluation anyway, as tactical speed is a typical advantage of the youth.

Is your fork public? Have an example of where Stockfish does not blunder with your change?

It's not public, but the diff is tiny: https://pastebin.com/Nd2sS0KX

Also - it's not only my change is making stockfish not blunder. Any little change in the code or execution will butterly-effect into no blunder.

To reproduce a bunder you have to be very deterministic. You need to specify threads = 1, and fixed cache sizes.

I have no logs and exact settings used anymore unfortunately.

One example move was 48. g4 from this game: https://lichess.org/osSoJGzO#95

> I also see the issue on Lichess’s GUI by going to this URL:

> https://lichess.org/lwiPq9wB#48

When I go to that link it shows Black is winning with -1.9 eval at depth 32 with SF19 1MB NNUE! It looks like it got this evaluation from the cloud database.

Click on the switch to the left of the evaluation to run a local Stockfish 19 eval of the position.

Showing Black winning in that position is clearly a bug; at 20 ply Stockfish 19 sees a +0.3 edge for White (which is also wrong; it’s about +5 for White, i.e. clearly winning) and only a +0.6 edge at 24 ply.

To be fair, Topalov, right after this game (1999, so before modern computer analysis), analyzed the game for hours with his second and thought Black still had strong drawing chances after accepting the rook sacrifice.

However, Stockfish 19 has gotten better at understanding fortresses:

https://lichess.org/analysis/1b3r2/8/4P3/2pP2k1/3nB2p/3P3P/5... (now finds 1...Rxf2 = immediately, the evaluation is closer to 0.0)

https://lichess.org/analysis/8/kp2q3/1Np5/pPPp1b2/P2Pp1p1/4P... (finds 1.Nc8+ = quicker (let it ponder four lines), the evaluation is closer to 0.0)

https://lichess.org/analysis/rk6/1p1b4/pBp5/PPPp4/3Pp3/4Pp1p... (The evaluation is closer to 0.0, though still far away)

https://lichess.org/analysis/7k/4p3/5p2/1p1p1p2/1PpP1PpB/1pP... (Evaluation is closer to 0.0)

Here is what I see typing this is Stockfish 18 and Stockfish 19:

  position fen 1b3r2/8/4P3/2pP2k1/3nB2p/3P3P/5PP1/5KRR b - - 0 1
  d
  go depth 35

In Stockfish 18: Black is 3.9 pawns ahead, and it wants to play Bg3

In Stockfish 19: Black is 1.2 pawns ahead, and it wants to play Rxf2!

So, yes, here Stockfish 18 can’t find Rxf2!, but Stockfish 19 does starting at depth 13.

To show you how strong Stockfish 19 is compared to 18, I used to lose 100% against 18, and I now lose 100% against 19, probably faster. Time to fire up En Croissant and see :)

Probably not faster if I have to guess. The nuances of why 19 is better than 18 are likely indistinguishable from noise when they play against someone at a much lower level (i.e. any human).

also in chess a better line is not required to be shorter

This seems like an interesting alternative line of research. Design the a chess engine where the goal is to beat the best human in the minimum number of turns, while still offering ~no chance of human victory.

The Leela odds bots are working on a similar problem: beat humans starting down material. It's roughly grand master level starting down a rook, and for fast time control even starting down a queen

Unfortunately Chess (unlike Go) doesn't have a good handicap system. When playing without a piece (or two) it is not chess anymore OR it doesn't really matter.. During how many games does Queen's rook come into play during early middle game?

Go, on the other hand has a nice handicap system (which doesn't change the nature of the game too much). One can have a fighting chance against a much stronger player with enough handicap (Upto 9 stones). In Chess terms, an IM can have a _decent_ game against a GM. Probably lose. But the GM has to work for it.

the rook is fairly useless early game (which is why rook and knight odds are surprisingly close), but the queen actually does a lot in the early midgame (just mostly not by moving). The queen is usually keeping a bunch of squares indirectly defended, the trick is that the queen starts in a pretty good position and is much more mobile, so its importance doesn't come from where it is, but where it could be in 2 movies

Losing faster to Stockfish 19 it'd mean youre an insanely good player, hundreds of times better than even Magnus.

For me (having written my own tiny 400 line chess engine, that now I'm not able to beat - sure I'm not a good player, but still), the reason to have Stockfish is not really to play against it, and then inevitably lose, but to use it to show which moves are best. I assume that's how most (hobby and professional) chess players use Stockfish nowadays.

I can't beat the famous 1k JavaScript version either. I was better when I was younger.

https://nanochess.org/archive/toledo_javascript_chess_3.html

this feels around 1000 elo

crazy to see 110mb.com links in 2026 (on the github repo project), it's been a while

Yeah, I blundered a knight and still managed to win.

What does it matter if you can see which moves are best if you can't understand it?

99.5% of the time you can understand it (percentage varies with rating).

You just try it yourself and go down the different lines. "Okay, Stockfish suggests this, but I wanted to do that, let's try my idea, oh, I didn't think of that response to my move". Etc., play along some lines, think about them. Sometimes it's obvious, other times it just sets up something far in the future you must be a grandmaster to see the subtlety of.

The Square One Chess channel on youtube turns this kind of exploration/commentary into quite digestible content, though the thumbnails are, like everything else these days, a little clickbaity.

It's been fun viewing for me and my chess-interested 9-year-old.

ChessNetwork on YouTube also does a great job with this: analysis of interesting games, using Stockfish as a tool along the way.

For some reason this question reminds me of all of the drama about Navier-Stokes from the last few days. Tangential to the ethical questions are tons of examples where in history when word gets out about a solution to a problem, not even the solution itself just rumors that there is a solution, suddenly competing solutions appear.

So maybe it matters like that? Just knowing that an oracle like stockfish is saying "this is the best move" may trigger pathways in your brain that promote understanding?

Because then you explore the line and try to learn what is so good about the move

Stockfish has created entirely new lines in chess which people incorporate into their play. For instance an early pawn push to h4 is now common among chess pros. It's value was not recognized before Stockfish.

h4 push was AlphaZero/Leela thing wasn't it?

It’s oddly fun to put maximum Stockfish in play against some horrible position and to be murdered by it.

Feels like a variant of "really bad chess". Would make for a fun game.

TIL: En Croissant https://encroissant.org/

love the name :P

Magnus cannot even beat his phone. We're long past the point of humans being able to compete with the `fish, any version, any platform.

I am a fairly good player but after playing 22000 games, I came to the conclusion that ideas and strategies from chess are context-dependent and don't much transfer to real life.

I'd rather spend all that time reading up on game theory and play some fun video games, like Red Dead Redemption.

I think there are some valuable lessons that chess can teach, if you approach it the right way:

- it teaches resiliency in bad situations.

- it teaches you to not be too harsh on yourself as a person. If you watch GMs playing online, they make bad mistakes all the time too - they're not godlike creatures who make no mistakes, far from it.

The flipside is that if you approach it the wrong way (I know I have at many points), it can have the opposite effects too.

To approach chess in a mentally healthy way was a long process for me. You need to get rid of your ego and be objective as much as possible when playing, or it will just make you sour - this in itself is a valuable lesson.

McFarter? LOL. That made me laugh.

Yes, I agree.

(Disclaimer, I’m the founder of chess67.com)

There is a very tangible benefit to people playing in chess that has nothing to do with the game itself. One example is for kids - I think kids playing chess (particularly over the board) learn quite a lot from the game. For example, etiquette (being respectful to your opponent), patience, logic, etc… but one of the best things they learn is how to lose. In chess you lose a lot. Asymptotically vast majority of players will lose approximately 50% of their games. Learning how to lose (respectfully), learn from your mistakes, then overcome them is really great process to learn. Doing this in real life rather than over the screen is also completely different.

Combine that with the game being fun and the social aspects (meeting new people, making friends)… chess very much transfers to real life in my view

I'm sorry, the founder of chess what dot com?

LMAO

I am too simple, could you pls share why that's funny...

> For example, etiquette (being respectful to your opponent), patience, logic, etc… but one of the best things they learn is how to lose.

That's valid for pretty much every kids sports/games though

So a lot of games/sports kids could cry out - "that is not fair". There is an element of luck in most board games.

Chess is the first board game you can introduce to your kid that has no element of luck whatsoever. It could be brutal. I have seen 7-9 year olds coming out crying after a game in a tournament. But it also may be requires very good parenting to handle all the emotional distress. Chess also is obviously accessible.

It is an emotional rollercoaster to engage kids in Chess tbh. We are ready to sometimes throw the towel every other week. :-)

i used to play poker casually, and the part that I found most interesting and most useful was not the reading other peoples' body language / tells or the probability calculations, but rather the practicing of regulating my own emotions

poker is different game-- more in tune with applied game theory but also, as you said, self-regulation and psychological introspection.

> (Disclaimer, I’m the founder of chess67.com)

You're so evil lol.

Created Feb or March this year it looks like lol.

LOL. So here is the truth: I named it because my oldest son was really into the “67” meme. I was trying to come up with a domain name, everything decent seemed taken, and chess67.com was available.

I figured it would just be a temporary name until I came up with something better.

Then it stuck. I will say it has been easy for my customers to remember the name!

Main thing I took away from chess when I used to play is it's just another exercise to work parts of your brain. I preferred 3+0 and 5+0 chess because of the low time commit, ease to move next, and the tradeoff between most correct move and most optimal move to cause the opponent to expand more of their time trying to figure out what you were doing.

I hit about 2k Elo on lichess playing the 3+0/5+0 and I have no regrets!

Yeah I have ADHD and can't bother play slow, long games. I only play bullet (60 sec) and would often defeat 2200-level players before I decided to give up. That means, most of those I will play (socially) in the future won't beat me.

That's true. Chess only teaches two general lessons about strategy: look more than one step ahead, and invent heuristics (abstractions that let you estimate if positions are good or bad) for yourself. Those are useful lessons, but you don't need to play thousands of games to grasp those ideas.

Other aspects of strategy, like shortening decision loops[1], forming alliances, or the exploration/exploitation tradeoff[3], are simply not represented by chess because it is a turn-based, zero-sum game with perfect information. That makes chess a poor model of real-world strategy.

Still, the things it does teach are real and useful, so as long as you don't think its the end-all, be-all of strategic thinking you can get something out of it.

[1]: https://en.wikipedia.org/wiki/OODA_loop

[2]: https://en.wikipedia.org/wiki/Cooperative_game_theory

[3]: https://en.wikipedia.org/wiki/Exploration%E2%80%93exploitati...

I think chess teaches a lot more. It teaches decision making and resource management. It's very important to consider your options and decide how much time you spend analyzing them. It teaches you things like elimination process (if you have 3 options, one complicated one and 2 simple ones then start analyzing simple ones because if they are bad you just saved a lot of time by not analyzing 3rd one).

It also teaches you to forget about the past and focus on the present. It teaches you long term plans usually don't work and relying on them is a bad idea and it's better to focus on execution here and now (this is what Soviet school got wrong although probably only officially, the players knew it all along - analogies to 5 years plans were just too tempting for book authors).

At human level it also teaches you to design strategy to exploit your opponents - play openings they hate, simplify vs players who like complicated positions, complicate vs players who like simple ones.

There is a lot of interesting things you learn about thought process and decision making as your level improve. It might not be the best game for that but it's pretty good considered how simple it is.

You (and others) should check out real time strategy games if your hope is to learn much of what you're sharing here!

RTS requires other computer-type skills that are pretty useless. APM (micro and macro management), changing screen focus quickly, etc.

Chess is a great game for removing all those external factors. It's just thinking

I'd argue that chess similarly has "useless skills". But learning "useless skills" is one of the most important things you can do for yourself to learn "useful skills".

Not everything has to be an absolute competition-or-not. For the same reason you can enjoy a casual game of chess with a friend, you can enjoy a casual game of some RTS. Helps that there are a lot of open source RTS communities that each have their own skills-emphasis.

> Chess only teaches two general lessons about strategy: look more than one step ahead, and invent heuristics

That is true for an engine.

Meta-games around studying your opponent, specific match preparation and all those game theoretic stuff are so pervasive people claim they are 'killing' chess at top levels. Anyway, for humans, chess at high levels of mastery is a striking display of bootstrapping some insane cognitive machinery around memory, processing speed, 'expanding' working memory via understanding:

[0] Alekhine blindfold simultaneous vs 32 opponents: https://www.blindfoldchess.net/assets/images/alekhine-excerp...

[1] Reingold, Charness, Pomplun & Stampe (2001), "Visual Span in Expert Chess Players": https://pubmed.ncbi.nlm.nih.gov/11294228/

[2] Connors, Burns & Campitelli (2011), "Expertise in Complex Decision Making: The Role of Search in Chess 70 Years After de Groot": https://onlinelibrary.wiley.com/doi/epdf/10.1111/j.1551-6709...

"Only teaches two useful lessons" seems a little certain.

Off the top of my head I can think of a number - the value of diversity and the contextual power of capabilities such as the power of weaker pieces used diligently, and the power of counterintuitive sacrifices.

Chess also makes no accommodation for your fantasies and ego. There is no luck, only skill.

There is also much to be garnered from the layered complexity of strategy, where some strategies will take you to a certain level of play that need to be completely replaced or expanded with nuance to reach a higher level.

There are many lessons to be learned from chess.

That said, I would not recommend anyone become good at chess to get good at life.

You can play Diplomacy for that.

Reminds me of: "The ability to play chess is the sign of a gentleman. The ability to play chess well is the sign of a wasted life." Paul Murphy.

I got to a 1900 bullet rating in two years, playing an average of 10 to 30 games a day--while not learning any theory! (I wanted to discover theories on my own).

And that's when I realized nah, it's not worth it.

That's a great quote, thanks for sharing.

Murphy was in general so dismissive of his own talent. Granted, he lived in the era before people played chess for a living.

*Morphy

> play some fun video games, like Red Dead Redemption

To each their own.

If you like the competitive aspect of chess, you may enjoy physical sports too. I started to find video games pretty boring compared to pushing myself physically.

I do like physical competitiveness as well-- I did box in my teens, and basketball for decades. It depends on the video games, too.

I agree. I think in general, any game that is 'turn-based' (in this case chess), with a limited set of options, quickly boils down to pure theory and in my opinion 'dumb intelligence' of just remembering many things and patterns (the AI of now). As soon as a game is in 'real-time' with many-many options, it reaches a level of complexity where you cannot ONLY remember theory and outsmart your opponent. It is far more about applying logic on the fly with the limited time and overwhelming set of options you are supplied. This skillset applies to almost everything in life and serves me personally very well thus far in life. Yes, I understand that chess has different modes that severely limit time per turns which transform it more into the 'real-time' games and value, but it is still turn-based and still has a limited set of options when compared to true 'real-time' games.

I've never studied any openings and still have loads of fun with chess. I think you have to be IM level or something to not immediately go out theory.

It can be a bit fun, sometime, to play someone that has studied an opening at my level, but since they aren't that deep into it, they only know the moves for some variations. So me, not knowing the opening, do something from the left-field, and they don't really know how to punish it anyway. Remember one time someone got a bit grumpy, heh.

Maths is "turn based" albeit it has practically infinite available moves per turn. But I think the latter matters less than it may appear at first: At least in practice there isn't much different (for both humans and computers) if there are infinitely many moves, or "just" 10,000. And while (as far as I know) there aren't any board games with that many moves, the exponential explosion means that if you conceptually "combine" let's say three moves in a row, you suddenly have not 30 moves, but 27,000. And that's practically infinite.

You can play a variant of math where you only have a limited set of options available each turn, and it wouldn't make much of a difference.

(Eg limit yourself to commands you can send to vim as you write your Lean proof or whatever.)

> I think in general, any game that is 'turn-based' (in this case chess), with a limited set of options, quickly boils down to pure theory and in my opinion 'dumb intelligence' of just remembering many things and patterns (the AI of now).

That's perhaps true for full information, two player games.

Try Diplomacy for a change. (Even if you play without explicit negotiations and only have the moves speak for themselves.)

I play chess and I'm basically with you

It's just a board game

Before transferring to life, it doesn't even transfer to a game like Poker.

Real life is a game of risk/reward rather than a game of deep calculation. You can't "calculate" life because there are too many variables and unknowns. You have to simply pick a path and either accept the risks or mitigate them.

Depends on what you do in life.

Zugzwang is everywhere.

Eh.. it did not originate in chess, this Zugzwang. It's a fairly well-known mathematical entity, goes by the name of Optionality. If you need to fly tonight, and the airliner knows it, they can charge you very high prices: you have no optionality.

Nassim Taleb say a lot on this in his Anti-fragile.

I don't play games for them to "transfer to real life". I play games for fun. Chess can be fun, it has an interesting history and the strategy of various kinds of play open up to you as you play/learn more, it has a structured competition community so you can go places and play with other real people and win fabulous prizes (not really; even some of the best professional players also have real jobs). It's played worldwide, so no matter where you are, you can find people to play with in the real world. It's popular, so there's always someone online you can play against.

Some folks find chess fun and interesting enough to play it and study it all their lives. Other games can also be fun, though it's rare for a game to hold someone's attention throughout their whole life.

I live my life such that the boundary between fun, work and thinking is usually fuzzy and blurred.

So, while I did play chess for fun, after a while it became boring--but I also impute this to my ADHD, perhaps for others it might not stale.

I would not mind a game or two with an interesting opponent, but playing daily is out of the question for me anymore. Time has opportunity cost, right?

Compared to chess, Magic: The Gathering ability transfers much better to real life. MtG has:

- dealing with imperfect information

- bluffing

- thinking ahead with an enormous number of possible moves (especially when you consider deck construction)

- tracking board state

- mentally implementing "triggers"

Why would you play 22000 games if you didn't enjoy it? Are you saying you went through all of that because you thought it would be good for you?

My comment may have been too terse. I didn't mean I didn't enjoy or else I wouldn't have played. I did, but it became stale after so many games and regularly defeating 2000+ opponents in bullet chess. The time and energy needed to reach 2500 is not worth the goal.

The GUI page is a bit out of date imo, or I'm not sure how does the selection work because there are countles GUI options nowadays. Anyways some of the best options are not even mentioned like

Nocheto https://nocheto.sallyx.org/ (this my favorite)

ChessMD https://chessmd.org/

PyChess https://pychess.github.io/about/

I have tried En Croissant and its personally its a nice GUI and easy to use. It allows you to download stockfish and others right from their GUI.

Amused me the first time I ran En Croissant, it lets you download existing chess databases and one of those (Caissabase) was my project from years ago :).

I don't play Chess much any more but others have taken up the Chess DB mantle (with slightly different criteria).

I do find it ironic that a project I did in a couple of days for myself and chucked online in case it was useful is still kicking around though.

If you like En Croissant there is a nice fork called Pawn Appetit that does all the same things and adds some other QoL bits and bobs.

I really like ChessX ( https://chessx.sourceforge.io/ ). Pychess was always unstable and slow for me.

Anyways. Now I can use an even better Stockfish to benchmark my own engine..

If you like uncomplicated GUIs: http://mayura.com/chess/

Shameless self promotion:

If you’re a Mac user and want something with native look and feel, there is also The Marvelous Brass Chessplaying Automaton (TestFlight beta):

https://testflight.apple.com/join/mqfhpmHY

There are also iOS and watchOS versions, but those can’t run 3rd party engines like Stockfish.

For when you need another decimal place of accuracy deciding what you should have moved instead of hanging your queen.

Would be nice if they included a few "notable" games in the release notes.

Presumably it's easy to highlight ones from benchmarks where the evaluation of Stockfish 19 is most significantly different from version 18?

They could but it's pretty much meaningless. You're bound to find outliers in favor of either engine, and the landscape is pretty chaotic

People perhaps won't remember the long history of computer chess at least how it played out in the 90s

There was a succession of interesting engines (I think The King used to win a number of computer tournaments, this is the one Chessmaster used, that wonderful audio-visual program, like Encarta but chess related) and then the big three of Fritz (match against Kramnik), Junior (match against Kasparov) and Shredder. And also HIARCS and later, Rybka (and Crafty, as well, touted as the open source chess engine)

Stockfish has been top dog (fish?) for a while now but there was a notable exception when AlphaZero (which later became Leela? I'm not sure) beat it, but these days of course Stockfish is far and away the best engine (I guess. I don't know how far behind Leela is)

And, yes, human vs computer chess has stopped being interesting for a while. There were some centaur chess (human and computer; humans using computers to analyze) competitions in the 2000s and some interesting things but nowadays of course your phone can handedly beat Magnus Carlsen and others without problems, reliably

> AlphaZero (which later became Leela? I'm not sure)

Pretty much. DeepMind never released their code or models, just the main ideas of their techniques and training.

Leela, which came out of the go-ai world, implemented the paper(s) and became the de-facto version everyone used for a while until other projects improved over it on both go (mainly katago) and chess (mainly stockfish).

The next closest is probably Torch currently

from the TCEC Season 29 Premier Division tournament, Stockfish was first, Reckless v0.9.0 was second, LCZero 0.33 was third, PlentyChess (won tiebreaker) tied with Torch for fourth.

see https://tcec-chess.com/#div=p&game=1&season=29

click on the Standings tab to see the results.

Alpha Zero Chess never made their code or models public. Leela is based on the same ideas as Alpha Zero Chess, but its model and code are public and uses crowd sourcing to generate and refine the model.

Stockfish uses a combination of classic Alpha-Beta pruning along with a type of AI called NNUE to evaluate positions.

Whenever I see another Stockfish update I dream of the day that I sit down and code my own chess bot that is good enough to beat me. I would be so happy if I manage to do that.

Wouldn't be that hard in my case

Same. Probably a few if-then statements would be enough to beat me :'(

I could hard code it to play the fried-liver and I'd still probably trip myself up occasionally.

Beaten by rand() chess

Depending on your feelings on being spoiled on the general techniques, and an approach to implement them, Sebastian Lague made a great two-part series on building a chess engine. It's specific enough that you have an idea of the techniques involved, without dwelling enough with the details to make implementing your own futile.

https://youtu.be/U4ogK0MIzqk

I did so and roughly twenty years later, my 4th attempt is "finished" in a sense and playing on lichess bots from time to time.

The process is very rewarding and there are great resources (chessprogramming wiki). To get an initial bot play random moves is not that much work but it is a good starting point because you can hook it into a gui and watch.

From there you do the move generator and the search and then optimization possibilities are endless :)

To me, the most interesting part is the architecture of the engine. How you encode a position or a move, how you design the move generator, etc.

Something as simple as encoding the 64-square board in a 64-bit number leads to a rabbit hole that can last years.

After that point, the "AI" part is already solved and you just need to implement it (alpha-beta + all the heuristics).

Most of the recent chess engines simply implement a neural network (by copying Stockfish) and suddenly are at 3000 elo with no effort! Such a waste.

Naive monte carlo on a phone from 10 years ago is likely enough unless you're quite good at chess

Join the Stockfish discord server! There is a very active community who builds their own chess engines there.

(If you ever end up going for it, don't use AI it takes away a lot of the fun)

It's a fun project, start here: https://chessprogramming.org

Correct me if I'm wrong but I think that a few years old version of Stockfish running on a modern MacBook Air with 1 min per move is unbeatable if the game starts from the initial position.

Even if the opponent is the latest Stockfish with 1000x more time per move, it will always be a draw.

I asked about this a few years ago in a Stockfish Discord and the above claim is what someone implied, if I remember correctly.

A faster engine is still very useful in practice, for example for position evaluations in game review.

Not that an engine evaluation will always be helpful (the line in question making a position strong or weak might be effectively unplayable/impossible to find by a human), but it’s useful for studying nevertheless.

This sounds plausible. But otherwise, there are "tournaments" like the TCEC which is for bots only and the way it works is, the bots are adjusted to play a specific opening and they're on their own after the Nth move.

Stockfish has been dominant for a long time. And a later version of Stockfish would outperform an earlier version given the same time control.

I believe this is not the case. The first reason is the optimization of the Transposition Table in stockfish. https://github.com/official-stockfish/Stockfish/blob/master/...

Basically this optimization means that stockfish will sometimes take a bad read from a cache (so it will take an evaluation of the position from cache which was calculated for a different position). It's fairly random when this will trigger a bad play, but if you brute force vs any deterministic setup you will find it eventually.

But even withtout it - you can just implement a kind of a brute-force heap of positions using stockfish's own eval and you can find wins from almost any position even on fairly high playout count.

I got as far as 32M from different even starting positions, but I spent compute to find as many games as possible rather than as deep as possible.

Example game: https://lichess.org/XVUjezMr#134

the collection of the games I found (vibecoded gui) https://kuboble.com/chess_ideas/

The post literally says

> against Stockfish 18, this new release brings an Elo gain of up to 44 points, and wins more than three times as many game pairs as it loses.

I'm talking about games that start from the initial position.

That point about Elo is irrelevant. Games on Fishtest use bullet-like time controls (60+0.6). Significantly shorter than 1 min per move.

As originally suggested, when running modern Stockfish at 60 seconds per move the games overwhelmingly tend to draw. Even when the older engines make slightly worse moves.

For those curious to see actual games played by top-level engines, the most famous French chess-commentator has a youtube plalist where he and french GM Matthieu Cornette analyzes some of these games.

https://youtube.com/playlist?list=PLiZ56cRXeIu3-zr-S7_XND3M3...

GM Mathieu Cornette is an excellent chess analyist, might not be the best at banalazing, but he does find some amazing computer games. Engine games are often boring, so it's nice that they pick the good ones.

Nice.

I wish these engines could explain why it is recommending certain move, or certain sequence of moves, using natural language - it would be very useful for learning.

> This release introduces the SFNNv16 network architecture, reducing binary size by removing redundant threat features…

This is nice to see. One of my few gripes with NNUE engines is their huge size compared with the old “HCE” engines I was accustomed to (which were usually hundreds of KiB or low single digit MiBs in size). I will trade off size for superior playing strength but at a certain point it becomes hard to keep justifying a 2x increase in size for what seems like ever diminishing practical returns. I am still searching for a strong NNUE engine that is ~10MiB or less. I have found Marvin, which is small at 1.8 MiB, but it is nowhere near as strong as Stockfish.

Question: does Stockfish beat more generic systems like AlphaZero?

I think so, see the tournament results: https://en.wikipedia.org/wiki/Top_Chess_Engine_Championship#...

yes and it has for a very long time. Stockfish took the neural net approach after AlphaZero showed it was a good idea as all modern chess engines have

They basically replaced the static position evaluation function with a NN approach and kept the rest of the tree search mostly intact.

I ran fishtest on an old cluster for a while. Was nice to watch the stockfish team methodically improving things. Basically, every commit has to show it can beat the main branch before it gets merged.

My understanding is that AlphaZero only really existed for a year or two; there's no objective way to compare it at the moment.

Leela Zero tried to open-source that work, but Stockfish incorporated a number of improvements from AlphaZero, including a neural network and a different search method, and consistently beats Leela Zero.

I have a book, "Game Changer", in which a chess expert calls out several instances where AlphaZero made moves surprising at the time; situations where all chess engines rated things one way and AlphaZero rated them differently. When I enter them into Stockfish now, it usually rates things more similarly to the way AlphaZero did, and often chooses the move chosen by AlphaZero.

The only real test of course would be to dig up AlphaZero and run it again; but I think based on the evidence we have, Stockfish of 2026 would probably trounce AlphaZero of 2018 with equivalent compute available.

I guess someone can twit Hassabis and ask him :D He certainly has access.

who is Hassabis?

Demis Hassabis [1], co-founder of DeepMind

[1] https://en.wikipedia.org/wiki/Demis_Hassabis

Isn't one problem that it's hard to determine what equivalent compute is, for CPU search vs a neural net based engine like AZ or Leela?

One way would be to calculate a cost per game, factoring in both electricity and an amortized cost of the hardware, maybe having a penalty too for extra time run (e.g., if focusing only on hardware depreciation and electricity, 1 minute of TPU would translate to 2 weeks of CPU, that 2 weeks of waiting still costs you something). Obviously this isn't stable, as relative prices of GPUs and memory shift over time, and it's somewhat sensitive to setup; but done right it's probably more "what a user actually wants to know", in terms of what it would take to get equivalent performance.

It's more fundamental than that. AlphaZero is a shallower search with a heavier evaluation function. Stockfish is a deeper search with a lighter evaluation function.

In chess, depth usually wins because of how narrow the search tree is compared e.g. to Go.

Interesting, and if you don't mind, where do we put humans (and superhumans like Magnus Carlsen)? I think they have a heavy evaluation function and do shallower search.

Way way shallower search. A human doesn't consider more than a few candidate moves. But they (we?) build up much better intuitions and heuristics

The Stockfish NNUE is completely unrelated to AlphaZero.

It's a neural network rather than a bunch of hard-coded rules. That turns out to make a big difference.

Actually, there's this interesting snippet from the release page:

> These techniques have been applied to hundreds of billions of training positions, all of which have been consistently rescored using a strong Leela net.

So Stockfish's neural network evaluator is actually trained using Leela Zero.

"Completely unrelated" is not quite true. Stockfish current NNUE models are trained on LC0 training data. LC0 is pretty much an open-source community replication of the ideas from AlphaZero.

I stand corrected. However, fundamentally, the idea of "tiny CPU-only neural network" combined with traditional alpha-beta search is substantially different from "big GPU network" combined with Monte Carlo Tree Search. And historically the NNUE came from a 2018 idea for shogi engines rather than from AlphaZero.

Leela Chess Zero did surpass Stockfish for a while, until Stockfish switched its eval to NNUE. Modern Stockfish would annihilate AlphaZero.

> My understanding is that AlphaZero only really existed for a year or two;

It still exists, but it's private / internal, and sometimes used for a few different things.

It was used by Kramnik to test the hypothesis whether no-castling chess was viable (basically chess, but disallowing castling). That was a year after DeepMind published the match they ran of AlphaZero versus Stockfish.

It was still in use last year, I remember seeing some Grandmasters with interests in chess studies were invited by DeepMind to judge the beauty of chess problems composed by AlphaZero (or whatever form the thing that used to be AlphaZero is now).

So it's still kicking around in the background.

> It was used by Kramnik to test the hypothesis whether no-castling chess was viable (basically chess, but disallowing castling).

Viable in what way? That it's advantageous to never castle if an engine learns to play with that directive? Or that it still makes for a fun game with that new rule?

The hope was that it would be a slight rule tweak that would lead to fewer draws without favoring white too strongly.

Technically nobody knows, because AlphaZero was never made public. In practice, we know Stockfish 19 would destroy it, because it destroys open-source reimplementations of AlphaZero.

> like AlphaZero?

Unknown, it was only around for a little while however there is Lila which is based on the same ideas as AlphaZero (and also a fantastic project generally) they generally trade blows with Stockfish coming out very slightly ahead.

That said Stockfish also uses an NN (NNUE) these days so the line is getting more blurry.

Like a red headed stepchild.

On average, yes. Stockfish is the strongest engine and beats AZ-like implementations like Lc0 and the like. But on a game to game basis Lc0 can still win some games, depending on the starting position. It's rare that Lc0 can win both black and white starting from the same position, tho.

There are a few yt content creators that cover great engine games, if you're curious.

It's widely believed so because Stockfish finds all of the right (or better moves) in the games from the original paper.

AlphaZero does not exist, as far as modern computer chess is concerned

Leela proved that the ideas worked, regardless of the conditions that the AlphaZero vs Stockfish matches were run in.

What's funny is that if computer chess had paid more attention to computer shogi, Stockfish could have already been using NNUE by that time, and not lost to AlphaZero. But perhaps that would have stifled Leela's development, and would have led to worse training data for Stockfish in that parallel universe's present day, making it weaker overall.

My favorite chess quote:

"The ability to play chess is the sign of a gentleman. The ability to play chess well is the sign of a wasted life."

I find this to be very valid.

Is doing something you love, and practicing to the point of expertise, wasting your life? Replace chess with playing an instrument, or basketball. This quote is just exclusionary.

Game engines, AI, robots and machines in general are establising a "goal horizon" for humans. This is unnecessary and counter productive way of using tools. Tools are supposed to work for us, not set goals or direction for us.

But as technology progresses and becomes more advanced, it does seem to set goals in all directions in that seems inevitable if we continue to 'collaborate' with modern technology. I mean, we no longer develop tools to help us out really, if you think about it. We develop tools to get ahead of other tools, to be faster than them, to protect ourselves from still other tools.

I'd argue that as we condition ourselves into a state of constant development, we shouldn't even call these things "tools" anymore, but pieces of an organism that spurn us to upgrade and complete it at every step.

Tools working for us only applies to local, simple tools that one person or a small communiity can basically understand and make, like a hammer. That's a tool. I wouldn't call a computer a tool any more just like I wouldn't say that our brain is merely a tool. It's more: it's an extension of us and even an integral part of us that modifies our very instincts and desires itself.

We shouldn't condition ourselves to keep transforming these tools unto our masters. That's the point I was making. We don't the "goal horizon" that is reachable to machines but too far away for humans. There is no point in having that. Goals should be reachable to humans. Humans never needed deal with numbers larger than a few thousand. Centralization is driving the demand for super-human capabilities.

I think an interesting problem would be for a LLM from OpenAI or Anthropic to try to learn chess from first principles and then use that knowledge to try and beat Stockfish.

I wonder if anyone at those companies is trying to tackle this problem.

It's probably quite challenging if you want to avoid chess information to be in the training data.

This may be a dumb question, but how do these engines get better when they're already amazing? Is it simply that they look further down the tree of possibilities?

Presumably a lot of it is in a better evaluation function. That is, when searching the tree, a better analysis of whether a specific position is beneficial and by how much. After all most of the tree searches do not end in checkmate, just in minute improvements which accumulate over many moves.

Yes. The ways to improve are: faster search (which lets you search more deeply) or better evaluation of a position or better heuristics.

How do you get better when you're already amazing?

When you're amazing, you understand what you're doing and where to improve. For very complex systems, the improvements will be in confined areas (not rewriting everything but noticing a bug or reading about another project which found that some other hyperparameter works better). Or even if you have no idea, the simplest way would be making random tweaks and seeing which ones are advantageous

Obligatory Tom7 video/papers: https://tom7.org/chess/

This was all I thought of when I read "Stockfish"

But does it use AI?

As a joke project, it would be nice to add a LLM module that adds a commentary bullshiting an explanation of why it is destroying the human player.

It's been tried, they feed the evaluation into the LLM and get it to commentate.

The results are absolute garbage (as you would expect) since the LLM's really don't understand chess or how to play it, someone posted a benchmark of the current SOTA ones playing Chess and the best was estimated at 1480.

I think OP meant trash talking. "Nice knight on the rim you got there buddy." "How long do you think you'll be able to hold onto that pawn?" "The Latvian gambit, really?"

Both, a nice and a naughty version :)

This is a very rough approximation of what the strong bots on chess.com do.

It's not really what you're looking for, but I made this lichess BOT that makes stupid (but relevant) comments on your moves from time to time: https://lichess.org/@/Annie_Archy/all

yes

This is good also https://chessdox.com

Is this an ad? How's this relevant, does it run the new Stockfish or a competing engine or something?

Reply on news.ycombinator.com