This is very cool and having stalemate is nice, however how much space would it take to implement the full ruleset?
As you write: not implemented: castling, en passant, promotion, repetition, 50-move rule - those are all required to call the game being played modern chess.
I could see an argument for skipping repetition and 50-move rule for tiny engines, but you do need castling, en pessant and promotion for pretty much any serious play.
Cool project. You could also use the front-end of GNU chess to save some lines, and implement only a back-end.
Bug report:
a b c d e f g h
8 r n b q k b n r 8
7 . . p p p p p p 7
6 . p . . . . . . 6
5 p . . . . . . . 5
4 P . . P P . . . 4
3 . . . . . . . . 3
2 . P P . . P P P 2
1 R N B Q K B N R 1
a b c d e f g h
move: b2b3
ai: b6b4
The pawn is not permitted to move two fields after it has already beeen moved once before: b6b4 isn't a valid move after b7b6. (First moving two fields, and then one would have been okay, in contrast.)
Cool! I just recently implemented a chess engine in ~400 (readable) lines, with all rules, first in Java and then ported to my own programming language "Bau" [1]. This is including a terminal UI. I'll measure the ELO, but I was never able to beat it :-) The castling moves are specially tricky to implement I think. I enjoyed the challenge as well.
If anyone is curious, the most common tool I've seen for ELO estimation among engine developers is cutechess [1], which uses SPRT [2]. Or ordo [3], haven't used this myself though
maybe for very low ratings it's plausible?
1 elo per byte might happen in a tiny range
but at a useful strength it would break fast, that's what i think
What's the snallest possible program that accepts a chess board state and prints any legal move? True randomness may only have a couple hundred ELO, but then, that's pretty big for golf
The program that resigns every time unfortunately does a lot worse than random. But it depends on the population it's pitted against - it should at least pick up a few points against copies of itself.
How many games did you have to throw away because stockfish wanted to castle? Or did you force stockfish to not castle? Castling seems like such a frequent move it is hard to draw any conclusions about the strength of an engine that does not support it.
zero games were thrown away for castling, because i forced stockfish not to castle (and not to play en passant/promotion) by filtering legal moves and only giving those filtered moves via root_moves
so every game stayed in the same no castling variant
and you're right, this rating is for that constrained variant, not full chess.
need to start measuring these things in the size of compiled functions so we can stop looking at oneliners (maybe wasm since it has an easy to read text representation)
It's wild to think that 4096 bytes are sufficient to play chess on a level beyond anything humans ever achieved. Makes you think what other difficult tasks are out there that take even highly gifted humans years or decades to master, but a superior algorithm would more or less fit into one of those big QR code formats.
These things always make me think back to Westworld season 2, where the finale revealed that human minds are much simpler than they themselves believe and fit completely into an algorithm that could be printed in an average book.
Well, one of the most fundamental algorithms for building a chess AI is minimax [1] (or variants like negamax), and that’s been around for close to a century. The key difference is that as compute power and available RAM have grown, it’s become possible to search much deeper and evaluate far more plies.
So while 4k is still very impressive for the code base, it comes with a significantly larger runtime footprint.
Min-max + alpha-beta pruning is the backbone of the chess engines. I think 2KiB or even 1KiB might be enough (I guess the last one would be a very challenging squeeze). But what separates the best engines from average ones, is the heuristics. Heuristics is the most complicated one, and I doubt it's possible to fit it into a single-digit kilobyte memory (even 2-digit). For heuristics, engines like Stockfish also use neural networks, in addition to hand crafted algorithms. Also huge tables are used for endgames, etc.
If you look at the current winner [1], it does a lot more than just brute force tree search. The space state for chess is simply too big to cover without good heuristics. Deep Blue may have been a pure brute force approach to beat Kasparov after Deep Thought failed using the same core algorithm, but modern chess engines search far deeper on the tree with far fewer nodes than Deep Blue ever could thanks to better heuristics.
I'm not suggesting that it's only brute force tree search, just that it's not very complicated to develop a theoretically perfect chess engine in direct response to the parent
> It's wild to think that 4096 bytes are sufficient to play chess on a level beyond anything humans ever achieved.
It's not just about the base algorithm. It's also about the memory needed to run it, and the clockspeed. For example, even the hardest problem you can imagine, if it has a verifier algorithm that fits in 4k (which means the solution itself can be much larger than 4k), then you can simply do a basic brute force search over the solution space. That doesn't mean this algorithm is very intelligent; it's only very capable if you have a sufficiently fast computer; although indeed brute force is only feasible for the simplest tasks in practice, so the idea that algorithms (of increasing sizes) enable (greater) intelligence is definitely a part of the story, but not the whole story. You can also think of DNA, which represents a recipe for our bodies and brain, which we then use (essentially as an "algorithm") to learn things, with degrees of freedom (memory) far surpassing what DNA stores.
Now if you had a very good chess program running in very constrained (dynamic/RAM) memory, then that'd be partially more revealing. From a cursory search there's a 1800 ELO engine for the C64, which seems very impressive but very far from the best human players.
I'd be interested to see a curve of ELO x Avaliable RAM for the best chess engines (up to given RAM), and how that compares to other games and activities.
On RAM vs ROM (program size) memory, I think at a high level dynamic memory helps you keep track of search paths in a large tree search, saving you some computation. Program size tends to enable improving the effectiveness of your search heuristic, as well as pre-computing e.g. initial and final game optimal moves (potentially saving arbitrarily much compute). I like thinking about those things because I think the search paradigm is pretty informative of computation (and even intelligence) in general. Almost every problem is basically some kind of heuristic search in some kind of space. And you tend to get better at things by refining your heuristics (usually through some experimental training process or theoretical insight), considering more options, exploring deeper consequences, etc..
I think what really defines humans isn't really our ability to solve problems or play chess well etc. (although that's extremely useful and also enjoyable most of the time), it's really our emotions and inner world. We are not really Thinking Machines in essence, we're Feeling Machines most significantly. The thinking part is a neat instrumental part :) We can delegate thinking to machines but what we cannot extinguish is feeling or the human "soul", because that is the source of all meaning.
The mailbox board representation is a good call for size-constrained engines. Bitboards give faster move generation but the manipulation code (shifts, masks, magic numbers for sliding pieces) eats a lot of bytes. With mailbox you just need offset tables and a sentinel check for board edges. Curious what your evaluation function looks like though. At 2KB you can't fit piece-square tables (that's 384 values minimum for both colors), so are you doing material-only eval or did you squeeze in some positional heuristics?
The gap between your 1200 Elo in 2KB and the TCEC 4K engines at ~3000 Elo is interesting. That extra 2KB buys a lot when it goes to better evaluation and move ordering. Even a simple captures-first sort in alpha-beta pruning costs only a few bytes of code but can roughly double your effective search depth.
This is very cool and having stalemate is nice, however how much space would it take to implement the full ruleset?
As you write: not implemented: castling, en passant, promotion, repetition, 50-move rule - those are all required to call the game being played modern chess.
I could see an argument for skipping repetition and 50-move rule for tiny engines, but you do need castling, en pessant and promotion for pretty much any serious play.
https://en.wikipedia.org/wiki/Video_Chess fit in 4k and supported fuller ruleset in 1980 did it not?
So I would ask what is the smallest fully UCI (https://www.chessprogramming.org/UCI) compliant engine available currently?
This would be a fun goal to beat - make something tiny that supports full ruleset.
PS my first chess computer in early 1980s was this: https://www.ismenio.com/chess_fidelity_cc3.html - it also supported castling, en pessant, not sure about 50 move rule.
ToledoChess [0] has a few implementations of this in different languages. Some highlights:
2KB of JavaScript with castling, en passant, promotion, search and even a GUI
326 bytes of assembly, without the special rules
I don't think the author has a UCI-compliant one, but it should be easier than the GUI. There are forks of the JS one that might do it.
[0] https://nanochess.org/chess6.html
Cool project. You could also use the front-end of GNU chess to save some lines, and implement only a back-end.
Bug report:
The pawn is not permitted to move two fields after it has already beeen moved once before: b6b4 isn't a valid move after b7b6. (First moving two fields, and then one would have been okay, in contrast.)Thanks for pointing it out! I will try to patch it.
Appreciate you taking the time to test it.
Cool! I just recently implemented a chess engine in ~400 (readable) lines, with all rules, first in Java and then ported to my own programming language "Bau" [1]. This is including a terminal UI. I'll measure the ELO, but I was never able to beat it :-) The castling moves are specially tricky to implement I think. I enjoyed the challenge as well.
[1] https://github.com/thomasmueller/bau-lang/blob/main/src/test...
If anyone is curious, the most common tool I've seen for ELO estimation among engine developers is cutechess [1], which uses SPRT [2]. Or ordo [3], haven't used this myself though
[1] https://cutechess.com/
[2] https://www.chessprogramming.org/Sequential_Probability_Rati...
[3] https://github.com/michiguel/Ordo
Do you think it would be possible to achieve 1:1 ELO:bytes? Even smaller, but can be less smart.
That's an awesome code golf challenge
maybe for very low ratings it's plausible? 1 elo per byte might happen in a tiny range but at a useful strength it would break fast, that's what i think
What's the snallest possible program that accepts a chess board state and prints any legal move? True randomness may only have a couple hundred ELO, but then, that's pretty big for golf
The program that resigns every time unfortunately does a lot worse than random. But it depends on the population it's pitted against - it should at least pick up a few points against copies of itself.
How many games did you have to throw away because stockfish wanted to castle? Or did you force stockfish to not castle? Castling seems like such a frequent move it is hard to draw any conclusions about the strength of an engine that does not support it.
zero games were thrown away for castling, because i forced stockfish not to castle (and not to play en passant/promotion) by filtering legal moves and only giving those filtered moves via root_moves
so every game stayed in the same no castling variant
and you're right, this rating is for that constrained variant, not full chess.
Wouldn't stockfish's position evaluation be incorrect in that case? (If it evaluated the position based on a formula that assumed normal rules)
https://www.chessprogramming.org/Toledo is a family a moderately strong tiny chess programs.
need to start measuring these things in the size of compiled functions so we can stop looking at oneliners (maybe wasm since it has an easy to read text representation)
I wonder how big 1300, 1400, ..., 2200 Elo chess engines are.
Cool that you could keep it under 2k but it would nice to have a readable version of the source code.
Do you work with it like this or do you have some sort of script you apply to get it down to a single line, single letter variable names?
What you’re describing is the typical output / function of a minifier
The real fun would be reverse-engineering the minified code (there are loads of tools to do this for chrome extensions)
not lossless
How did you handle games where Stockfish would castle or promote?
i forced stockfish to play only non castling, non en passant, non promotion moves by filtering legal moves and passing only those as root_moves
also removed castling/EP rights from FEN
I'd call that cheating but the size and capability is impressive nonetheless.
If you ever spent much time at a chess club, you've seen why 2kB is a really disturbing number.
I have not. Can you please tell me why?
He's just trying to trick HN readers to join chess clubs.
Not really. You have to see it for yourself.
(Partial answer, 2kB is a very small fraction of what we'd like to think counts as human.)
Humans don't have much capacity for systematic tree search. It's sort of amazing that humans can do as well as they can, given that limitation.
I don't think what you're saying has any connection to chess or chess clubs.
2kB is a very small fraction of what we'd like to think counts as human
This doesn't seem to mean anything. Why would 2KB have any relation to "counting as human". It's the data of about 10 comments.
This is amazing! Thanks for sharing. What would be the elo gain for 4KB engine?
P.S. I assume 1200 elo in chess com scale (not lichess / fide elo) and bullet chess variant?
There is a TCEC category for 4k engines. The top ones are ~3000 Elo.
It's wild to think that 4096 bytes are sufficient to play chess on a level beyond anything humans ever achieved. Makes you think what other difficult tasks are out there that take even highly gifted humans years or decades to master, but a superior algorithm would more or less fit into one of those big QR code formats.
These things always make me think back to Westworld season 2, where the finale revealed that human minds are much simpler than they themselves believe and fit completely into an algorithm that could be printed in an average book.
Well, one of the most fundamental algorithms for building a chess AI is minimax [1] (or variants like negamax), and that’s been around for close to a century. The key difference is that as compute power and available RAM have grown, it’s become possible to search much deeper and evaluate far more plies.
So while 4k is still very impressive for the code base, it comes with a significantly larger runtime footprint.
[1] - https://en.wikipedia.org/wiki/Minimax
Min-max + alpha-beta pruning is the backbone of the chess engines. I think 2KiB or even 1KiB might be enough (I guess the last one would be a very challenging squeeze). But what separates the best engines from average ones, is the heuristics. Heuristics is the most complicated one, and I doubt it's possible to fit it into a single-digit kilobyte memory (even 2-digit). For heuristics, engines like Stockfish also use neural networks, in addition to hand crafted algorithms. Also huge tables are used for endgames, etc.
The core search algorithm is very simple though. 4KB engines may not run that fast if they do exhaustive search, but they’ll be quite accurate.
According to TCEC the time control is 30 mins + 3 sec, that’s a lot of compute!
If you look at the current winner [1], it does a lot more than just brute force tree search. The space state for chess is simply too big to cover without good heuristics. Deep Blue may have been a pure brute force approach to beat Kasparov after Deep Thought failed using the same core algorithm, but modern chess engines search far deeper on the tree with far fewer nodes than Deep Blue ever could thanks to better heuristics.
[1] https://github.com/MinusKelvin/ice4
I'm not suggesting that it's only brute force tree search, just that it's not very complicated to develop a theoretically perfect chess engine in direct response to the parent
> It's wild to think that 4096 bytes are sufficient to play chess on a level beyond anything humans ever achieved.
It's not just about the base algorithm. It's also about the memory needed to run it, and the clockspeed. For example, even the hardest problem you can imagine, if it has a verifier algorithm that fits in 4k (which means the solution itself can be much larger than 4k), then you can simply do a basic brute force search over the solution space. That doesn't mean this algorithm is very intelligent; it's only very capable if you have a sufficiently fast computer; although indeed brute force is only feasible for the simplest tasks in practice, so the idea that algorithms (of increasing sizes) enable (greater) intelligence is definitely a part of the story, but not the whole story. You can also think of DNA, which represents a recipe for our bodies and brain, which we then use (essentially as an "algorithm") to learn things, with degrees of freedom (memory) far surpassing what DNA stores.
Now if you had a very good chess program running in very constrained (dynamic/RAM) memory, then that'd be partially more revealing. From a cursory search there's a 1800 ELO engine for the C64, which seems very impressive but very far from the best human players.
I'd be interested to see a curve of ELO x Avaliable RAM for the best chess engines (up to given RAM), and how that compares to other games and activities.
On RAM vs ROM (program size) memory, I think at a high level dynamic memory helps you keep track of search paths in a large tree search, saving you some computation. Program size tends to enable improving the effectiveness of your search heuristic, as well as pre-computing e.g. initial and final game optimal moves (potentially saving arbitrarily much compute). I like thinking about those things because I think the search paradigm is pretty informative of computation (and even intelligence) in general. Almost every problem is basically some kind of heuristic search in some kind of space. And you tend to get better at things by refining your heuristics (usually through some experimental training process or theoretical insight), considering more options, exploring deeper consequences, etc..
I think what really defines humans isn't really our ability to solve problems or play chess well etc. (although that's extremely useful and also enjoyable most of the time), it's really our emotions and inner world. We are not really Thinking Machines in essence, we're Feeling Machines most significantly. The thinking part is a neat instrumental part :) We can delegate thinking to machines but what we cannot extinguish is feeling or the human "soul", because that is the source of all meaning.
Oh my god the source is so tiny! It's really hard to parse because of it being minified but I love it to bits.
Good job! I love how you obfuscated your code, really in a spirit of FOSS!
Oh well, the file initially looked like https://github.com/datavorous/sameshi/blob/7ab4e47144f96becd...
It is hideous now!
It's not about being hideous, it's about being useless.
Your code is useless to anyone that wants to contribute, or maybe make something better by improving on the idea.
Coworker: “hey if you have a second, I have a one-liner PR open”
The PR:
The mailbox board representation is a good call for size-constrained engines. Bitboards give faster move generation but the manipulation code (shifts, masks, magic numbers for sliding pieces) eats a lot of bytes. With mailbox you just need offset tables and a sentinel check for board edges. Curious what your evaluation function looks like though. At 2KB you can't fit piece-square tables (that's 384 values minimum for both colors), so are you doing material-only eval or did you squeeze in some positional heuristics?
The gap between your 1200 Elo in 2KB and the TCEC 4K engines at ~3000 Elo is interesting. That extra 2KB buys a lot when it goes to better evaluation and move ordering. Even a simple captures-first sort in alpha-beta pruning costs only a few bytes of code but can roughly double your effective search depth.
Codex or Claude Code?
none.
scribbling long enough on a piece of paper is more enjoyable than prompting.
a thousand times this.
Isn't it bad enough they beat us at chess, do you have to make it even worse? ;p