- Blind mode tutorial
lichess.org
Donate

Why Leela Plays Bad Moves (That Work)

#2

I don't know if studying amazing engine moves solves the problem of those who want to develop in chess, you will probably even understand these complex theories, being able to do them in practice is what will not happen, unless you make use of engines, when cheaters in Lichess tournaments try to play without an engine, it is a real disaster, I believe that this also applies to those who are GM, memorize some theories and obviously in OTB tournaments they have no way to resort to help, so it is clear that there is no chess by memorization, it is practice, practice, practice, thoughts of a beginner who thinks he knows something, lol...

I don't know if studying amazing engine moves solves the problem of those who want to develop in chess, you will probably even understand these complex theories, being able to do them in practice is what will not happen, unless you make use of engines, when cheaters in Lichess tournaments try to play without an engine, it is a real disaster, I believe that this also applies to those who are GM, memorize some theories and obviously in OTB tournaments they have no way to resort to help, so it is clear that there is no chess by memorization, it is practice, practice, practice, thoughts of a beginner who thinks he knows something, lol...
#3

That's fair; while it's fun to notice when Leela uses unusual strategic setups, it's worth pointing out that Leela also silently is ducking all your tactical blows and raining down fire upon you whenever you miss something. No human can imitate that level of tactical vision.

That's fair; while it's fun to notice when Leela uses unusual strategic setups, it's worth pointing out that Leela also silently is ducking all your tactical blows and raining down fire upon you whenever you miss something. No human can imitate that level of tactical vision.
#4

You know what's impressive, the marketing makes us believe that the engines are almost human beings, but in fact they are just giant calculators, according to one of the creators of Leela, https://matthewsadler.me.uk/, she reached such a high ELO level that GMs can't understand her strategy, obviously after so much meeting one day they will be able to, but then she will also have gone further, and the great irony is that a calculator won't teach you how she arrived at the calculations and formulas, it's a job that only humans can do, that is, ultimately an AI is just a program that makes summaries, being ironic, Kasparov thought he was retired when Deep Blue beat him, it turns out that now he will die working trying to decipher literally the millions of lines of Leela variants... lol

You know what's impressive, the marketing makes us believe that the engines are almost human beings, but in fact they are just giant calculators, according to one of the creators of Leela, https://matthewsadler.me.uk/, she reached such a high ELO level that GMs can't understand her strategy, obviously after so much meeting one day they will be able to, but then she will also have gone further, and the great irony is that a calculator won't teach you how she arrived at the calculations and formulas, it's a job that only humans can do, that is, ultimately an AI is just a program that makes summaries, being ironic, Kasparov thought he was retired when Deep Blue beat him, it turns out that now he will die working trying to decipher literally the millions of lines of Leela variants... lol
#5

Really enjoyed reading this. The differences between Leela and SF play are fascinating and I think provide some neat insights into the kinds of errors humans of different ability tend to make. Great stuff!

Really enjoyed reading this. The differences between Leela and SF play are fascinating and I think provide some neat insights into the kinds of errors humans of different ability tend to make. Great stuff!
#7

La resistencia es inútil - The Spanish Borg (Nobody expects the Spanish Borg!)

To me the interesting bit here is how differently skilled players would react to this. I have no idea what pawn structures are about, so the joke's on Leela. My reaction would probably be a catchy "Leela invented a new opening!" YouTube-like thing.

One of the largest differences between humans and computers is that we struggle to come up with a plan, then tend to stick to it, even if it becomes detrimental later. Computers just go through moves, coming up with new plans every move, or basically not having any.

I would love to see a computer chess engine that would be plan based, guiding its search space based on previously concocted plans and being reluctant to change them. I believe this is "human play", even at low levels where you just want to take the opponent queen and you miss they have mate. And the way to exploit the humans would also be to figure out their plans, then confidently play "brilliant" moves which are objectively bad, but in the context are devastating. How else would computers learn to gambit?

La resistencia es inútil - The Spanish Borg (Nobody expects the Spanish Borg!) To me the interesting bit here is how differently skilled players would react to this. I have no idea what pawn structures are about, so the joke's on Leela. My reaction would probably be a catchy "Leela invented a new opening!" YouTube-like thing. One of the largest differences between humans and computers is that we struggle to come up with a plan, then tend to stick to it, even if it becomes detrimental later. Computers just go through moves, coming up with new plans every move, or basically not having any. I would love to see a computer chess engine that would be plan based, guiding its search space based on previously concocted plans and being reluctant to change them. I believe this is "human play", even at low levels where you just want to take the opponent queen and you miss they have mate. And the way to exploit the humans would also be to figure out their plans, then confidently play "brilliant" moves which are objectively bad, but in the context are devastating. How else would computers learn to gambit?
#8

I'm not that skilled, but when you realize the engine is targeting your king by moving your queen/bishop/rook toward it, applying pressure, if you unexpectedly change its position, the engine loses its way. The engine also can't stop you from forcing the knights to switch, which greatly complicates its plan, because they know how to move knights divinely, and for a human, it's very difficult to calculate the knights' movements several moves ahead. And it lets you make the switch because it doesn't perceive your intention, it JUST follows the theory of the best possible move. So when the human inverts this logic, they manage to trick the engine at some point. Obviously, in the end, you lose, because the engine, unlike you, never makes mistakes. Now, what's the purpose of an engine that never makes mistakes, when chess is precisely the game of mistakes, even a GM makes mistakes? I understand this as anti-game, because the challenge of chess is to win by discovering your opponent's mistakes and, above all, AVOIDING your own. It seems to me a lot like the issue of the 1st division of Formula 1 compared to the 2nd division. In the 2nd division, the cars are all the same and the driver needs to prove his worth, while in the 1st division, stockfish is used a lot, the cars are practically driven from the pit, the driver just holds the steering wheel.

I'm not that skilled, but when you realize the engine is targeting your king by moving your queen/bishop/rook toward it, applying pressure, if you unexpectedly change its position, the engine loses its way. The engine also can't stop you from forcing the knights to switch, which greatly complicates its plan, because they know how to move knights divinely, and for a human, it's very difficult to calculate the knights' movements several moves ahead. And it lets you make the switch because it doesn't perceive your intention, it JUST follows the theory of the best possible move. So when the human inverts this logic, they manage to trick the engine at some point. Obviously, in the end, you lose, because the engine, unlike you, never makes mistakes. Now, what's the purpose of an engine that never makes mistakes, when chess is precisely the game of mistakes, even a GM makes mistakes? I understand this as anti-game, because the challenge of chess is to win by discovering your opponent's mistakes and, above all, AVOIDING your own. It seems to me a lot like the issue of the 1st division of Formula 1 compared to the 2nd division. In the 2nd division, the cars are all the same and the driver needs to prove his worth, while in the 1st division, stockfish is used a lot, the cars are practically driven from the pit, the driver just holds the steering wheel.
#9

@TotalNoob69 said in #7:

La resistencia es inútil - The Spanish Borg (Nobody expects the Spanish Borg!)

Haha yes, that's right, our chief weapon is surprise...surprise and fear...

It's a great point about sticking to plans vs. being fluid. I have seen Leela display some remarkable adaptability in plans, especially in some of its unusual pawn structure play. The way LeelaRookOdds handles 1.f4 as White is really fascinating—it chooses between a number of contrasting formations depending on how its human opponent responds.

I wonder if an engine based more on an LLM architecture (where it may have learned whatever it knows about chess primarily through scraping PGN data and human-written books) could display that plan-based behavior you're speculating about. That might be interesting to see.

@TotalNoob69 said in #7: > La resistencia es inútil - The Spanish Borg (Nobody expects the Spanish Borg!) Haha yes, that's right, our chief weapon is surprise...surprise and fear... It's a great point about sticking to plans vs. being fluid. I have seen Leela display some remarkable adaptability in plans, especially in some of its unusual pawn structure play. The way LeelaRookOdds handles 1.f4 as White is really fascinating—it chooses between a number of contrasting formations depending on how its human opponent responds. I wonder if an engine based more on an LLM architecture (where it may have learned whatever it knows about chess primarily through scraping PGN data and human-written books) could display that plan-based behavior you're speculating about. That might be interesting to see.
#10

@HarpSeal said in #9:

I wonder if an engine based more on an LLM architecture (where it may have learned whatever it knows about chess primarily through scraping PGN data and human-written books) could display that plan-based behavior you're speculating about. That might be interesting to see.

One could use an LLM (but who would invest money to train billion parameter models for chess and then run it on very expensive GPUs?) to tweak the value function of a chess engine. It would also make it really slow. I believe the solution is more like tweaking the Monte Carlo search to be "greedy" towards promising directions. So go 140 moves deep in some directions, but ignore 10 ply deep lines in other. So maybe use it in the pruning algorithm?

I don't know.

@HarpSeal said in #9: > I wonder if an engine based more on an LLM architecture (where it may have learned whatever it knows about chess primarily through scraping PGN data and human-written books) could display that plan-based behavior you're speculating about. That might be interesting to see. One could use an LLM (but who would invest money to train billion parameter models for chess and then run it on very expensive GPUs?) to tweak the value function of a chess engine. It would also make it really slow. I believe the solution is more like tweaking the Monte Carlo search to be "greedy" towards promising directions. So go 140 moves deep in some directions, but ignore 10 ply deep lines in other. So maybe use it in the pruning algorithm? I don't know.