chrisw wrote: ↑Fri Jan 02, 2026 3:21 pmIt’s every engine newbie that tells us SF is somehow fundamentally flawed and his new wonder approach (often padded out with BS, philosophical rambling and unprovable terminology, as here) is going to discover some deep patterns miraculously that SF (after aeons of development) is unable to see through either idiocy or just ineptitude on the part of its developers.
Most of the time, incremental progress works better than entirely new approaches. Over time, incremental progress gets you a long way.
Sometimes, though, a new idea simply makes the old ideas obsolete. Then incremental progress begins again from the start of the new idea.
Human chess is partly about tactics and strategy, but mostly about memory
chrisw wrote: ↑Fri Jan 02, 2026 3:21 pmIt’s every engine newbie that tells us SF is somehow fundamentally flawed and his new wonder approach (often padded out with BS, philosophical rambling and unprovable terminology, as here) is going to discover some deep patterns miraculously that SF (after aeons of development) is unable to see through either idiocy or just ineptitude on the part of its developers.
Most of the time, incremental progress works better than entirely new approaches. Over time, incremental progress gets you a long way.
Sometimes, though, a new idea simply makes the old ideas obsolete. Then incremental progress begins again from the start of the new idea.
But empty ideas are empty ideas. Come up with something. Until then it's all rambling.
hgm wrote: ↑Fri Jan 02, 2026 8:28 amThis is not just hypothetical. LC0 trained for high Elo plays like crap in Knight-odds games. Just training for high Elo keeps it oblivious from even the slightest idea what you can do best when you are down a piece early in the game.
OK, I agree that finding the "best" move in a non-artificial position that is objectively clearly lost is an interesting and useful goal.
I could imagine having different networks for "undecided" and "decided" positions.
However, an alpha/beta-type engine might simply not be the best choice for finding moves that might trick a winning opponent into a draw. The better its search, the quicker it will realize that no move can prevent the position from going further downhill in the next 10-15 moves. So I doubt that loading a specially trained NNUE network into SF could do the trick.
So there is room for chess programmers to develop a new type of engine (in so far as Leela Odds hasn't already solved this problem). Perhaps one should try to train an LLM for this. Something that is inherently bad at a deep tactical searches.
Engines like Dragon do fine in odds games, this trend of performing poorly in odds games seems unique to Stockfish, and might have something to do with training methodology itself being conceptually flawed.
hgm wrote: ↑Fri Jan 02, 2026 8:28 amThis is not just hypothetical. LC0 trained for high Elo plays like crap in Knight-odds games. Just training for high Elo keeps it oblivious from even the slightest idea what you can do best when you are down a piece early in the game.
OK, I agree that finding the "best" move in a non-artificial position that is objectively clearly lost is an interesting and useful goal.
I could imagine having different networks for "undecided" and "decided" positions.
However, an alpha/beta-type engine might simply not be the best choice for finding moves that might trick a winning opponent into a draw. The better its search, the quicker it will realize that no move can prevent the position from going further downhill in the next 10-15 moves. So I doubt that loading a specially trained NNUE network into SF could do the trick.
So there is room for chess programmers to develop a new type of engine (in so far as Leela Odds hasn't already solved this problem). Perhaps one should try to train an LLM for this. Something that is inherently bad at a deep tactical searches.
Engines like Dragon do fine in odds games, this trend of performing poorly in odds games seems unique to Stockfish, and might have something to do with training methodology itself being conceptually flawed.
No the trend of performing poorly in odds games is not unique to stockfish.
A lot of NN engines do poorly with odds.
The exception is NN engine that does not do poorly with odds.
Berserk13 also does poorly with odds and it seems everybody copy some bad ideas from stockfish only because stockfish has a bigger elo.
Training a network for exactly reproducing win probability in principle desroys the possibility for the engine to convert a non-trivial win. E.g. a KBNK mate typically takes some 60 ply, but, as long as no B or N is blundered away, is always a certain win. There is no way a search would reach 60 ply if it has no guidance for what to prune; without pruning you must be happy to reach 8 ply. And there is no guidance, neither for prunig as for striving for intermediate goals (like getting the bare King into the corner), as all non-sacrificial end-leaves will have an identical 100% score. This reduces the engine to a random mover biased against giving away any material. No way its random walk through state space will ever bring a checkmate within the horizon. At some point it will stumble into the 50-move barrier, and there is nothing it can be done to avoid it.
A corrolary is that when an engine is trained to reproduce the true win rate on the games played by the latest version, it can never achieve the exact win rate. When it is still ignorant it will discover that it will have a higher chance of winning in positions where the mate is close, just because it has a larger chance to stumble on a position that has it within the search horizon. Those will then get higher score, which can be seen by the search from further away, so that the win rate there will also improve. But it has to keep a score gradient large enough to guide the play towards the mate from far away. And there is some minimum gradient for which this still works, because there will always be evaluation noise. So you will get some equilibrium, where positions far away from any mate only get, say, 90% win rate. Because that will indeed be their win rate, because the gradient from 90% to 100% is spread over so many moves that it has become so small that it fails to follow it in 10% of the cases.
hgm wrote: ↑Mon Jan 05, 2026 9:56 am
Training a network for exactly reproducing win probability in principle desroys the possibility for the engine to convert a non-trivial win. E.g. a KBNK mate typically takes some 60 ply, but, as long as no B or N is blundered away, is always a certain win. There is no way a search would reach 60 ply if it has no guidance for what to prune; without pruning you must be happy to reach 8 ply. And there is no guidance, neither for prunig as for striving for intermediate goals (like getting the bare King into the corner), as all non-sacrificial end-leaves will have an identical 100% score. This reduces the engine to a random mover biased against giving away any material. No way its random walk through state space will ever bring a checkmate within the horizon. At some point it will stumble into the 50-move barrier, and there is nothing it can be done to avoid it.
A corrolary is that when an engine is trained to reproduce the true win rate on the games played by the latest version, it can never achieve the exact win rate. When it is still ignorant it will discover that it will have a higher chance of winning in positions where the mate is close, just because it has a larger chance to stumble on a position that has it within the search horizon. Those will then get higher score, which can be seen by the search from further away, so that the win rate there will also improve. But it has to keep a score gradient large enough to guide the play towards the mate from far away. And there is some minimum gradient for which this still works, because there will always be evaluation noise. So you will get some equilibrium, where positions far away from any mate only get, say, 90% win rate. Because that will indeed be their win rate, because the gradient from 90% to 100% is spread over so many moves that it has become so small that it fails to follow it in 10% of the cases.
Are you saying an engine can have precise, accurate analysis, or dominating play strength, but not really both at once?
hgm wrote: ↑Mon Jan 05, 2026 9:56 am
Training a network for exactly reproducing win probability in principle desroys the possibility for the engine to convert a non-trivial win. E.g. a KBNK mate typically takes some 60 ply, but, as long as no B or N is blundered away, is always a certain win. There is no way a search would reach 60 ply if it has no guidance for what to prune; without pruning you must be happy to reach 8 ply. And there is no guidance, neither for prunig as for striving for intermediate goals (like getting the bare King into the corner), as all non-sacrificial end-leaves will have an identical 100% score. This reduces the engine to a random mover biased against giving away any material. No way its random walk through state space will ever bring a checkmate within the horizon. At some point it will stumble into the 50-move barrier, and there is nothing it can be done to avoid it.
A corrolary is that when an engine is trained to reproduce the true win rate on the games played by the latest version, it can never achieve the exact win rate. When it is still ignorant it will discover that it will have a higher chance of winning in positions where the mate is close, just because it has a larger chance to stumble on a position that has it within the search horizon. Those will then get higher score, which can be seen by the search from further away, so that the win rate there will also improve. But it has to keep a score gradient large enough to guide the play towards the mate from far away. And there is some minimum gradient for which this still works, because there will always be evaluation noise. So you will get some equilibrium, where positions far away from any mate only get, say, 90% win rate. Because that will indeed be their win rate, because the gradient from 90% to 100% is spread over so many moves that it has become so small that it fails to follow it in 10% of the cases.
Are you saying an engine can have precise, accurate analysis, or dominating play strength, but not really both at once?
The enemy's king is hiding in a desert with lots of high dunes obscuring vision at ground level. Two different companies of soldiers will attempt to kill him:
1. Stockfish (SF)
2. Quick checkmater (QC)
SF parachute into the middle of the desert, quickly getting close. Now they're there, though, they don't know which direction to travel in: they just have to guess, try a direction, and hope to uncover some clues.
QC start their assault from the edge of the desert, carefully charting their position as they go. This will lead them directly to the king - but it's too obvious and predictable, so they will get ambushed and killed.
Human chess is partly about tactics and strategy, but mostly about memory
hgm wrote: ↑Mon Jan 05, 2026 9:56 am
Training a network for exactly reproducing win probability in principle desroys the possibility for the engine to convert a non-trivial win. E.g. a KBNK mate typically takes some 60 ply, but, as long as no B or N is blundered away, is always a certain win. There is no way a search would reach 60 ply if it has no guidance for what to prune; without pruning you must be happy to reach 8 ply. And there is no guidance, neither for prunig as for striving for intermediate goals (like getting the bare King into the corner), as all non-sacrificial end-leaves will have an identical 100% score. This reduces the engine to a random mover biased against giving away any material. No way its random walk through state space will ever bring a checkmate within the horizon. At some point it will stumble into the 50-move barrier, and there is nothing it can be done to avoid it.
A corrolary is that when an engine is trained to reproduce the true win rate on the games played by the latest version, it can never achieve the exact win rate. When it is still ignorant it will discover that it will have a higher chance of winning in positions where the mate is close, just because it has a larger chance to stumble on a position that has it within the search horizon. Those will then get higher score, which can be seen by the search from further away, so that the win rate there will also improve. But it has to keep a score gradient large enough to guide the play towards the mate from far away. And there is some minimum gradient for which this still works, because there will always be evaluation noise. So you will get some equilibrium, where positions far away from any mate only get, say, 90% win rate. Because that will indeed be their win rate, because the gradient from 90% to 100% is spread over so many moves that it has become so small that it fails to follow it in 10% of the cases.
Are you saying an engine can have precise, accurate analysis, or dominating play strength, but not really both at once?
The enemy's king is hiding in a desert with lots of high dunes obscuring vision at ground level. Two different companies of soldiers will attempt to kill him:
1. Stockfish (SF)
2. Quick checkmater (QC)
SF parachute into the middle of the desert, quickly getting close. Now they're there, though, they don't know which direction to travel in: they just have to guess, try a direction, and hope to uncover some clues.
QC start their assault from the edge of the desert, carefully charting their position as they go. This will lead them directly to the king - but it's too obvious and predictable, so they will get ambushed and killed.
Unbelievable BS. Chess is not life nor war, both of which are massively complex, unbounded and anything can happen. They are neither two sided, take it in turns to move operating on a tiny space of 64 squares with pieces that can move in six ways only. The "chess is life" concept is and always was a nonsense fantasy. Tal included.
FireDragon761138 wrote: ↑Mon Jan 05, 2026 3:16 pm
Are you saying an engine can have precise, accurate analysis, or dominating play strength, but not really both at once?
Not really. It depends on how you interpret the score. Interpreting the maximum of the score range as 100% winning probability would produce this dilemma. But if you interpret a score (say) halfway the maximum score already as certainly won, (and train the engine's evaluation accordingly), you still have room to use the rest of the range to indicate progress towards actually converting the win (based on expected remaining game duration). That would of course also have to be trained, by using the game durations rather than just WDL statistics.
