[Outside the Box] Can brilliant Grandmasters crush Stockfish?

It's not a psychological barrier, humans simply have zero chance against modern engines.

@BrilliantAdam said in #5:

@KenulL_76 I already knew that Zug won against Stockfish8 (I actually was watching his stream when that happened) but that game doesn't prove anything due to the lack of time for the engine to think and also not optimum hardware (lichess engine servers) no brilliancy needed to defeat engines under time pressure indeed.

Stockfish level 8 is a handicapped Stockfish that plays suboptimal on purpose. Stockfish level 8 is not the same as Stockfish 8, it's much much weaker.

@BrilliantAdam said in #10:

1.Engines follow their numbers to get a material advantage in the future. Engines could reject (overlook) a long combination line because they followed their numbers to an extent where they didn't find a material advantage so stopped calculating (they actually might stop before the winning move by a move or two and doesn't matter how easy it is in the end BECAUSE THEY DON'T UNDERSTAND, they follow numbers to a fixed extent).

2.Humans understand chess, see ideas, calculate and execute them.

This only betrays your ignorance of how engines work. Engine evaluation is much better than just some material count. Engines are much better at finding long combinations leading to a positional advantage than any human.

@BrilliantAdam said in #10:

Alphazero isn't an engine, it's an AI

That's wrong. It's a chess engine, and either all chess engines are AI (although some are bad), or none are. The measuring criteria of AI is the achieved ability (playing chess well), not the underlying methods used to reach it. NNs don't make an engine inherently more AI, and unlike what you seem to think, NNs are all about number crunching. Stockfish is the strongest chess AI.

Consciousness is not a calculation. ~ Roger Penrose

@IndigoEngun said in #12:

It's not a psychological barrier, humans simply have zero chance against modern engines.

That's what the whole discussion about. I'd love to hear your opinion Don't you think there is any chance a human could play better than AlphaZero did?

Stockfish level 8 is a handicapped Stockfish that plays suboptimal on purpose. Stockfish level 8 is not the same as Stockfish 8, it's much much weaker.

True, Stockfish level 8 and Stockfish 8 are two different things, We're on the same page so where is the issue?

This only betrays your ignorance of how engines work. Engine evaluation is much better than just some material count. Engines are much better at finding long combinations leading to a positional advantage than any human.

Engines are divided into 3 categories:
1.Classical (rely on hard-wired rules and brute-force calculation of variations).
2.Neural networks/AI (AI engines use self-play reinforcement learning to train a neural network).
3.Hybrid (Both Classical and AI)

"Chess Engine" as a term I use it to describe the classical engines that had no AI implemented at all (before 2019), Stockfish 12 was the first version to use AI and turn to be a hybrid engine (I didn't know the information of the hybrid thing exists until I made a research after reading your comment), so that's why I was thinking and referring to all engines as classical engines (As you said, my ignorance betrayed me).

That's wrong. It's a chess engine, and either all chess engines are AI (although some are bad), or none are. The measuring criteria of AI is the achieved ability (playing chess well), not the underlying methods used to reach it. NNs don't make an engine inherently more AI, and unlike what you seem to think, NNs are all about number crunching. Stockfish is the strongest chess AI.

As I said, I had no clue that Stockfish is using AI now. Thank you for adding up new information and ideas to me.

@BrilliantAdam said in #14:

That's what the whole discussion about. I'd love to hear your opinion Don't you think there is any chance a human could play better than AlphaZero did?

None. The strongest possible human play would be to play a memorized top line and have the engine stay in book long enough that a draw may be possible, but there is no creativity in it. And it's simple to make an engine that will have slightly more speculative play creating complications that will make any human lose.

True, Stockfish level 8 and Stockfish 8 are two different things, We're on the same page so where is the issue?

Earlier posts in this thread were not clear in this regard, and I've seen many people on the internet making the confusion. I thought it's important to clarify.

Engines are divided into 3 categories:
1.Classical (rely on hard-wired rules and brute-force calculation of variations).
2.Neural networks/AI (AI engines use self-play reinforcement learning to train a neural network).
3.Hybrid (Both Classical and AI)

"Chess Engine" as a term I use it to describe the classical engines that had no AI implemented at all (before 2019), Stockfish 12 was the first version to use AI and turn to be a hybrid engine (I didn't know the information of the hybrid thing exists until I made a research after reading your comment), so that's why I was thinking and referring to all engines as classical engines (As you said, my ignorance betrayed me).

This is more accurate than your earlier take, but you still have some misconceptions.

Artificial Intelligence and Neural Networks are not the same thing, but I already made this point earlier. NNs are a tool that can help to create AIs and challenge some difficult problems we didn't have good method for, but AI is larger than NNs, and the field of chess engines has been considered AI for decades.

The second is brute force. Engines look at stupendous amounts of positions, and this calculating ability is obviously one of their key strength. This is true even for the slowest NN engines (such as Leela and A0) that still look at tens of thousands of positions per second. But doing a lot of calculation is not what define brute force. There is a technical meaning and a less accurate layman meaning, but both qualify an approach that's simplistic and relies on raw computing power. Although the evaluation of strong classical engines is quite inaccurate (despite best attempts from humans at finding simple rules, but still much better than naïve evals), their search is where the true magic happens. The difference between the search of Stockfish and naïve alpha-beta is more than a thousand elo. Search is very refined and sophisticated.

Consider that Stockfish's search, that was designed first and foremost for chess, has been lifted and used by japanese Shogi engine authors, and all top Shogi engines are more or less Stockfish search with a custom evaluation on top (that's where NNUE was born). Not what you'd expect from a dumb brute force method.

Chess engines gained much more elo during the 2000-2019 era from software improvement (a large majority of it happening in search) than they did from hardware improvement (although these did help testing, and enabled the creation of NN engines).

The third (to a lesser degree) is your point about hard-wired rules. It's true of course that a compiled version of a classical engine has hard-wired eval and search. And projects like Leela are training a huge amount of neural nets trying to create stronger nets with more training data. You have thousands of nets, and from one to the next you'll have slight changes in favored moves, evaluations, blindspots. But it has its own set of "hard-wired rules". The algorithm used to build the search tree based on the eval and move preference output is hard-wired for example. And of course a given net will compute the same millions of matrices the exact same way every single time it sees a given position. Unlike a human, who will often tend to try out different things in a given position, Leela and other such engines will stick to one idea (usually good, but not always). Ironically, multi-threading variability means that classical and NNUE (hybrid) AB engines have less predictable play, given a position with several close moves there is a luck element to which will be picked.

But back to your main topic: all three kind of engines (handcrafted eval + classical search, A0-like, NNUE-eval + classical search) are more than capable, for their strongest representatives, to give pawn odds to a grandmaster and destroy him in a match.

Look at the odds matches that have been played by Komodo/Dragon against Nakamura and other masters to get an idea.

Well, thinking outside the box a little, yes a grandmaster could beat the top chess engine. All they have to do is start with rook odds and play carefully.

@IndigoEngun
I'm truly grateful for your time and effort explaining that to me, your comment was extremely helpful and instructive.

I consider myself very lucky to be a part of lichess community where discussion could be really constructive and helpful unlike most of current communities.

Defeat the fish? Maybe here and there.
Crush? Thats a bit too much to ask for.

Re. Morphy, Alekhine and Fischer: I suspect that if the likes of Kasparov can't beat chess engines, then these other player (and Capablanca, while we're naming names) probably couldn't win either.

@Shreksify said in #4:

@BrilliantAdam
Does Fisher, Morphy, Alekhine ..etc have a gun?
It better be or It's stockfish in three minutes less.
As for Stockfish 8 vs Alphazero
Today's Stockfish (15) Is going to smack Alphazero like a fly on a swatter. Place your bet? A hundred to zero. Stockfish 8 was still able to beat Alphazero. Even if the statistic shows 290 games lost. Stockfish won 24 games.
Stockfish 15 vs Stockfish 8, who's gonna win? Even Stockfish 8 cannot weasel a draw in contrast to that 886 games that were dead heat against Alphazero. Alphazero sure was special considering that it ran on a multi-million (company) machine. It also had NNUE which was a novelty during that time. You might find some Youtube videos where it shows Alphazero playing against Stockfish 15, Dragon 3, or even Leelazero. To those who believe that it's Alphazero who is playing. It's not, It's a dupe.

AlphaZero was the best chess engine of its day but I think Stockfish is still catching up. In that hundred-game match against Stockfish, it won 28, drew 72, and lost zero.