AlphaZero v Stockfish (2018): Lecture by GM Ben Finegold
ฝัง
- เผยแพร่เมื่อ 28 ธ.ค. 2023
- This lecture was recorded on December 20, 2023 in Roswell, Georgia. Thank you to the Zatloukal Family for sponsoring!
Games/Positions:
09:07 AlphaZero vs Stockfish
24:56 Stockfish vs AlphaZero
37:01 Stockfish vs AlphaZero
Check out Ben's Chessable courses here! www.chessable.com/author/BenF...
If you're interested in sponsoring a lecture of your choice, email Karen at karen@atlchessclub.com
Signup or gift a chess.com Premium membership to get access to their lessons, remove ads, improve your chess, AND help Ben at the same time!! Upgrade your chess.com membership or join! - go.chess.com/finegold.
Buy Merch today! ccscatlmerch.com/
Watch live at / itsbenandkaren
Donations are appreciated at: streamlabs.com/itsBenandKaren...
Follow me on Twitter: / ben_finegold
#benfinegold #chess #AlphaZero #Stockfish - เกม
V. Kramnik: 'Alpha 0 beats Stockfish 8? With only 9 hours of self-learning? And without openings book? How interesting ...'
It has a chip in its butt lol
Beating stockfish is easy. It’s cheating to use an engine. Stockfish is an engine. I automatically win as a result. Checkmate stockfish.
But can you prove that Stonkfish ls cheating?
Also: Cheating basically means getting help from a third party? Like when I as a 500 elo fish get hints from a master while I play some ofishel chessgame like a real nerd and stuff.
Stockfish doesn't cheat. Its chess speaks for itself.
It's also illegal to get human assistance during a game, so wouldn't you get disqualified too? : )
And if the assistance provided by your own brain doesn't count... you see where it is going then; does Stockfish use itself?
Congrats on using the word "suffice" twice in a row in a grammatically correct sentence. Achievement unlocked.
GM Finegold at his funniest while talking about computer chess.. Way better than the stockfish channels cheesy scripts
That channel is just loads of AI generated crap. The games they show are probably just Stockfish vs Stockfish itself.
The main reason A0 is more willing to sac material is that since it's using AI/NN to evaluate the position, it can more or less see the compensation, where as SF can't due to the horizon effect. Very similar things happen in some of the SF v Leela0 games. One of the main things at play I think is long term piece mobility.
Has anybody consulted Vladimir Kramnik and his mathematicians about these games?
Kramnik is several hundred points lower rated. His opinions dont apply here. lol
come on, millionaires, sponsor GM Finegold analyzing all 1000 games in full
best comment 2023
I don't even think they made all 1000 games public, I think just around 100 or so are available
AlphaZero was great for it's time. But I don't think most people realize Stockfish is way stronger than AlphaZero ever was now.
I think StockFish would roflstomp AlphaZero even from the starting position. We live in crazy times.
For real. It incorporated techniques from AlphaZero, just didn't replace itself with them entirely. Current stockfish is like old stockfish and AlphaZero combined.
It just takes time for Alpha Zero. It has a much stronger computer that Stockfish will never see. It a matter of minutes Alpha Zero will always surpass Stockfish.
Wasn't Alphazero trained for just 5 hours? It's just that they stopped after it became the best in the world and moved on to other things, but had they continued, with 5 years of development/training it's hard to see how Alphazero wouldn't be on top today...
@@kingsolo6241
Alphazero is not a computer and neither is Stockfish. Both are computer programs.
Alphazero advantage at the time was the fact that it was a NNUE (Efficiently updatable neural network) Once Stockfish updated it's evaluation function too it surpassed AlphaZero. And Since then Stockfish has gotten way stronger still.
AlphaZero doesn't have any advantage now, while Stockfish has the advantage of being open source and is constantly tested by it's community.
Not only did I stay awake long enough to see the Stockfish win, I also stood awake long enough to write this comme
Show ALL the games! Five THOUSAND minutes with Ben!
Great lecture. I remember the AZ headlines and even saw a video of your first relentless example. Your explanation filled my swiss cheese understanding of why AZ isn't mopping up the chess world and leaving the game in tatters. Again my thanks for giving much needed perspective. Your series is quite refreshing.
Oh yes! I was waiting for these!
Great video. Love the Berlin Defense. They are never as drawish as top players make it look. Especially this endgame is very tricky. I actually like much less the lines that don't head to this early queen trade. The old mainline. Then the structure is just symmetrical. The dull exchange french basically.
Thanks GM!
Those endgames by AlphaZero remind me of Capablanca, but on steroids. Capa loved to push his pawns and gain space.
Thanks , very instructive . Your comments are so funny. Keep going
IIRC, the resignation rule was that both engines had to agree that one side was up by at least 10 pawn-equivalents. So, one engine could be forced to play on, even though the other engine was the only one that had hope for its position
Just like human chess 😂😂😂
Go Ben!
But stay right there!
hi Ben, this game is awesome, its almost as awesome as the games you play are!
Thanks, Ben! Great video as always.
You might be interested to know that I've just downloaded a free version of LeelaZero, and my first test was to setup board position: 1. d4 d5 2. Nc3 Nf6 3. Bg4...
Onwards from here I let LeelaZero play both sides. Strangley enough it accepted a 3-fold-repetition at move 29. Maybe I'm supposed to tweak some settings, I don't know.
the setting to look for is "Contempt", which basically just tells the engine how much it should avoid draws, or in other terms, how much worse of a position it's willing to accept to avoid a draw (on the assumption that it can outplay the opponent in the long term and win anyways)
@@sage5296 oh yeah. I think i remember ben using that word now. Thanx!
There is a very good reason for not using opening book.
chess engines at that time played at high depths and evaluated positions based on materials (Queen 9 points, rook 5, pawn 1) or simple rules which were coded into the engine by humans.
This made chess engines great at tactics but chess grandmasters would still outplay them because the engines didn’t understand strategy.
This is the reason stockfish uses opening book to compensate for lack of understanding chess
People at that time used to believe that learning chess was therefore a sign of human intelligence.
The point AlphaZero was trying to make was an AI can learn to play chess at a grandmasters level with zero human training just by getting the rules of the game.
Pitting Stockfish against AplhaZero without opening book is a test of “Artificial Intelligence” learning strategies that beat calculations
Even in 2018 the eval functions were more sophisticated than what you outline.
I would guess they used a genetic algorithm with parameters as genes to converge upon optimum values, and to determine which to jettison to improve speed.
Your claim that engines were great at tactics but they would still be outplayed by grandmasters due to a lack of strategical understanding is just plainly wrong. Even years before this match, the best engines had reached a point far beyond human ability to match. There might have been deficiencies, but they were not going to be outplayed by humans.
I guess I could frame it better, the main point is after finishing AlphaGo, they made AplhaZero which didn’t use human games and trained entirely on self play aka it learned long term strategies on its own. They could quickly train it to play other games than go (including chess)
They were not trying to compete with stockfish and that’s why they quickly moved to other things instead of improving AplhaZero unlike stockfish
Evaluating based only on material will give you a significantly inferior engine than Stockfish 8. HCE was the result of three decades of development and this is just disrespectful
You've oversimplified it to the point where it's blatantly untrue.
The rules were anything but trivial.
If you ever looked into (old at least) Stockfish source code, it contained a huge pile of "manual" corrections - surprisingly specific, like "in an endgame with just Queens and minor pieces value the mobility of minor pieces 16% less" etc. Not literally ofc, but this sort of stuff.
Of course it's still heuristics (albeit very complex and multi-layered), and it doesn't quite convey the notion of long-term strategical sense (this part is correct).
But modern engines were very far from evaluating positions only based on the material value of pieces.
This is really unfair to their creators (who btw often were strong players themselves, or strong players assisted with the development).
If you built an engine based on such simple criteria, it would never be stronger than an amateur.
It's true that engines relied on opening books because they weren't very good at grasping the strategical implications of opening choices (the "fog" was too dense from the starting position).
But still, even if you disabled an opening book, an engine would still play a pretty strong opening based on general principles (mobility, control of the center etc.).
The main reason for including an opening book is just optimization.
Similarly, (serious) engines use endgame tables, even though endgame is "pure" calculation, and relatively easy at that (as you can calculate much farther when there are only a few pieces left on the board). But the reason is optimization. In an engine vs. engine match time matters: just like in human chess, they both have only got so much time allocated for the game.
So why waste time on computations when you can replace them with the equivalent of "home preparation".
Stockfish 16 really likes what AlphaZero did and predicts pretty much all moves. I expect it's about twice as good as Stockfish 8.
8 * 2 = 16
If you can now tell me what "twice as good means" :'D
@@nilsp9426 the answer is fries.
Way more than twice as good but ok
@@DarthVader-wk9sdHow do you quantify twice as good
Obviously elo only went up some hundreds of points. Not to 6000
What does it mean to be twice as good? A 2 to 1 win probability? Then you could use the elo formula to compute the exact elo gap, should be about +100 points.
Most versions of stockfish will accept the berlin draw, so magnus(or anyone else) can draw it with white, but I don't know that i'd really count that, but if a GM said that I'd assume that's what they meant, since it's obviously absurd otherwise.
46:44 my guess AZ saw the idea of promoting the H pawn and didn't want to open the a1-h8 diagonal for the dark square bishop, so that he can put his rook to h8.
Notice how the Bishop on f2 can't get to f6 or g7 any way
So, I was curious about the last game and it turns out that Ben was right about a couple of things. I let SF 14 take a look at the position where, at 47:18, Ben says that it looks to him like it should still be a draw and the newer SF engine agrees, and when Ben says that, after white pushes h5, he would go g:h5, well, he's absolutely right. The way to draw was to exchange pawns on h5 and then go Kc4, with the intention of bringing it to d3. Instead, Kb3 that the A0 played instead of g:h5 straight up loses, the eval jumps from something like 0.03 to around 3.5.
Incorrect off the bat. Alphazero is/has an engine. It's just that at the time AZ used neural net tech whereas SF didn't use NN tech, (SF started using NN tech starting with SF 12). Technically any machine that plays chess has an engine built into it.
Does anyone know a game of this match (or the one from 2017) where AlphaZero sacrificed and Stockfish won by defending and holding on to the extra material? Like, did it ever misevaluate / overestimate its attacking chances?
Precisely this, it lost most of its games by attacking in a drawn position, but A0 believed it had an overwhelming attack but stockfish had a strong defence to counter.
My nasty comment is that Janet Jackson had a song years ago called "Nasty Boys".
Lol. I dont think anyone even knows who that is now!
I wonder why at 30:30 stock fish chose not to capture the dark squared bishop with the knight eliminating the bishop pair. Is the light squared bishop that much stronger in this position, maybe because of the A2 target?
It wouldnt have used human reasoning like you and I. The answer lies in huge branches of calculations
8:31 Some versions of Stockfish will gladly play the Berlin draw most of the time with black so you can get lots of easy draws that way
At a certain point black's best idea is to give away all his pieces every move in order to stop white's attack. That's why being single minded about gaining a material advantage was the flaw in the armor of Stockfish 8. At a certain point, it was willing to trade all that material back to stop getting its ass beat, and considered giving away material like that to be its best move. Alphazero was definitely on the right track with its judgment call that the attack was strong enough even at a 4 pawn deficit to be worthwhile.
nice
i like BF discussion about compensation and judgement
Remember, stockfish, at that time, was calculating around 4 million positions a second to alpha zeros 80 thousand... If AO were to have the same calculating power as SF then no stockfish would be able beat Alpha Zero😮
They probably had the same computational power on hand. The thing is that A0's evaluation function is much more complex and much more computationally demanding than that of a traditional engine, so if A0 is to match SF8 in term of positions per second, A0 has to have many, many times more computational power at its disposal
They do have the same calculating power
From my look at AlphaZero vs. Stockfish it seems A0 is increadibly good at punishing any setup with b7, Bc8 and e6. Black not getting their bishop out is such a common theme in these games.
Interesting
If I remember correctly, when they didn’t have the opening book, both “engines” decided that the Evans gambit was optimal.
Which is awesome because I only play 1.e4 with the hopes of being able to go 4.b4
Extremely incorrect. Maybe at very low depth, but certainly not at the massive depths they are typically run.
@@ethangilworth7891
You’re right.
It was just alphazero that did it. Not stockfish.
@@jefftaylor1186 I’m an Evan’s player myself, so I would love to see your source for this.
AlphaZero actually flipped the board before Stockfish was able to play Be7
22:00 black decides that the white bishop is worth more than the black rook, and wants to trade those two pieces. White also decides that the white bishop is worth more than the black rook, and decides to take the black rook with his own rook instead of with his bishop. Both computer programs agree that white's bishop is worth more than a rook. It's not three points versus five points at all, each piece has a fluid value depending on where it is placed in the current position. Sometimes it's basically worthless, like when it is trapped, or when its only value is in stopping checkmate by blocking an attack and putting itself in an absolute pin to stop the attack. Sometimes, like this bishop, the piece is so strongly placed that it is worth more than a supposedly much stronger piece that is having to play defense to stop the assault of the bishops. Being passive and forced into the role of being a meat shield for the king truly lowers the value of a piece that normally gets its value from the range of spaces it can move to and the speed at which it can attack an opponent's position. Lowering that mobility and putting it on passive guard-only duty really wrecks the value of the piece, to the point you'd gladly trade it away for the minor piece that is utterly dominating your king.
One thing i love about computer chess is when the engines radically disagree about the evaluation.
Didn't stockfish "resign" if its own evaluation of its position was like, -5 whole points or something like that? I may be getting the number wrong, but I thought they called the game lost if either program felt it was losing by a sufficient amount. I could be confusing this with a different computer chess matchup event, but I know I didn't invent this out of whole cloth, I definitely read or heard about it somewhere.
edit- literally one second after I wrote this comment you mentioned white resigned. Great timing on my part. /sarcasm
_"Didn't stockfish "resign" if its own evaluation of its position was like, -5 whole points or something"_
They can set the resign score threshold for the match. Once it reaches that threshold (e.g. -8.00), and that score stays or worsens over three more subsequent moves, then it resigns. You can download a good chess gui like Arena 3.5 (it's free) and run your own matches with Stockfish 8 and other newer NNUE engines.
I would have been more interested in the opening sequences
I always wonder at what point, what move, the engine realizes it was wrong and its losing. What moves did one engine miss that the other saw.
With Stockfish, it uses Alpha Beta pruning for its tree search, so it saw every move. The question is more “what move did it evaluate incorrectly.”
@@ethangilworth7891 not exactly, it can easily be the case that it prunes away the key line too early. Ig that would be an incorrect evaluation to some degree but non-NN engines rely more on seeing far ahead than evaluating accurately in any given position
@@ethangilworth7891 that's pretty much the whole definition (and the purpose) of pruning - to NOT see every move.
AlphaZero "cheated". Stockfish 2018 would have drawn almost all the lost games had it been given a proper computer to run on. The deep mind team put Stockfish on the crappiest computer they could find, the worst time controls Stockfish played at, and only published the results where AlphaZero won. Oh and they removed Stockfish's opening book and endgame database. AlphaZero on the other ran on a custom gpu-driven machine. The match was a complete farce.
Idk it’s pretty hard to come up with fair hardware for GPU/CPU engine matches. Even modern engine tournaments get criticized over them.
@@user-ws7kp1yh9l I agree it is hard to compare apples to apples CPU vs GPU, but look at what Stockfish was given vs. Google's machine. It was totally lopsided.
If you Google "Hardware used in AlphaZero vs Stockfish match", and read this line from the StackExchange discussion:
"[AlphaZero] used 4 TPUs for the games, so a processing power of 180 TFLOPS. Note TFLOPS = 1000 billion floating point operations per second."
The GPU network AlphaZero ran on was roughly 10,000 times more powerful than the hardware SF was allowed, and AlphaZero was *designed* to work without an opening book, whereas Stockfish was designed to work with an opening book an endgame database, and at classical time controls. The game structure was obviously designed handicap SF to the maximum.
OG EBN!
Nasty comments! Rawr!
37:03
I will allow the ads to run only for some FineGold.
You dont use ad blockers?
You are not cool Ben. There you go! 1:58
Stockfish 13 was the last version of stockfish to be weaker than Alphazero. Stockfish 14 was already considered stronger. Stockfish 13 and Alphazero do not maximize their advantages from the opening, while stockfish 14, 14.1 have near to complete drawable foresight to the game. Stockfish 15 does better sacrifices slightly. 15.1 and 16 reached a new super level by world record sacrifices and depth to the game. Leelazero, Dragon, and even Torch are a bit behind since they are still only around the level of brilliancy as Stockfish 15.1.
I'd agree with your assessment. When I tested SF13 by having it analyze some of AZ's wins, it couldnt find some of the moves that AZ played in critical positions. But SF15 and 16 are most definitely stronger than AZ. I wish someone would hack DeepMInd, steal the code for AZ, and make a a new "Frankenstein" version of AZ. 😂 I reckon it'll be even stronger!
@@CorteumGoogle Leela Chess Zero
I'd really like to show one of these games to someone like Capablanca or Alekhine (without telling them where it came from), i wonder what they'd think of it.: ) would they "get it"? would they think it was played by genius players? or dismiss it as nonsense
Kasparov's claim was that GM's on the Deep Blue team were inserting specific recommendations for certain positions, and the key position in Game 6 was one of them. GK knew his move wasn't sound but he knew DB wouldn't respond correctly - which it wouldn't have without the specific human input.
In any case, my favourite bit is that a few days after Kasparov bumps into Charles Bronson, they have a chat, GK says something like "I think I can beat it in a rematch". Bronson says "they aren't going to let you have a rematch". And he was right.
Y they didn’t let him rematch
@@archsys307because they won
They had never been obliged to keep on granting him rematches until he's satisfied with the outcome.
And why would they grant him a rematch anyway after he publicly accused them of cheating without providing a shred of evidence? "I just know that computers can't play that way" is an allegation, nothing else.
never play f3, never start a land war in asia, and never trust a headline
(and 7 times never kill man)
In final position of last game, after Bf8 Rxf8 . h8Q Rxh8 Rxh8 is still hard to break the fortress as there are no pawns available for attack.
A human brain works on roughly 20 Watts of power. I wonder how strong an engine is when limiting it's computing power to a cell phone's vs a GM. @Ben: thoughts?
Stockfish on your phone is already stronger than all GMs out there
Noted, but what if the brain power required to play chess is between 1% and 10% the computing power of a phone. Limiting computing power to that: Who would win? It'd be a cool experiment 😇
That black pushing pawn to e4 (move 28.) is an instant loss. It's like a ??? type of a blunder for a strong engine. There is really no reason to analyse the game any more since after that move it's practically a forced loss (even though engines of that era - 5 years ago didn't really have a clue).
My favorite game in this match is when Alpha zero plays as white in the Queen's indian
Ya engines don't get tired and won't mind playing a 24 hours long game because they ain't got no life. Just like me, except for the tired part.
Can A0 play F3 - only as white?
It can if it's forced to. But it will never choose to play f3 if left to its own calaculations.
21:14 hahaha 🤣
1000 games? Tyler1: Hold by beer...
Leela is also free right ?
Indeed. But you need a decent GPU to run it.
has Alpha Zero run away hiding from new Stockfishes
The Stockfish derivative engines (engines based on Stokcfish but with modifications), are stronger than standard SF, for sure. Have you played around with any of them?
And AlphaZero says no kibitzing.
ChessGenius for the win...
Ginger GM was pushing Harry up the board before A0!
So is this "Great Engines of the Past"?
Or ... "Great Systems of the Past"... No, I don't think we have a good word that specifically includes AI(ML) and also engines/expert systems, but is still narrow enough...
Ah well, I like the first idea, anyway
Chess engine is basically any program that plays chess. Pretty much all strong chess engines (including Stockfish) incorporate NNUE these days. So whether it's based on neural networks or manually encoded heuristics, it's still a chess engine.
"It seems like AlphaZero chooses the move that's most likely to win" is an almost spot-on description of how these deep-learning game AIs actually work. They don't really "evaluate" any position; they just simulate the entire game to the end and see what happened. My understanding is that getting AIs to give a traditional "-2.3" type of evaluation pretty much required a hack; what it's actually "thinking" is "from this position, Black wins 74% of games" or whatever.
my understanding is that it's basically using the standard alpha-beta pruning, but being much more selective/aggressive with the pruning. The AI part is the algorithm that evaluates a position, which is what basically assigns that %win or +- score. The tree is then pruned accordingly in the same fashion as SF would, except it checks less positions overall (often 100x or higher less positions), generally to a comparable depth tho
In other words, it plays moves towards a position that's most likely to win, but it still evaluates many moves deep
@@sage5296No, Alphazero does not use alpha-beta pruning. Evaluation is too computationally expensive for that
No, chess AIs (or engines based on neural networks) don't work that way.
They don't simulate the entire game to the end (that's known as a Monte Carlo simulation, this technique predates the AI as we know it, and it can only be executed with a rather low accuracy / randomness, due to computational complexity).
Quite the contrary - they've got a highly developed (via deep learning) evaluation function, which is essentially a black-box algorithm that "wrote itself" (by trial and error, in the training phase).
Noone really knows how it works, and what patterns trigger the AI into thinking a position is "promising" or "seeds of defeat". It's the same as with eg. image recognition.
Noone knows what exactly it was that made the AI recognize a dog in the picture.
AZ, in a way, played more like a human (human brains are biological neural networks, after all).
It uses highly sophisticated pattern recognition insteadf of a huge tree search with a slimmer (manually encoded) evaluation.
That's why it analyzes much fewer nodes per second than the likes of Stockfish. But its analysis is of higher quality, and it compensates for that.
@@vibovitold In fact, AlphaZero (and all similarly-architected game-playing AIs) use Monte Carlo Tree Search (MCTS) as a fundamental part of their training (and, it seems, during actual play, but see below). MCTS is exactly how AZ learns "by trial and error, in the training phase", and is indeed based on playing a large number of games to the end to determine win/loss efficacy for a given move, which it uses to update the network. It's surprisingly difficult to find clear info on whether current AIs use MCTS during play to choose candidate moves. DeepMind's paper on AlphaGo Zero (the Go-playing AI which preceded generalized AlphaZero) states specifically that it does not, but the AlphaZero description is unclear on the matter, and Leela Zero definitely does use runtime MCTS (i.e., "playouts" or "rollouts"). Apparently without them its Elo was only ~2500. In any case, though, the "evaluation function" of deep learning AIs does literally boil down to "Will this move lead to a win, on average?"
@@Rubrickety as part of the training, sure, i meant the "over the board" phase.
According to LZ website
"Many people call this MCTS (Monte-Carlo Tree Search), because it is very similar to the search algorithm the Go programs started using in 2006. But the PUCT [Predictor + Upper Confidence Bound tree search] used in AGZ and Lc0 replaces rollouts (sampling playouts to a terminal game state) with a neural network that estimates what a rollout would do"
so if it doesn't "sample playouts to a terminal game state", only "estimates what a rollout would do", it directly contradicts the claim that "they just simulate the entire game to the end and see what happened". no?
i do admit that while i'm a programmer with some interest in chess (and conversely chess engines), i am not working on chess engines, let alone an expert in the field
actually, 155-6, with even a million draws DOES qualify as a "crush"
+155 -6 =1000000 is just a 0.05 elo difference
1000 Games is just too small of a sample size.
Thats the old stockfish, 12 mybe, but now you have stockfish 15 who can easly defeat alpha
Rawrrrr
kf1!!
I believe what is cool about AZ is that it doesn't know the concept of material in a way that humans do and therefore Stonkfish.
what is cool is that AZ doesn't know ANYTHING at all. it only knows how to learn.
It didn't know the concept of *anything* outside of the rules. That's the amazing part. It's just here are the rules, play a billion games against yourself and see what works
Smartest fish in the world is Stockfish
Everyone is ignoring the elephant in the room. Alpha zero surpassed stockfish in a MUCH shorter development period. Stockfish had years of development at the time alpha zero was better, but alpha zero had mere months of development. So yes BECAUSE alpha zero was shut down in 2018 and stockfish continued developing for years after, stockfish is better now than alpha zero was then. This is like comparing Magnus Carlson to Paul Morphy or even Bobby Fischer. Yes, Magnus Carlson of today could beat them then, but they were so much farther ahead of their competitors than Magnus is from his competitors. Now, if we look at leela chess zero, an AI that has been in development for a few years instead of months, we see that it CAN beat the best modern stockfish. Now think about how much more advanced alpha zero was than leela and apply the same or faster advancement for alpha zero. I don’t understand how anyone would think that if alpha zero had continued development, it wouldn’t be the best chess entity in the world. It just doesn’t make sense.
Carlsen, not Carlson.
And the fact that every strong language (including Stockfish) includes a neural network now (NNUE) pretty much answers the question of whether the approach prototyped by AZ is superior or not.
It is even more obvious where you look at Go. Due to (much) greater computational complexity, there have NEVER been any "traditional" Go engines capable of beating the best humans (without handicap). No "Go-Stockfish" existed.
Which is why AlphaGo was tested against the best human players (beating them).
This has only become possible once deep learning / neural networks were used, as opposed to traditional algorithms.
The stockfish16(GM) on my phone says 2850. So, there's another version that is 3350? Very difficult to even get a draw with this one.
What app are you using
Chessify
@@mrcleanisin Chessify’s Stockfish seems to be handicapped to 2800 elo
If you want to play the non handicapped version use either Droidfish or SmallFish (depending on whether you use iOS or Android)
I'm not sure I understand you. I have chessify on my android phone that says it uses stockfish rated 2850. It beats me really easy, but I have gotten a draw when I take back some weak moves. I would imagine this chessify on my phone would beat Magnus Carlson. Why don't you try it?
StockFish doesn't know how to play chess. It uses a negamax function that mini maxes thousands of positions per second with alpha beta pruning and iterative deepening, by comparing static evaluations that come from a static evaluation function that doesn't necesserily understand how chess is being played, its just good enough to beat all Humans and most other engines...
But if a Human GM had the ability to remember a perfect oppening book, had the ability to have a perfect ending book, and the ability to calculate equal amount of positions per second and acess them perfectly at any second... Then the Human GM would always win against StockFish and he would in fact prove that the engine doesn't know what it is doing...
But a neural network does! A neural network indeed thinks like a Human, and sort of understands that chess more deeply... Of course only a saintient being could understand what chess is, but neural networks show sparks of AGI - artificial general inteligence, and that is much closer to playing stronger chess overall.
Maybe StockFish 16 or whatever we have now, is already a bruteforce - neural network AI hybrid and that is the only way to reach such high elo. And by StockFish 16 I don't mean the one you have on your phones...
Stockfish 16 uses NN and Stockfish 16 runs on your phone. It also has a branching factor of under 1.7, which is way less “brute force” than even human GMs
AlphaZero never existed (at least as a fully functional and strong chess engine). Nobody has seen it, the chess players were never allowed to "play with it" (although many of them wanted to), the source code was never released (why not? the engine is abandoned), and the notation of 862 games (out of 1072) which were allegedly played against Stockfish 8 were never published.
Classic google move
According to my info (my friend works at Google), "AlphaZero" was Hans Niemann disguised as a neural network.
In the last game, a GM playing black would've taken the draw by repetition. Probably all those 800+ draws went a lot like this. Which begs the question: why this game?
I think AlphaZero evaluated the positron as winning every time. By the barest margin. Somewhere deep in its behavior model, there's gotta be a place where it decided "when I'm winning by any margin, I refuse to repeat." And that's even if the move that breaks repetition is worse, as long as the overall evaluation is equal or in its favor. So it allowed its position to become worse from every refusal, until its usual model didn't work. Truly a 0.6% error.
Hey Ben,
Leela is also an engine just like Stockfish and they are both "AI". The difference is mainly how they evaluate positions.
And even though they are similar strength, I find Leela much better for analysis since it can give you stats (such as estimated win loss draw probabilities) that Stockfish cant (Stockfish "thinks" in centipawns). So I think the chess world/coverage could benefit from adapting Leela as the main engine.
Cheers, thanks for the content
Stockfish's algorithm was written manually. (It's got a NNUE component now, but it didn't back then).
What we refer to as "AI" in this context are implementations based on neural networks, meaning they're black-box algorithms which have kind of "written themselves" by trial and error..
No Talking
So many idiots in comment. Nice video Ben! Keep it up 😊
Leaving a nasty comment even before you started explaining anything and everything.
Having looked at the games, I can say that AlphaZero was better than SF8, SF9, SF10, and I'd say even SF11. But SF12 would be pretty close to AlphaZero's level. AZ was at least 3600-3650 FIDE classical level. SF8 was about 3400-3450 classical on the rating lists.
SF8 I think was, in reality, 3200-3250. Many of its moves were predictable. And still, Stockfish had years and years of development and learning and openings repertoire.
All of that sunk at the bottom of the sea when torpedoed by mere 9 hours of self-learning by A0. And A0 moves were utterly mind-blowing and unpredictable. We got an all new insight into the game.
@@zvonimirtosic6171 I think if you had SF8 competing in disguise in human elite events, it would wipe the floor clean. Former WCCC Rybka 3 was already over 3000. Rybka 4 was not quite as strong as SF8 and had a rating of around 3350. But i remember clearly SF8 having been rated at over 3400 on multiple ratings lists.
A good way to compare AZ with today's engines is to have SF12 or higher play a match against SF8 using the exact same openings as it played vs AZ. You'll learn a lot from those games and have a much better idea of how strong SF12 (or higher) is compared to AZ.
@@zvonimirtosic6171 I've gone through a lot of A'Z's wins vs SF8, and even SF12 struggled to find some of AZ's best moves. In some cases, it never found the moves at all. Only later versions of SF did find AZ's moves.
What I wanted to say, perhaps, is that chess engines do not perform at presumed ELO rating across all possible aspects of playing. On some aspects they are good and maybe better than their alleged ELO rating, but on some they are not as good as their alleged rating. Maybe this is "human way" of assessing engine's performance (which does not suffer from human issues, but it does have its own weaknesses). I do remember using engines before A0, and they played "predictably boring" hi-level chess, without blunders. But with A0, though, for the first time I saw something I never saw before, and which Magnus Carlsen called "something we can mistake for creativity". A0 approached the game in a wholly different way and that chess was exciting to watch.
@@zvonimirtosic6171 A0 is definitely something else. I remember testing Stockfish 12 NNUE on some of AO's brilliant moves, and SF12 struggled to find the correct solution. It wasn't until SF14 and 15 that it started to find A0's moves.
I agree that the top engines have their strengths and a few (very few) weaknesses. But in general, their performance on different aspects of chess is superior to elite GMs. In saying that, there are still some positions they struggle with that i've come across. But those are very few.
It would be interesting if they ran an average engine (e.g. Houdini 1.5a or Rybka 3 - both former WCCC's), using just 1-cpu and a small hash, in a top level human tournament, without the human players knowing about it. And just see what the approximate human elo level is for these engines. Then do the same with a more modern NNUE neuralnetwork engine. My guess is that a few GMs may get some draws against Rybka or Houdini, but probably no wins. Meanwhile, the NNUE engine will not lose, or even draw, a single game vs elite GMs. Maybe more advanced AI's in the near future will be able to accurately model the performance and style of different elite GMs from past or present, and then test to see how they might have performed against chess computers from past or present using different hardware configurations. Would love to see it. 👍
I have only watched 12 seconds of this video but I feel strangely compelled to leave a nasty comment for no reason. Odd since I usually damn this channel with praise coming from the likes of someone such as me. My approval brings you shame, etc.
nasty comment #1
Rawr!!
Stockfish 2018 ? That's to outdates, not worth watching.
obligatory nasty comment: *shakes fist*
They didn't use Stockfish 9 because they couldn't afford it, and so with Stockfish 10, and stockfish 11, and stockfish 12, and stockfish 13, and stockfish 14.
Only SF13 ror later wouldve had a chance against AZ.
@@Corteum I'm pretty sure Stockfish 16 is good enough to defeat her consistently.
@@sugarcravings1797 Most definitely. SF16 is super strong. But there are some SF derivatives that are even stronger.
@@CorteumDerivatives are just scam
Hi guys, I'm here to leave a nasty comment.
Very nasty comment. Nice glasses though Ben
Stockfish is just overrated!
Stockfish’s tournament wins speaks for itself
A very nasty comment
leaving a nasty comment for good measure
zero had over 100 cpu and sf 1 cpu
sf 8 elo 3400 sf 16 elo 3650
You seem to be confusing two things here.
AZ was TRAINED on a strong machne, but when it played, it didn't run on one anymore.
Training a neural network is pretty much the equivalent of home preparation.
@@vibovitold thanks for the info was something l read on an other post and did not look into it any more than that
Very nasty comment
booo!!! ben sucks!!!1111 🥰😍
But in the most enjoyable nevermind.
A bit much chatter and not enough analysis. Just get into it! Typically enjoy your work but not this.