Daniel, there is an interesting Concept that I have brought up on a couple chess channels and that has not yet found much traction: Would it not be even more ground-breaking, if AlphaZero self-plays an order of magnitude MORE games with roughly an order of magnitude LESS calculating ability? Mightn't this Concept unlock more positional secrets and repeatable themes??..This AZ puts itself in positions that humans often cannot calculate their way out of..Why?? Because AZ knows that it CAN...and so what if a New AZ would More Often put itself in what you yourself might refer to as Comprehensible or Practical over-the-board positions?...This Concept might lead to an order of magnitude more prescriptive chess understanding..then again, is the chess world Ready to go That Nuclear??...or to risk finding a Chess-Singularity??? -Huge Fan and Art Theorist
Best AlphaZero commentary I've seen. Really clear for the 1800 level player like myself who isn't that skilled but can follow deeper play if explained clearly. Thanks a ton, very fun to watch.
Well and it does not have a need for closure or certainty, like humans do. We like to know for sure, that we are fine or better. But in such a position, where you sac a piece for long term compensation, you cannot be sure as a human. If you misassesed one variation, or one feature of the position, you are immediately busted...
on the contrary this is actually a very human way of assessing a position which is what is scary. you could almost call it 'having a feel' for the position.
Reshevsky jokingly said that his style was somewhere between that of Tal and Petrosian. In Alphazero's case, this is probably true: it is a fantastic positional player, master of both prophylaxis and attack (at the same time!), and makes long-term "intuitive" sacrifices. A truly beautiful machine.
I find this whole thing so incredibly interesting. Great commentary as usual. I've decided to do monthly support on patreon for this channel, and I recommend you all do the same. We need to keep the best chess channel on RU-vid on the air.
Justin, I truly find it strange that very good chess RU-vidrs generally have between 5,000 and 50,000 subscribers even after the channel has been up for years. Then some channels like MatoJelic and Agadmator have 100,000+ subs but the players are under 2100 and have a very low understanding of chess and barely anything to teach the viewers. Chess is just about the only skill game I have seen where people learning don't want to learn from the best on RU-vid. I don't know why that is, but I am glad you support Danny's channel. I am sure it is a lot of work for him. Not easy.
A0 is almost romantic in its style: piece activity, developing, material sacrifices for attacks without immediate recouping of material. Engines may have encouraged a dry, conservative style in human play, but A0 may just change that, finding explosive, dynamic positions that traditional engines avoid.
It's not taking any risk. It's calculates and CAN SEE, that the bind (positional or tactical) is stable or can only increase. It can also see, that there is no way to break that bind, otherwise it wouldn't play the move. It's like comparing 2300 player vs 2600 player. Both can calculate equally well, but 2600 can evaluate way better than 2300 player.
Wasn't it Paul Murphy who dominated, even gave the opponent one of his rooks before the game, because he knew it didn't matter in his attack? AlphaZero seems able to determine ways to lock away opponents pieces and also dominate with less material on the board. Truely a paradigm shift.
I have become acutely aware of how and when to give back material to complete development and get equality and easy game to play. It mostly happens with the queen side pieces where you struggle to develop the bishop or the knight and thereby locking in your rook or even getting uncoordinated pieces when you manage to develop them. When you face a strong opponent, and you get some material advantage causing a potential queen side stifle, look for ways to return ill-gotten material quickly. It's very embarrassing to lose with material advantage. At least, for me, that's my motivation. Don't lose a game with such embarrassment.
Guess one important advantage that AlphaZero has is it does not have "horizon effect". The fact that it sacrificed the knight on h6 goes to show that it "calculated" way beyond the Stockfish did. I am just guessing here, but that to me is causing SF to lose
I love Mr. King's sober and professional demeanor. If a patzer like myself would comment on this game, I'd scratch my head at Qh1 and say, "OK, let me try to explain this." At c4, I'd simply go: "Alright, I'm done."
I don't really get it - Alpha Zero against Stockfish and on move 18 Komodo thinks, white is better?! I imagine all the chess engines sitting around, commentating the game...
Is this the most outrageous, ridiculous series of moves ever played in a game of chess? For me, I have never ever seen anything as calm and vain as Alpha Zero.
Playing for piece power is how many of the great world champions of the past have won their games. In ways a game like this gives us all so many extra possibilities in many openings. If in this case, would we be brave/instinctive enough to play this way?
I must hand it to you Daniel, this is a good presentation. Though I did like the video "How Stockfish 8 Could Have Drawn Each Game Against AlphaZero" by IM Kislik where he showed with very deep (depth 40+ ish) analysis how Black could have drawn in this game and that it was just a practical sacrifice. That, to my eyes shows the beautiful depth of chess. Yes, AlphaZero did something cool, but logically and fundamentally Black was okay and could have drawn. Too many people got carried away in trusting the result. I think GM Boros also had a video saying the same thing as Kislik later on.
Just wondering, what kind of sentences would come out, if you could "make the beast talk"... that is connect it to a "language module", that uses common language (English sentences), to teach us, what AlphaZero thinks about a certain chess position!
Did somebody try to pass positions before 26. Qh1 and before 32. c4 to Stockfish or other “conventional” engines to analyze positions for white? How does Stockfish evaluate these moves (first line, centipawns)? We can say that these unobvious moves are something specific to AlphaZero only if other engines failed to find these moves.
Thanks Daniel, amazing video. I think it's a case of dedicated hardware and software is compiting to software made for a lot of hardware. So, the results seem logical to my eyes. I was reading that the stockfish version was for a year ago and without an opening book and a final table. That was a mistake for Deepmind because Stockfish was running at a suboptimal level. Maybe they should make another match with (Komodo, Houdini or Stockfish) playing at full capacity. But the results and the games would be less spectacular. The main objective of the paper is achieved, which shows the potential and high level of understanding of the neronal self-learning network. But for chess lovers, the fact that the engine runs at a non-full power is a small dark spot. Maybe we never know because i think that there will not be an AlphaZero in every home, like there was not a DeepBlue. Have a Happy Holidays Daniel, i deeply enjoy your channel, Cheers From Argentina.
With conventional engines at full power, they would still have the problem of horizon, which for A0 seems to be close to non-existent. So i am not sure if it would change much at all. Maybe the strange ways in which Stockfish treated some of the already lost positions would be prevented.
SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.
A lot of the horizon problem should be fixed with the finals table. Stockfish 2017 is +50 in ELO that Stockfish 2016, and a optimized opening book for the engine is +30-40 in ELO, and using table bases for endings is +60-70 in ELO, the results should be diferents, so the Deepmind guys should put AlphaZero to run in more TPUs and get the same difference in ELO!
A0 is not at full strength either or fully tuned.. That is the point of saying it was "only" trained for 4 hours. If openng db was a problem or causes sf to lose, it would be easy to point out what move was a opening mid or late move blunder in postmortem analysis cause by unfair handicap with a "100%" sf. I only read sf fans handwaving whining but no hard evidence where stockfish made a mistake because of Match conditions. They only have hurt feelings and no computer analysis even weeks later.
Great videos, Daniel! Lovely English accent in addition to keen chess analysis. My only suggestion: Please add the opening played, when appropriate, to video descriptions. Thanks!
GM King, you said you think, if it's made available, Alphazero may change the way professional chess players play. Do you think that'll be a good thing or a bad thing? Just thinking of all the draws in the early part of the recent London Classic.
Players these days don't play, but simply out-prepare each other. The one who can memorize better is the one with advantage. Fischer was right. Chess960 would be the real way to compare strength of players and not how good they (or their seconds) home-prepare the games.
Has anyone asked what the result would be if two identically set up A0's played each other? That's what I would like to see, not just A0 whomping on stockfish (with the emphasis on fish). I think that the end game being played by the owners of A0 to monetize it would put A0 itself to shame. I also would be willing to bet that there are buildings full of people in Moscow and Washington trying to figure out all of the military applications.
@@JamesGadbury AlphaZero lost 3 games in the 100 game match and that was against a weakened SF8. Google only published wins since it was a promotional event, impressive nonetheless but NOT unbeatable.
Indeed. I think AlphaZero is awesome, and I'm very excited that it managed to master chess to this extent with the amount of time it had. Still, I think all of the hype journalism surrounding this result is just that - PR hype. There was an amateur test conducted with the handicapped Stockfish pitted against a better optimized version, with similar results as to the AlphaZero match. The lack of time management in Stockfish especially handicapped the engine I think, as well as the very limited 1GB hash (it would realistically need 16, 32, probably even 64 GB hash to work as intended - but then, it has never been tested with such powerful hardware)
SF still manages it's time well at various time controls. By the way time management is a non-factor at fixed time/move controls. The book-Even when Stockfish followed theory in these published games AZ outplayed SF after the opening. EGTB-Stockfish had lost positions before the EGTB would be of any use. Hash transposition size-Try it yourself give SF a big hash and see how long it takes SF to see that AZ's sacs were sound. Based on what I have seen in the published games my theory why AZ outplayed SF is that AZ has vastly superior move ordering. This is supported by the fact that SF was doing 70,000,000 nps while AZ was doing only 80,000 nps. Even if SF has a huge transposition hash table that won't be enough to compensate for much inferior move ordering. Inferior move ordering results in too much time wasted on pointless variations. SF will miss crucial variations because of that.
Does Alpha-Zero really use the Monte-Carlo Tree Search algorithm? I find it very hard to believe, and the 4-hour thing is certainly not true. A pure MCTS engine should almost always lose against any decent player, not to mention against stockfish.
Thanks for the wonderful video but I disagree. Human ideas and methods for playing chess is shaped by our mental capacity. We cannot play like Alphazero because at the absolute best we can remember just nine slabs of information at the same time in short term memory (this includes Magnus and all the super GMs). When we get tired or stressed, we remember even less. The game of chess is popular for ever more because at 8x8 squares, it tests the upper memory limit of every possible human that is born. Future AI's will show that chess was a little over engineered by our ancestors so that we would always have a challenge. From the AI's perspective, it will show us a couple of styles of play we could not have foreseen and that is it.
Impressive but chess is way too complicated to think A0 solved the chess or cracked it. In a game vs chess game itself :-) A0 would be destroyed 100:0. Also I agree that match was not fair - too many limits imposed on SF and A0 in the full throttle. Thoughts about match differ - according to some SF played very badly with all these limitations put in place whilst others considered A0 overall performance rating of more than 4000 which is mind blowing. Nevertheless very entertaining games but for me as chess lover game itself is safe from being solved of any other similar term applied by any existing or future engine or AI. It's complexity is simply too much to handle even for silicon beasts or any type of AI.
Such scattershot explanation...one can't hardly follow...BECAUSE YOU JUST BOUNCE SO FAST THAT ONE CANT KEEP UP WITH YOUR STARTING POSITIONS... TERRIBLE EXPLANATORY
I think the reason why Alpha Zero beats Stockfish 8. Is One they took away its Book Openings .Second No hash,Stockfish 8 was given 1 GB hash table size When 32 or 64 or More .The hardware could make a great difference . You need to have at least 1024 of hash in order for Stockfish 8 to play very strong . .No endgame table bases. The pounder is off. You need to run Stockfish 8 on at least 64 cores . Mean while Alpha Zero is running on a super computers. Stockfish 8 did not . This game is not a fair game at all .
