When alpha zero started playing Starcraft, he literaly killed his units, to make room for another....everyone was little bit scared, imaging Alpha ledaing actual armies..
She is a retired Chess player, streamer, RUclipsr, can speak Hungarian, Spanish, English. She is one of the bestsoft spoken person I have seen in my life. 😊 Enjoy.
I concur. I'm just starting out in chess and most videos are way too fast. These videos are perfect and you never get flustered, lol. A real joy to watch and listen to.
e6, g7, h4, moves to A3 to sacrifice G6 easy to understand if your a chess player but if your a normal person this ain't even english! god damn elitist chess douchbags 😂
@@abdqs853 "In computer chess, a chess engine is a computer program that analyzes chess or chess variant positions, and generates a move or list of moves that it regards as strongest. " Alpha Zero can do this, thus is a chess engine, among other things.
It looks like AlphaZero's idea was to trap their higher level pieces even if it means sacrificing a pawn. It effectively took their pieces since they weren't able to move.
so typical of humans....trying to find water on Mars wasting a billion dollars in the process while being disrespectful of the abundant water sources on earth. Typical of humans.
@@vvenkat111 the reason to find water on mars Is to use it later on In conolies In mars, no to bring it to earth. And yes, earth is a paradise, but our chances of extintion are pretty high, if we colonize mars. those chances goes to 0
@@gergogaal568 I dont know about that. I think most players now and days rely too much on memorising defenceve stratagys and all the quick win moves amongst other things. Theres verry little originality in play style now and days.
Another way a0 wins over SF: We already know it likes mobility - and this is a huge advantage over SF-type engines. They don't do well with mobility because it increases their depth exponentially. a0 has basically learned to "stretch" SF and use its brute-force calculations against it :D
@@Duduvianna3 interesting. To be honest, when writing this, I was thinking about the kind of opponent used to train a0, which would have been anyway a traditional style (dare I say "brute force with strategy" calculator), of the same type as StockFish. So, maybe not StockFish itself but an opponent that plays the same?
Eduardo Vianna what I think is really interesting is that in the learning process there are multiple Ai instances of a0 playing each other at the same time. And there is this phenomenon that in certain badges of AIs that play each other certain metagames form where one of the AIs learns to prefer a strategy that is strong against what the other AIs in their badge plays and strongly overvalues that strategy because it wins so much with it. So the fact that A0 tries to learn to play on its opponents weaknesses has to be and is controlled for in order to make A0 learn the games instead of learning to play its oppenents.
Eduardo Vianna Eduardo Vianna no, from my understanding it‘s not that easy. just trusting that the natural probabilistic distribution let‘s THE strongest strategy come out on top would create loops and hinder learning. ruclips.net/video/eZCI7zu_DlM/видео.html check out this video from the 25:50 minute mark. It’s about the deepmind AI for Starcraft, but I am pretty sure this problem occurs in every highly complex game. The problem is that every strategy depends on what your opponent does and there are many different possibilities to act. That means that there is the problem that some sort of rock paper scissors situation developing, where strategy 2 beats strategy 1 and most instances start playing strategy 2 until strategy 3 gets discovered that wins against strategy 2 but loses against strategy 1. To prevent that the instances get caught up in a loop with no real learning progress, that cycle is disrupted by introducing instances that play a certain way every time and don‘t evolve in order to make a0 remember that it has to be able to beat those strategies as well even though they are not as popular within the other a0 instances. I agree with the second point. The learning happens within a* itself, although games at the top level are used to initialize the instances. They are basically used as some kind of starting point for the ais to develop from.
Eduardo Vianna Eduardo Vianna the paper you linked does not go deep enough into the process of how deepmind implemented the self-learning in a0. Moreover it refers to the algorithms implemented by Ago0 unless stated otherwise. hackernoon.com/the-3-tricks-that-made-alphago-zero-work-f3d47b6686ef This blockpost about Ago0 however specifically mentions multi-task learning with hard parameter sharing, which is pretty much what I mentioned above. On the deepmind website, there are also papers that deal with exactly this topic so I am still not sure where you take your confidence from that this common problem in ai self learning suddenly was not relevant for a0? And yes, they did not use external agents to overcome the problem I mentioned, but that is also not what I said. I said they used external games by top players to initialize their agents which is mentioned in the beginning of the video I linked.
Insane. It was literally watching black squirm around until it felt it was the right moment to attack. This goes against everything I have ever learned about tempo.
As an amateur player (rather poor at that) who is interested in AlpahZero and the recently publicized competition, I've got to say that Anna made a great selection of a game to analyze, and just brought a lot of insight and excitement to it. This should go down as a classic video. Way to go, and thanks, Anna!
@Nathan Sindlinger No, it's not. Neither design for basic fission devices it particularly hard to understand. It *may* be going a bit far to say you could build it in your basement but with the right person I'm sure it's doable. The problem is that the best way to get enriched uranium or plutonium is as byproducts of nuclear fission and building a controlled nuclear reactor is a much more complicated process. I'm not aware of marketplaces where the general public can just buy lumps of barely subcritical weapons grade uranium and plutonium, but if they exist I will stand corrected.
Should be careful. One company is setting out to ruin backgammon with the AI info. They deduct blunders from the score and say winning the game is not enough?! 😂 That's going too far when the whole point is to win the game
UPDATE: Hey folks! Thank you for your kind words about my AlphaZero videos! I don't have new AZ videos on this channel but I have created something even bigger: an entire chess course (Attack like a machine!) highly influenced by AlphaZero's play, with the main chapter dedicated to AlphaZero's attacking chess. You can check out the preview video and find out more about the course here: www.chessable.com/anna-1 *Also: it's an "it", I know I was wrong to correct myself and apologise often. You can check out in the rest of the AlphaZero videos if I had learned from it: ruclips.net/video/rkt51PTCQG4/видео.html
HEY!! Thank You!! I find the A.I. games very fascinating! I just watched the "AlphaGo" movie, and while I do not know the game "Go" very well, I DO know chess quite well. The thing for me is that I realise the machine has several advantages over the human brain in terms of speed, and THOROUGHNESS of analysis, processing, and calculation. This ultimately places the Gen.A.I. in a whole different class. This is why I fail to understand WHY some people actually get upset and or offended by the powers of the machine when they are defeated "IT". BTW Why are YOU struggling with the irrelevant gender of "IT"??? It is an IT, gender has NO Meaning here!!! LOL!! Anyway, I enjoyed this video. Thank you again. Oh BTW, do you know how I might obtain a good chess program for my P.C.? Free or pay no matter.
Very good descriptions. A lot of chess videos I don't feel like it's improving my game but the way you pause for a moment and ask what should be done at this point then explain out the possibilities really helps.
I think its interesting that AlphaZero uses known openings that noone programmed it with. Says a lot for book, but you also notice AlphaZero changes those openings pretty quickly too. this is absolutely fascinating. Anna thank you for your commentary on the game. I'm not an exceptionally good chess player so Its nice to have a professional player's interpretation of the moves.
+@c b My point is that it makes sense because when a human wins we win *despite* our flaws achieving greatness. How good a computer performs is completely and utterly irrelevant to how much chess players should be paid. A human can inspire other humans as we can emulate them and achieve the same things as they can. You can never become a machine however. At least not with current technology. Your comment don't make sense in my mind till human-computer interfaces reaches a point where we can freely transfer our minds too and from a computer and fully utilize a computer wirelessly with our mind. And even then chess might be interesting with such a interface turned off and played with our regular flawed pieces of meat.
+@@solomonlubega8337 Several of the novelities makes sense to at least *some* humans looking at the games. We're going to see more players emulate the best parts of AI play and traditional engines slowly adopt more and more of those into their code base till they reach parity.
Starts like Morphy, grinds it out like Magnus. Stockfish calculates; AlphaZero learns from experience, like a human. Who said that romantic chess is dead?
yeah that's what i also find so interesting about these leela zero and alpha zero vs engine games.. this friggin ai's play these engines like a gm an amateur.. the engine is trapped in its own position with no way of getting pieces into the game.. they don't care about any chess principles and somehow gain advantage due to some seemingly weird and pointless move and then bam 35 moves later they are getting the upper hand because of this.. the positional game of these ai's is so incredible.. and then you realize this is fucking stockfish they play against.. an engine so strong it plays every human like a fiddle with ease.. an engine you will never win a game against within your lifetime if you play against the strongest version..
To win in chess you need to checkmate the enemy king. The fact that we value the other pieces is just an assumption we made. AI engines don't make those assumptions.
@@charimuvilla8693 Or has it assumed that the opposition will do or has been programmed to do exactly that and then plays that knowledge back against it?
Having watched so many video commentaries of the AlphaZero vs Stockfish series, yours is by far the most illuminating and easy to understand one. I'd love to see more of those from you, for sure. Keep that up.
Thank you Anna for this analysis, you explained the game and its nuances perfectly but, you know, many chess analysts can also do that, what you managed to do that was unique for me was to also convey the excitement of the game. I found myself smiling along with your enthusiastic presentation and sharing your sense of admiration for AlphaZero's audacity. I liked the way that you anthropomorphize AlphaZero, it is clear that you see him as a bold, dashing hero and you made me see him that way too. Wonderful.
Probably my most favorite chess video I've ever seen. Just the enthusiasm alone. 22:10 I feel like the knight to d7 just hurts that bishop even more with its zero mobility, and Qc1 was typical AlphaZero protecting the c5 square, further imposing difficulties onto black. Clearly AlphaZero is just pure evil with its idc about giving you even a hint of a chance.
Köszönöm Anna, hogy ezt a valóban meglepő játszmát ilyen szépen prezentálta. Öröm volt nézni és hallani, hogy egy ember mennyire másképpen gondolkodik mint egy gép; mert hiszen bennünk számos pszichológiai tényező dolgozik, amik a döntéseink szerves részét képezik. Az egyén pszichológiai alkata és a személyisége a szűrő amin kereszül egy helyzetet megítél - még a sakkban is. Ezt egy neurális hálózat, mint AlphaZero nem tudja, de Stockfish8-ban benne van az emberi gondolkodás lenyomata. Nagyon érdekes volt. Köszönöm! :)
Gosh, I love your assessment of this game. It reminds of Bobby Fisher, he always did things that no one could understand. In fact in a tournament one time, he moved his knight into a position that everyone(chess experts), in the chess community watching the game live, gasp, called it a huge blunder. The chess experts explained how it has weakened his position, etc.. Then 9-12 move later the game was over Fischer destroyed his opponent with that knight move. Fischer too was superhuman.
Fischer 20-0 sequence with players from top 15 cannot be measured with Elo. His last three years were 100+30=8- with top players while playing reckless to win with a no draw mentality.
I rarely see commentary on games, but I've definitely been quite impressed with AlphaZero so thought I'd check this out. Your presentation was enchanting and I have a feeling that you're on the money about why AlphaZero does so much better than Stockfish- it is unhampered by human theories on how to play the game to win or at least draw. I imagine it has come up with its own set of theories on how to go about winning/at least drawing, theories that have far surpassed the human theories that are essentially tying Stockfish down to a lower bar.
Rather than "theories," AlphaZero has *intuition* built up by trying things repeatedly, and recording what did and did not work. It does not attempt to distill this into a rule set you could explain or communicate - it just "knows" what to do. Here's a summary article (though it would be nice if they described the "policy network" in part 2): www.chess.com/article/view/how-does-alphazero-play-chess
that was thrilling. a good sign of a good teacher is that they are enthusiastic about what they teach. it is really engaging and gets the student excited about the material also.
Oh the passion! A haven't seen ANYONE talk about chess like this! One can see in your eyes that you love the game and new discoveries it offers. This was really enlightening, even for someone who plays chess as a fun afternoon brain exercise. I caught your stream the other day and loved your way of teaching a friend to play chess. Keep up the good work and keep loving the game!
@@megamindblue1732 True. Videos are a form of passive learning, hence learning with a book and board along with many pauses in the between to check the position out for yourself is the best way to study. Do not waste your time watching videos.
She explains the moves so simply, beautifully and with such passion that even I as a 1500 player can understand and appreciate the beauty of this game. My new fav chess channel!
Oh my God! This is just mind blowing. Thank you so much for your detail explanation of this beautiful game. It is such an illustrative game for me that shows how we humans developed our own rules of chess based on our own weaknesses, mostly our limited ability of doing deep and precise calculations which keeps us away from playing what we consider risky movements like loosing a lot of material. As you said AlphaZero looks so confident entering in this terrain knowing that just improving and opening the position increase the odds of winning in the long term. It feels like we’re witnessing the God of mathematics playing chess for the firs time. Just amazing.
Just so you're aware, the newest chess engines, including Stockfish 12, would wipe the floor with AlphaZero, but that doesn't take away from the incredible achievement of AZ, and the amount of data that chess engine developers were able to take away from what AZ accomplished, and used it to strengthen their own chess engines. Of course, AZ was never a chess engine, and has not made any attempts to keep up in that department, it was simply a stepping stone for them in the grand scheme of machine learning.
@@joshs7160 thanks for your reply, I totally agree Stockfish and chess engines are in general extraordinary human achievements, and now Stockfish is also partly an AlphaZero achievement. We’re clearly walking into a completely new era not only for chess but for science overall.
Hopefully it sacrifices you first....or maybe you could be the final one sacrificed, so you get a front row seat to the extermination of half your species.
This is an amazing coverage and walkthrough of this game. I hope to see more content like this from you in the future. You have a talent in explaining high level chess in a very engaging way.
I understand your excitement. All those years we have made several assumptions like "more pieces is a good thing" or "rook if better than bishop" and it's good to see an AI ignore those rules and play in a purely strategic way without assumptions made by humans.
I loved your analysis of this fascinating game, Anna. So many "alien intelligence" ideas at work here - the novelty (ideas Stockfish doesn't see) and depth of strategic play is spellbinding. Although humans cannot play like Stockfish, I do wonder if humans *could* learn to play like AlphaZero if we could learn the same deep strategic principles and positional logic.
A0 to me is like a human GM that played chess for thousands of years and has gotten to the 3600 level. Of course it just played against itself for 8 hours to get to this level lol
I wonder if the human concept of a value system assigned to chess pieces is a result of capitalism ( everything on the planet is a commodity and has a value based on utility & scarcity ). Therefore humans have a flawed understanding of the game of chess by virtue of the economic system that we live under. I have an exciting idea... in Russia they start chess training around age 5 or something? For the next generation of chess players, remove the concept of a value system from the lesson and we can have some crazy chess games.
@@MrAgmoore All economic systems are value systems, thats what economic means. The value of pieces changes depending on the position. Alpha zero assigns values at a much higher level then the general principles we teach beginners. You talk like adding values to pieces is bad for the game, when thats literally how A0 learned. The values you're thinking of are just general basic principles to help out beginners. Economics has nothing to do with it wtf is wrong with you. Try teaching a 5 yr old how to play the game without telling them that they should usually value a rook over a knight, because in most games a rook will be more powerful than a knight long term. But no you're right, because capitalism assigns value to things that fluctuate based on the situation of the world (just like chess pieces change in value depending on their relevance in a position) we've been playing chess wrong this whole time! A0 knows that a rook is more valuable than a knight or bishop long term, but it also knows when it can sack material for a different long term advantage. The basic principles of general piece value still applies to chess at the GM lv and the A0 level, but so do many other principles. And the higher up the rating ladder you go, the better understanding of all principles is grasped. Denying probably THE most obvious and basic principle of the game is just ludicrous. Jesus Christ man, how about you learn a little about the world
I followed this game with Stockfish 10+ on lichess after we got halfway into it, it actually suggests many of the same moves (not all, but most - including the masterful pawn d5 move that SF8 didn't have in its top 10 moves) that AlphaZero took. Its time for another A0 v SF matchup
Thanks very much for this, Anna. One of the things I always enjoy in addition to your chess analysis and commentary, is that you speak perfect English. I speak a smattering of several languages, but am not quite fluent in them, and I am at a loss at how someone could master a language so well which is not their 'native' language.
Talk to native speakers as often as you can. Use the language you want to be fluent in frequently. Watch movies in the language you want to learn and use subtitles in the language too. All of these may feel like slow progress, but it is the fastest method that I know.
The scary part is even the NN developers don't quite understand how the algorithm works. Hopefully we have heard the last from people saying chess has reached a draw death. At last we humans have found a great teacher. There was an old master who considered that pawns just obstructed his pieces. Perhaps he was right all along. The engines are toast. Long live AI.
I think you're wrong about that last sentence. Given equal computational resources and equal knowlege of chess principles a traditional chess engine would come out on top because it makes more efficient use of the computational resources. The strength of a NN on the other hand is that it can add new chess principles on its own without getting those hardcoded in by a human. In other words, we're likely to see a back and forward shifting of the "balance of power" between NNs and AIs in chess as chess engine developers work hard to catch up with the NNs. Each time they catch up in encoded chess principles they'll be ahead till the NNs discover some new principles the engines are unaware of.
@@Luredreier The engines are weaker than AI because they rely on human understanding of chess. Alpha is beyond our understanding but we can learn better from it than brute force calculation engines.
@@proflaxis6968 That's not a correct understanding. One could easily rebuild brute force alpha-beta on a tabula rasa basis with modern computers, with far less input of human expertise (certainly nothing that required input from an expert player), and you'd probably make some gains, too, but it likely wouldn't prove decisive. Even the "old school" programs like Stockfish are tuned using self-play and advanced algorithms from machine learning. But certain trade-offs between the domain of the static evaluation function and management of the search tree are built into the code base-especially around the issue of quiescence-and no amount of self-play can improve these choices, without the manual intervention of a competent chess programmer. This is why some people on the classical computer chess side of this regard the fundamental innovation from AlphaZero as the Monte Carlo tree search (MCTS) algorithm rather than the neural network component. AlphaZero was _forced_ to adopt a fully general search algorithm, because there's no other way to fully automate self-learning. Now that MCTS has been proven viable at the very top of the heap, it's hard to say how much the old tradition might also be impacted. What you get from the old fashioned fast-and-dirty search trees is extremely good peripheral vision. You're rarely surprised by anything. (This was good in competitive chess, because it means you can't be sandbagged by clever preparation on the other side.) What you sacrifice is visual acuity on the primary lines. Your forte is to play solidly enough and never fall short of tactical excellence (capturing is almost always a solid move, if uninspired, so we saw a lot of that for twenty years). Leela, who is less well trained right now that the AlphaZero, already displays much of the same brilliance, but her peripheral vision is poor (due to her reliance on a narrow and slow search that's not yet fully mature) so she gets beaten by weaker programs far more than you'd expect for a program that sometimes trounces Stockfish decisively. With deep recurrent networks, the brilliance comes first, and the wisdom of years arrives later (this is also true of offensive-minded rookie defensemen in the NHL). Note how much emphasis you see in this video on AlphaZero's quiet moves. Leela hasn't got this trick yet-not to this degree-in the videos I've watched. But she sure shrugs off counterattacks (often at great material loss) to open up lines in an eerily non-human way. She's just raring to jump into the play-often to a fault. Compared to AlphaZero, she's the Karate Kid right now. Over time, it's expected that her inner zen will improve. She needs to augment her hyper-aggressive goal-crease directed style with just enough 360° peripheral vision to stay out of trouble (those blind behind-the-back drop passes that every gifted rookie short of Gretzky seems obligated to try out at least once per game). The legacy trap you cited is almost always dominated by cost-benefit effects. You _could_ go back and start over, tabula rasa, but it will be a long slog and dry spell before your new program can defeat your old program. The incentives for making this choice are rarely high. Especially when the programmers involved don't fundamentally believe that the legacy of human expertise is a particularly large problem; if tabula rasa Stockfish appeared likely to gain hundreds of rating points from the "purity" budget alone-fumigate all taint from those confounding narrow-minded human GMs-work along those lines would probably commence immediately. One of the reasons that Stockfish is so much more powerful than similar programs is that Stockfish employs a very carefully tuned and narrower search: more acuity, less peripheral vision. We observed that in this game, where Stockfish completely fails to see the central pawn push. Narrow search is potent, but risky. Of course, it's a huge win to have no human legacy at all (from either chess experts or chess programming experts). That's what Google sees in AlphaZero and the real reason AlphaZero exists in the first place. To reiterate: Stockfish could easily be rebuilt tabula rasa with no human chess expertise at all, and with modern computers you'd expect this to work out better, not worse (microwave ovens have more computing power these days than chess computers from the 1970s). But the re-investment of chess _programming_ expertise would be non-trivial, and merely erasing the negative legacy of human chess-expert tunnel-vision would not pay enough dividends to support a ground-zero rework on this scale. At least, not unless adopting MCTS at the same time changes everything on that side of the chess world, as well (without any contribution from deep learning or a neural network). Keep your eye on Komodo MCTS. Hybrid vigor could yet put traditional engines on a whole new track, all by itself.
@@afterthesmash Who wants to stick to black and white TV when they have seen colour? It seems obvious to me that Alpha sees at least three times further than the Fish. We humans, like Alpha, can see the long term benefits of early strategic decisions. Yes The Fish has excellent peripheral vision and can't be easily surprised but is fatally handicapped by its event horizon as Alpha is demonstrating. Its a shame Alpha can't explain its plans to us or how it comes up with them. But it is providing the best learning material ever produced so far. Hopefully our GMs can take it on board and get over the 3000 threshold.
Insane. It was literally watching black squirm around until it felt it was the right moment to attack. This goes against everything I have ever learned about tempo.
This is late... but its playing chess like it plays Go. its moves are designed to increase its potential to win if it plays perfectly, if it can increase its odds of winning by giving up pieces, it will since greater odds of winning is better then pieces. as people can see, it reduces the moves available to its opponent in any way possible, increasing its odds of winning, eventually as it plays out this results in the enemy having no place to move, or essentially 'checking' as many of their pieces as possible, which is sort of the point of the game. After seeing it play just this one game, its quite clear that its the correct/best way to play, but also the most difficult as its literally playing on a tightrope, if it were to screw up and allow pieces to escape, it could ruin its pin and result in a loss. its like playing with absolute confidence that you wont lose with every move you play, monitoring more pieces then seems to be humanly possible. It would be interesting to see alpha zero play halfway through the game if a human screwed up its game by moving a piece badly, what would alpha zero do? would it be able to recover?
@@jessiejanson1528 It's just incredible that moves like Kb1, throwing away all the extra time/tempo it had gained, where actually the most optimal moves it could've made. It's like it's some kind of sadist that just likes watching the opponent miserably fail to activate their pieces.
you can not compare an A.I. to a human player. humans are bound to make mistakes along the countless calculations of possibilities opening up. so it is advisable to use advantages and exchange pieces since you will mess your advantage up by overlooking some detail at some point. you can play like alphago zero if your game is without mistakes, flawless. but it isn't since you're human after all. try to play that way and get shattered.
Greetings from Ecuador. Loved the explanation and even though I am new to chess, the way you presented the game made it understandable. You gained a new subscription. Keep up the great work.
Great video, Anna. I loved all your comments, they really highlighted and brought to life the ideas and approach AZ had in this wonderful game. (Side note: Looking good, Anna!).
I'm analyzing this match with stockfish 10, and it immediately found that d5 move. Plus, SF10 sees that AZ is better and disagrees with some of SF8 moves, which is logical. I really want to see a match with that new version of SF
@@Rainnnny Stockfish 10 didn't exist when they were doing this match. Stockfish 9 didn't even exist. They did this near the beginning of 2018. It's a peer reviewed paper so it took awhile to be released.
Stockfish: Life should be planned carefully to be most fulfilling AlphaZero: IF YOU HAVEN'T SUFFERED OR FELT THE RAIN POUR AS YOU LAUGH, HAVE YOU TRULY LIVED?
I love this video because of Anna's skillful understanding and in-depth explanation; also Alpha-0 astounded. Chess will never be the same. Bishops on side near the back rank with the Queen close at hand to help and the Rooks on the opposite side to the Bishops is what I try to emulate. And sacrifice Pawns to open files. Truly a concert of Chess.
something wonderful about chess is that despite knowing only the rules, it is still captivating to understand the thought process and understand it once explained. Harder to do in Go.
This has been the most charming presentation of chess I have ever personally experienced. Anna Rudolph makes chess fun...and fun to watch and analyze! Consider me a fan. Look me up when you find yourself in LA. Prophylactic move. That's hilarious.
Jeffrey Gold She’s not using the word in its sexual sense, but its meaning as “a precautionary or preventive measure”! PS check out KWVE.com- excellent Christian teaching!
Satisfactorily & interestingly instructive a tactical & gambit presentation. Please feel free to post more of those AlphaZero chess games' beneficial video demonstrations.
After watching a bunch of these AI games, it is easy to find yourself starting to think like that: who cares about material! The most important thing is making sure your pieces have a lot of potential threats, and that your opponent's pieces are stuck in the corner. A bunch of opponent pieces huddled in the corner unable to move seems to be the AI engine trademark. We can all learn from this strategy :)
Can't know for sure what goes on in the AI's mind, but I think its game is more complex than just that. We humans, with our limited brain power, are only able to think of simple strategies - like "sacrifice material for position". I think AZ's model, even if encoded into its neural network and not being available as an actual mathematical model, is sort of numeric - its neural network has found a way to evaluate everything - material, position of its own pieces and of those of the opponent, and the relations between them on the board - as a single function. Its play is to just maximize that function.
@@a0flj0 Sure, but what we're talking about here is really the relative weights given to different types of advantage. AlphaZero and Leela seem to give more weight to closing off opponent pieces (especially bishops) vs. material than most humans, and thus also human-designed brute force heuristics. Of course when I say "most humans" there are exceptions...I've often heard AI strategy compared to Tal, and it seems fitting, but Tal would often make such "wild" moves without being able to consider the possible outcomes should his opponent act positionally. In effect, his games seem to exploit the human bias towards material over preventing "really bad" looking positions. But the AIs are trained against themselves, so it makes sense that the AI makes these moves in a slower, more long term manner. Similarly, the AI engines seem to value bishops over knights more than most humans would, except in extremely locked up positions. It does raise a valid question: why do we still teach beginners the old "knights and bishops are 3 points" dogma? Is that really a useful guide, or does that hurt more than help? Of course, advanced human players already evaluate the relative weight of material based on the potential action of each type of piece, to some extent, but the AI seems to take that to an extreme. It also seems possible that programming in "traditional" material weights to a brute-force heuristic engine might hurt it more than it helps, when playing against other powerful engines. Are the AIs seeing positions that are ignored by the normal engine because the normal engine does not traverse that tree, due to "material imbalance"?
I don’t know chess much but I like her style. It’s very clear and understandable. Not like she assumes everyone is a chess master and she moves 10 pieces in 5 second.
This was an excellent analysis, and I really appreciated your point on AlphaZero being a fresh perspective in an age-old game! Thank you for the splendid content!
Bobby Fischer had so much in common with Paul Morphy--both rose to the very top I'm meteoric fashion, then stopped playing competitively and died insane!
AO: Here, take this pawn. I don't really need it that much. GM Ben Finegold: "Great job... Give up a piece because you don't need those." (Incredulous sarcastic 'what the heck are you thinking' look) "Stupid machine." AO: Wins the game *GM Ben Finegold has left the chat*
Philidor: Pawns are the soul of chess.
AlphaZero: What are all these tiny pieces doing in my way?!
That made me laugh 🤣🤣
😂🤣😂
alpha zero makes worthy sacrifices, does necessary evils.
alpha zero is all seeing, alpha zero is utopia
When alpha zero started playing Starcraft, he literaly killed his units, to make room for another....everyone was little bit scared, imaging Alpha ledaing actual armies..
Normal chess: let me set up a fence with my pawns so I can develop my pieces
A0: let me sweep all this trash out of the way so I can get to my desk
Your comment made me actually laugh out loud. Thanks for that.
Ok. The MacReady defence, that worked..(The Thing,1982)
Hahaha, love it.
Awesome haha.
@@evolveofwow jajaja SAME
When I see alpha zero playing, I realize that I'm just a chimpanze that can talk.
yeah may be you are. But we humans aren't cause this alpha shit is made by us
Chimpanzee with an inferior short-term memory.
I don't see any proof that you can talk
True, but I like this feeling.
If make you feel better, alphazero is just a bunch of sand that can play chess.
Alphazero in reality: "I have no idea of what I'm doing."
LOL
xD
It's true. :D That's how neural networks work.
AlphaZero : Have i won ?
We are not there yet. But lmo
me: Haha, I'm a pawn up!
AlphaZero: Pawns are worthless, earthling! You have already lost, mate in 29
Jamesons...likes The Thing...cheating bitch
Yeah... Sometimes you just have to sacrifice material just to open up your opponent's kingside. EVEN QUEENS!!!!
AlphaZero: pawns are worthless, peasant. You merely adopted chess, I was born into it. Molded by it. How could it favor you.
that's my greatest fear haha
Lol
"And it was in this position that uh, Stockfish resigned the game"
-Agadmator
I read this in his accent lmaooo
Osas this your name burst my head .finish !! hahahaa happy to see my fellow benin man quoting Agadmator
@@Scriabin_fan same haha
And uh, this position is totally winning for white
Call your name
I am not an English speaking person.Your English is very easy to hear.Your commentary was very polite and easy to understand.
She is a retired Chess player, streamer, RUclipsr, can speak Hungarian, Spanish, English. She is one of the bestsoft spoken person I have seen in my life. 😊
Enjoy.
I concur. I'm just starting out in chess and most videos are way too fast. These videos are perfect and you never get flustered, lol. A real joy to watch and listen to.
e6, g7, h4, moves to A3 to sacrifice G6
easy to understand if your a chess player but if your a normal person this ain't even english! god damn elitist chess douchbags 😂
I agree with you. I also think her English is easier to understand than others.I'm also a non English speaker
It's called Hungarian accent. I always said that we speak much clearer English than the English.🤣
Stockfish analysis engine to stockfish : Why did you play that move ? Its a bad move.
Stockfish : You are the one who played it.
SF: "You're weak"
SF: "I'm you"
Stockfish : I am the best chess engine
AlphaZero : Hold my pawns...
AlphaZero: You can keep them
well A0 is not an engine. It's a Neural Network so SF is still the strongest engine
@@abdqs853 "In computer chess, a chess engine is a computer program that analyzes chess or chess variant positions, and generates a move or list of moves that it regards as strongest. "
Alpha Zero can do this, thus is a chess engine, among other things.
HAHAHAHA ! You killed me xD
@jyloup85 so it took Stockfish 12 attempts to get where Alpha got in zero?
AlphaZero: *King to B6*
StockFish: "Understandable, have a nice day"
AlphaZero loves activity!
*King B1*
It also likes to keep it's opponent pinned!
*King A1*
hahhahahahaha
Lmao
Needed space for the b2 a2-b1 c1 setup
Phillidor : Pawns are the soul of chess
AlphaZero : Let me just sacrifice all of my pawns
@schindler91589 Humans : (pikachu's meme face)
I guess you missed the significance of the White pawn on H7. Phillidor is still correct.
Alpha zero sacrifices the soul(pawns) for Absolute control and Complete Domination
@@Ceeewolf and you missed the joke...
@@Ceeewolf I was just joking ;)
> HAHA, you already blundered 25% of your pieces!
> I did? Well, time to win.
> What?
Jotaro: By now, I have finally learned to play the game.
@@luizquevedo6580 hahahah
Pieces: New chess master is in town. His name is AlphaZero.
Pawns: *sweating profusely*
:-))
Lol
King: developing projectile vomiting
Queen: developing explosive diareah.
It looks like AlphaZero's idea was to trap their higher level pieces even if it means sacrificing a pawn. It effectively took their pieces since they weren't able to move.
It’s like watching a SWAT team storming a medieval castle...
underrated comment bro! Nice to meet you
Damm that's exactly what i thought...
yeah, but hella slowly.
Perfect metaphore! 😆🤘💀🔥
If ur funding the enslavement and slaughter of animals, you are an animal abuser
Humans: I can make a queen with a pawn
AlphaZero: pawns get in the way of my queen
so typical of humans....trying to find water on Mars wasting a billion dollars in the process while being disrespectful of the abundant water sources on earth. Typical of humans.
@@vvenkat111 the reason to find water on mars Is to use it later on In conolies In mars, no to bring it to earth. And yes, earth is a paradise, but our chances of extintion are pretty high, if we colonize mars. those chances goes to 0
@@v44n7 0 is a little low i think
Pretty sure our chances of extinction is exactly 1 no matter what we do
@@thoughtsofapeer extinction looking a distinct possibility with covid panda (;-)) Our death is made in China too ;-)
AlphaZero trolling Stockfish, sacrificing 3 pawns and playing a prophylactic move.
Stockfish: Am I a joke to you?
He is a fish
yep, easy to do DeepTroll when you can see few moves more than your oponent
Your analisys is excellent!!
Stockfish: Give up the Knight.
ALphaZero: Rg1
Stockfish: ** Surprised Pikachu face **
I think this is the wrong game, because it violates all the foundations of chess, I think we must sue the creators of AZ0
AlphaZero: *my goals are beyond your understanding*
me: *my goals are beyond my understanding*
This was one of the most entertaining and interesting video I've watched recently ...Anna, you rock! Thank you.
Hot babe
Under every wholesome comment lurks a creepy one. One truly has to admire the tough skin women on the internet need to have.
6:22 "AlphaZero, goes for the attack with a very strong...
Kb1"
That was hilarious!
@@dominicsellers1813 and very aggressive one
Interesting is, my Stockfish 11 on a normal computer plays also 21. Kb1 (with the evaluation 0.00, so eval even that black ist 2 pawns up).
Damn she has a crush on Alphazero.
so do i
Who doesn't?
I know a bunch of people who talk about Alpha Zero like he's their favorite high school crush.
And the rest of the world on Anna.
@@michaelfortunato1860 You took the words out of my mind.:)
Always play Kb1.
Alphazero was secretly trained by watching all of Ben Finegold's lectures.
;)
Rawr!
Always retreat.
You forgot that one.
You with the right answer
Terry Leddra incorrect
Guess I'll chuck my 'chess theory' books written by mere finite mortals. LOL
humans cant play like this, thats why they dont. they can improve theory by watching these games but will never play this calculated
@@gergogaal568 I dont know about that. I think most players now and days rely too much on memorising defenceve stratagys and all the quick win moves amongst other things. Theres verry little originality in play style now and days.
NO. Pls don't. Very few humans can convert a position that a program sees as "winning".
Another way a0 wins over SF: We already know it likes mobility - and this is a huge advantage over SF-type engines. They don't do well with mobility because it increases their depth exponentially. a0 has basically learned to "stretch" SF and use its brute-force calculations against it :D
@@Duduvianna3 interesting. To be honest, when writing this, I was thinking about the kind of opponent used to train a0, which would have been anyway a traditional style (dare I say "brute force with strategy" calculator), of the same type as StockFish.
So, maybe not StockFish itself but an opponent that plays the same?
Eduardo Vianna what I think is really interesting is that in the learning process there are multiple Ai instances of a0 playing each other at the same time. And there is this phenomenon that in certain badges of AIs that play each other certain metagames form where one of the AIs learns to prefer a strategy that is strong against what the other AIs in their badge plays and strongly overvalues that strategy because it wins so much with it. So the fact that A0 tries to learn to play on its opponents weaknesses has to be and is controlled for in order to make A0 learn the games instead of learning to play its oppenents.
Eduardo Vianna Eduardo Vianna no, from my understanding it‘s not that easy. just trusting that the natural probabilistic distribution let‘s THE strongest strategy come out on top would create loops and hinder learning.
ruclips.net/video/eZCI7zu_DlM/видео.html
check out this video from the 25:50 minute mark. It’s about the deepmind AI for Starcraft, but I am pretty sure this problem occurs in every highly complex game. The problem is that every strategy depends on what your opponent does and there are many different possibilities to act. That means that there is the problem that some sort of rock paper scissors situation developing, where strategy 2 beats strategy 1 and most instances start playing strategy 2 until strategy 3 gets discovered that wins against strategy 2 but loses against strategy 1. To prevent that the instances get caught up in a loop with no real learning progress, that cycle is disrupted by introducing instances that play a certain way every time and don‘t evolve in order to make a0 remember that it has to be able to beat those strategies as well even though they are not as popular within the other a0 instances.
I agree with the second point. The learning happens within a* itself, although games at the top level are used to initialize the instances. They are basically used as some kind of starting point for the ais to develop from.
Eduardo Vianna mind to share the link?
Eduardo Vianna Eduardo Vianna the paper you linked does not go deep enough into the process of how deepmind implemented the self-learning in a0. Moreover it refers to the algorithms implemented by Ago0 unless stated otherwise.
hackernoon.com/the-3-tricks-that-made-alphago-zero-work-f3d47b6686ef
This blockpost about Ago0 however specifically mentions multi-task learning with hard parameter sharing, which is pretty much what I mentioned above. On the deepmind website, there are also papers that deal with exactly this topic so I am still not sure where you take your confidence from that this common problem in ai self learning suddenly was not relevant for a0?
And yes, they did not use external agents to overcome the problem I mentioned, but that is also not what I said. I said they used external games by top players to initialize their agents which is mentioned in the beginning of the video I linked.
Ex chess player, loved the analysis and how accessible you made it. Nice job. Subscribed :)
Peter Tench congratulations on your sobriety
You never stop being a chess player it's like riding a bike. You just take a break from it
Insane. It was literally watching black squirm around until it felt it was the right moment to attack. This goes against everything I have ever learned about tempo.
As an amateur player (rather poor at that) who is interested in AlpahZero and the recently publicized competition, I've got to say that Anna made a great selection of a game to analyze, and just brought a lot of insight and excitement to it. This should go down as a classic video. Way to go, and thanks, Anna!
I concur 100%
Amazing explanation showing how much can we learn if we free our minds from rules we take for granted🤗
it's like the matrix when morpheus talks to neo in combat training
@One Blue Boi Building a nuke isn't actually that complicated, getting the materials is the hard part.
It shows how much a computer can learn. Not we.
@Nathan Sindlinger No, it's not. Neither design for basic fission devices it particularly hard to understand. It *may* be going a bit far to say you could build it in your basement but with the right person I'm sure it's doable. The problem is that the best way to get enriched uranium or plutonium is as byproducts of nuclear fission and building a controlled nuclear reactor is a much more complicated process. I'm not aware of marketplaces where the general public can just buy lumps of barely subcritical weapons grade uranium and plutonium, but if they exist I will stand corrected.
Should be careful. One company is setting out to ruin backgammon with the AI info. They deduct blunders from the score and say winning the game is not enough?! 😂 That's going too far when the whole point is to win the game
UPDATE: Hey folks! Thank you for your kind words about my AlphaZero videos! I don't have new AZ videos on this channel but I have created something even bigger: an entire chess course (Attack like a machine!) highly influenced by AlphaZero's play, with the main chapter dedicated to AlphaZero's attacking chess. You can check out the preview video and find out more about the course here: www.chessable.com/anna-1
*Also: it's an "it", I know I was wrong to correct myself and apologise often. You can check out in the rest of the AlphaZero videos if I had learned from it: ruclips.net/video/rkt51PTCQG4/видео.html
HEY!! Thank You!! I find the A.I. games very fascinating! I just watched the "AlphaGo" movie, and while I do not know the game "Go" very well, I DO know chess quite well. The thing for me is that I realise the machine has several advantages over the human brain in terms of speed, and THOROUGHNESS of analysis, processing, and calculation. This ultimately places the Gen.A.I. in a whole different class. This is why I fail to understand WHY some people actually get upset and or offended by the powers of the machine when they are defeated "IT". BTW Why are YOU struggling with the irrelevant gender of "IT"??? It is an IT, gender has NO Meaning here!!! LOL!! Anyway, I enjoyed this video. Thank you again. Oh BTW, do you know how I might obtain a good chess program for my P.C.? Free or pay no matter.
Maybe we should stop using words like "sacrifice" as a descriptor. A move is a move is a move. Nothing more.
Do not apologize for calling AZ "he." However, it would have been way more badass if you called AZ a "she".
I am very impressed with Alpha Zero's performance at the chess board!!
Thank you for sharing.😄
@@CyanBlackflower Chessmaster9000? If u like older chess games..♛..Black queen is white though..😒))
Very good descriptions. A lot of chess videos I don't feel like it's improving my game but the way you pause for a moment and ask what should be done at this point then explain out the possibilities really helps.
You can beat AlphaZero if you pour beer over it's circuit boards.
not going to work ...it will just sacrifice it and win
Counts for humans too :P
"Cheating bitch..."
This works for humans as well. We are immune to non-alcoholic beer tough while he isn't.
just unplug the damn machine and you will win
Alpha Zero's entire strategy is to strangle his opponents by ruining any possible development.
Absolutely haha
I think its interesting that AlphaZero uses known openings that noone programmed it with.
Says a lot for book, but you also notice AlphaZero changes those openings pretty quickly too.
this is absolutely fascinating. Anna thank you for your commentary on the game. I'm not an
exceptionally good chess player so Its nice to have a professional player's interpretation of the
moves.
+@c b Let's just agree to dissagree on that one.
Human games are interesting exactly *because* of our flaws.
Alpha Zero is improving them but there is no proper explanation for the novelties
+@c b My point is that it makes sense because when a human wins we win *despite* our flaws achieving greatness.
How good a computer performs is completely and utterly irrelevant to how much chess players should be paid.
A human can inspire other humans as we can emulate them and achieve the same things as they can.
You can never become a machine however.
At least not with current technology.
Your comment don't make sense in my mind till human-computer interfaces reaches a point where we can freely transfer our minds too and from a computer and fully utilize a computer wirelessly with our mind.
And even then chess might be interesting with such a interface turned off and played with our regular flawed pieces of meat.
+@@solomonlubega8337 Several of the novelities makes sense to at least *some* humans looking at the games.
We're going to see more players emulate the best parts of AI play and traditional engines slowly adopt more and more of those into their code base till they reach parity.
@@Luredreier So you think the engines can fight back ... from this? Bring it on!
Starts like Morphy, grinds it out like Magnus. Stockfish calculates; AlphaZero learns from experience, like a human. Who said that romantic chess is dead?
Oh god so true. All through the game im thinking "just like Morphy". A0 is the pride and pride of chess.
It's 2021. Me, analyzing this game with the Lichess Stockfish: All moves that AlphaZero played are the first recommend moves of Stockfish.
You really did that??
Nice😂 guess somebody learned from his mistakes
Stockfish learns from played games, unlike Alphazero. After this game is in the database it'll remember the better moves
This is absolutely savage, alpha zero is toying with stockfish like a well trained dog owner offering it delicious pawns for it to play passive
yeah that's what i also find so interesting about these leela zero and alpha zero vs engine games.. this friggin ai's play these engines like a gm an amateur.. the engine is trapped in its own position with no way of getting pieces into the game.. they don't care about any chess principles and somehow gain advantage due to some seemingly weird and pointless move and then bam 35 moves later they are getting the upper hand because of this.. the positional game of these ai's is so incredible.. and then you realize this is fucking stockfish they play against.. an engine so strong it plays every human like a fiddle with ease.. an engine you will never win a game against within your lifetime if you play against the strongest version..
To win in chess you need to checkmate the enemy king. The fact that we value the other pieces is just an assumption we made. AI engines don't make those assumptions.
@@charimuvilla8693 Or has it assumed that the opposition will do or has been programmed to do exactly that and then plays that knowledge back against it?
@@Honeysucklebommie It should be possible to train an "anti-stochfish" AI but after seeing how positionally alpha plays I don't think that's the case.
All after being shown the rules for the first time just four hours ago.
Having watched so many video commentaries of the AlphaZero vs Stockfish series, yours is by far the most illuminating and easy to understand one. I'd love to see more of those from you, for sure. Keep that up.
Thank you Anna for this analysis, you explained the game and its nuances perfectly but, you know, many chess analysts can also do that, what you managed to do that was unique for me was to also convey the excitement of the game.
I found myself smiling along with your enthusiastic presentation and sharing your sense of admiration for AlphaZero's audacity. I liked the way that you anthropomorphize AlphaZero, it is clear that you see him as a bold, dashing hero and you made me see him that way too. Wonderful.
Probably my most favorite chess video I've ever seen. Just the enthusiasm alone.
22:10 I feel like the knight to d7 just hurts that bishop even more with its zero mobility, and Qc1 was typical AlphaZero protecting the c5 square, further imposing difficulties onto black.
Clearly AlphaZero is just pure evil with its idc about giving you even a hint of a chance.
Köszönöm Anna, hogy ezt a valóban meglepő játszmát ilyen szépen prezentálta. Öröm volt nézni és hallani, hogy egy ember mennyire másképpen gondolkodik mint egy gép; mert hiszen bennünk számos pszichológiai tényező dolgozik, amik a döntéseink szerves részét képezik. Az egyén pszichológiai alkata és a személyisége a szűrő amin kereszül egy helyzetet megítél - még a sakkban is. Ezt egy neurális hálózat, mint AlphaZero nem tudja, de Stockfish8-ban benne van az emberi gondolkodás lenyomata. Nagyon érdekes volt. Köszönöm! :)
She seems so sincerely excited about this, it's adorable!
I have NEVER so thoroughly enjoyed a chess vid. Thank you, Anna for your charming, entertaining and enthusiastic commentary
Hot babe.
21:45 great summary of Alpha's strategy. Excellent explanation and analysis, very enjoyable thank you!
Gosh, I love your assessment of this game.
It reminds of Bobby Fisher, he always did things that no one could understand.
In fact in a tournament one time, he moved his knight into a position that everyone(chess experts), in the chess community watching the game live, gasp, called it a huge blunder.
The chess experts explained how it has weakened his position, etc..
Then 9-12 move later the game was over Fischer destroyed his opponent with that knight move.
Fischer too was superhuman.
A game with Portisch has a such story.
Fischer 20-0 sequence with players from top 15 cannot be measured with Elo. His last three years were 100+30=8- with top players while playing reckless to win with a no draw mentality.
You described this game so charmingly. It almost felt like a bed time story.
literally watching this while lying in bed to sleep xD she's so soft spoken, it's really nice to listen to her.
I love her analysis, but I also love her voice. It's super soothing!
I love her sexy smile, her brain, her everything !!... ;-)
most awesome chess voice since yasser seirawan !! ^^
her voice is feminine, sweet, innocent and persuasive. it is very soothing, she could put a baby to sleep by whispering lullabies in their ears.
@@jackriver1999 well said !
Deffo, I'd like a nosh as well!
Great job with this, Anna. Just wonderful explanation and enthusiasm with the “patient attack.” Helps my son and I so much. Kudos.
I rarely see commentary on games, but I've definitely been quite impressed with AlphaZero so thought I'd check this out. Your presentation was enchanting and I have a feeling that you're on the money about why AlphaZero does so much better than Stockfish- it is unhampered by human theories on how to play the game to win or at least draw. I imagine it has come up with its own set of theories on how to go about winning/at least drawing, theories that have far surpassed the human theories that are essentially tying Stockfish down to a lower bar.
Rather than "theories," AlphaZero has *intuition* built up by trying things repeatedly, and recording what did and did not work. It does not attempt to distill this into a rule set you could explain or communicate - it just "knows" what to do. Here's a summary article (though it would be nice if they described the "policy network" in part 2): www.chess.com/article/view/how-does-alphazero-play-chess
@@James-nr9gm Good article, thanks. I believe I read the second part, but not sure I read that first one.
I agree !!!!
that was thrilling. a good sign of a good teacher is that they are enthusiastic about what they teach. it is really engaging and gets the student excited about the material also.
Oh the passion! A haven't seen ANYONE talk about chess like this! One can see in your eyes that you love the game and new discoveries it offers. This was really enlightening, even for someone who plays chess as a fun afternoon brain exercise. I caught your stream the other day and loved your way of teaching a friend to play chess. Keep up the good work and keep loving the game!
Chess Guru: Remember Pawns are the heart of the game...
Alpha zero: IF pawns are the heart then let the brain ( other pieces) win!!!!!!!!!
I'm a 1000 rated player but I've watched a lot of games from the best GM's but this brought tears to my eyes.
Megamind Blue how many hours have you wasted?
@@manit77 Prolly 100+
lts not waste if he learns something from it ignorant kid
@@ausar3852 I've been watching such videos forever but I never seem to learn.🙁
@@megamindblue1732 True. Videos are a form of passive learning, hence learning with a book and board along with many pauses in the between to check the position out for yourself is the best way to study. Do not waste your time watching videos.
Dear Anna - thank you for covering this game although AZ makes my head spin! I do love your commentary and please cover more of these.
She explains the moves so simply, beautifully and with such passion that even I as a 1500 player can understand and appreciate the beauty of this game. My new fav chess channel!
Yeah, my moves were pawn sacrifices. I'm just like AlphaZero! _>
In some years AI will play the Human Gambit!
Lol
No
a logical conclusion, since we do no good to anything. besides inventing sth that solves us.
Lolll
GENIUS!!
i always thought her voice was computer generated
Please do more AlphaZero analasys. Great video!
Oh my God! This is just mind blowing. Thank you so much for your detail explanation of this beautiful game. It is such an illustrative game for me that shows how we humans developed our own rules of chess based on our own weaknesses, mostly our limited ability of doing deep and precise calculations which keeps us away from playing what we consider risky movements like loosing a lot of material. As you said AlphaZero looks so confident entering in this terrain knowing that just improving and opening the position increase the odds of winning in the long term. It feels like we’re witnessing the God of mathematics playing chess for the firs time. Just amazing.
Just so you're aware, the newest chess engines, including Stockfish 12, would wipe the floor with AlphaZero, but that doesn't take away from the incredible achievement of AZ, and the amount of data that chess engine developers were able to take away from what AZ accomplished, and used it to strengthen their own chess engines. Of course, AZ was never a chess engine, and has not made any attempts to keep up in that department, it was simply a stepping stone for them in the grand scheme of machine learning.
@@joshs7160 thanks for your reply, I totally agree Stockfish and chess engines are in general extraordinary human achievements, and now Stockfish is also partly an AlphaZero achievement. We’re clearly walking into a completely new era not only for chess but for science overall.
Anna Rudolf is inspiring.
I'm not great at chess, but she explained the principles of AZ's play in a way that makes me want to try this for myself.
I don't even play chess (or any other games), but I was captivated for the first 30 minutes of this video.
AI will improve the world by sacrificing the humans.
Only half joking.
Hopefully it sacrifices you first....or maybe you could be the final one sacrificed, so you get a front row seat to the extermination of half your species.
Sacrifice the tall pawns
How would you know I am human?
@@albinorhino6 You're just a specist.
Cool channel and very good analysis.
This is an amazing coverage and walkthrough of this game. I hope to see more content like this from you in the future. You have a talent in explaining high level chess in a very engaging way.
I understand your excitement. All those years we have made several assumptions like "more pieces is a good thing" or "rook if better than bishop" and it's good to see an AI ignore those rules and play in a purely strategic way without assumptions made by humans.
I ain't forgetting how awkward the WCC closing ceremony was. "Now we invite you to take a picture of the winners"
Thanks for helping us enjoy this video.
One of the greatest chess videos I've ever watched. Thank you for this.
I loved your analysis of this fascinating game, Anna. So many "alien intelligence" ideas at work here - the novelty (ideas Stockfish doesn't see) and depth of strategic play is spellbinding. Although humans cannot play like Stockfish, I do wonder if humans *could* learn to play like AlphaZero if we could learn the same deep strategic principles and positional logic.
Great point, Mark! I think we can and should learn from the AlphaZero way of seeing chess!
A0 to me is like a human GM that played chess for thousands of years and has gotten to the 3600 level. Of course it just played against itself for 8 hours to get to this level lol
@@AnnaRudolfChess its more likely to learn to play like A0 than SF. A0 was modelled after human brains, after all.
I wonder if the human concept of a value system assigned to chess pieces is a result of capitalism ( everything on the planet is a commodity and has a value based on utility & scarcity ). Therefore humans have a flawed understanding of the game of chess by virtue of the economic system that we live under. I have an exciting idea... in Russia they start chess training around age 5 or something? For the next generation of chess players, remove the concept of a value system from the lesson and we can have some crazy chess games.
@@MrAgmoore All economic systems are value systems, thats what economic means. The value of pieces changes depending on the position. Alpha zero assigns values at a much higher level then the general principles we teach beginners. You talk like adding values to pieces is bad for the game, when thats literally how A0 learned. The values you're thinking of are just general basic principles to help out beginners. Economics has nothing to do with it wtf is wrong with you.
Try teaching a 5 yr old how to play the game without telling them that they should usually value a rook over a knight, because in most games a rook will be more powerful than a knight long term. But no you're right, because capitalism assigns value to things that fluctuate based on the situation of the world (just like chess pieces change in value depending on their relevance in a position) we've been playing chess wrong this whole time! A0 knows that a rook is more valuable than a knight or bishop long term, but it also knows when it can sack material for a different long term advantage. The basic principles of general piece value still applies to chess at the GM lv and the A0 level, but so do many other principles. And the higher up the rating ladder you go, the better understanding of all principles is grasped. Denying probably THE most obvious and basic principle of the game is just ludicrous. Jesus Christ man, how about you learn a little about the world
I followed this game with Stockfish 10+ on lichess after we got halfway into it, it actually suggests many of the same moves (not all, but most - including the masterful pawn d5 move that SF8 didn't have in its top 10 moves) that AlphaZero took. Its time for another A0 v SF matchup
Wow, I just wanted to check this out in like 5 minutes...ended up watching the whole thing in one go.
Thanks very much for this, Anna. One of the things I always enjoy in addition to your chess analysis and commentary, is that you speak perfect English. I speak a smattering of several languages, but am not quite fluent in them, and I am at a loss at how someone could master a language so well which is not their 'native' language.
Agreed. Peter Svidler is someone else wiho has a very impressive grasp of English
Talk to native speakers as often as you can. Use the language you want to be fluent in frequently. Watch movies in the language you want to learn and use subtitles in the language too.
All of these may feel like slow progress, but it is the fastest method that I know.
The scary part is even the NN developers don't quite understand how the algorithm works.
Hopefully we have heard the last from people saying chess has reached a draw death.
At last we humans have found a great teacher.
There was an old master who considered that pawns just obstructed his pieces.
Perhaps he was right all along.
The engines are toast. Long live AI.
I think you're wrong about that last sentence.
Given equal computational resources and equal knowlege of chess principles a traditional chess engine would come out on top because it makes more efficient use of the computational resources.
The strength of a NN on the other hand is that it can add new chess principles on its own without getting those hardcoded in by a human.
In other words, we're likely to see a back and forward shifting of the "balance of power" between NNs and AIs in chess as chess engine developers work hard to catch up with the NNs.
Each time they catch up in encoded chess principles they'll be ahead till the NNs discover some new principles the engines are unaware of.
@@Luredreier The engines are weaker than AI because they rely on human understanding of chess.
Alpha is beyond our understanding but we can learn better from it than brute force calculation engines.
@@proflaxis6968 That's not a correct understanding. One could easily rebuild brute force alpha-beta on a tabula rasa basis with modern computers, with far less input of human expertise (certainly nothing that required input from an expert player), and you'd probably make some gains, too, but it likely wouldn't prove decisive.
Even the "old school" programs like Stockfish are tuned using self-play and advanced algorithms from machine learning. But certain trade-offs between the domain of the static evaluation function and management of the search tree are built into the code base-especially around the issue of quiescence-and no amount of self-play can improve these choices, without the manual intervention of a competent chess programmer. This is why some people on the classical computer chess side of this regard the fundamental innovation from AlphaZero as the Monte Carlo tree search (MCTS) algorithm rather than the neural network component. AlphaZero was _forced_ to adopt a fully general search algorithm, because there's no other way to fully automate self-learning.
Now that MCTS has been proven viable at the very top of the heap, it's hard to say how much the old tradition might also be impacted. What you get from the old fashioned fast-and-dirty search trees is extremely good peripheral vision. You're rarely surprised by anything. (This was good in competitive chess, because it means you can't be sandbagged by clever preparation on the other side.) What you sacrifice is visual acuity on the primary lines. Your forte is to play solidly enough and never fall short of tactical excellence (capturing is almost always a solid move, if uninspired, so we saw a lot of that for twenty years).
Leela, who is less well trained right now that the AlphaZero, already displays much of the same brilliance, but her peripheral vision is poor (due to her reliance on a narrow and slow search that's not yet fully mature) so she gets beaten by weaker programs far more than you'd expect for a program that sometimes trounces Stockfish decisively. With deep recurrent networks, the brilliance comes first, and the wisdom of years arrives later (this is also true of offensive-minded rookie defensemen in the NHL).
Note how much emphasis you see in this video on AlphaZero's quiet moves. Leela hasn't got this trick yet-not to this degree-in the videos I've watched. But she sure shrugs off counterattacks (often at great material loss) to open up lines in an eerily non-human way. She's just raring to jump into the play-often to a fault. Compared to AlphaZero, she's the Karate Kid right now. Over time, it's expected that her inner zen will improve. She needs to augment her hyper-aggressive goal-crease directed style with just enough 360° peripheral vision to stay out of trouble (those blind behind-the-back drop passes that every gifted rookie short of Gretzky seems obligated to try out at least once per game).
The legacy trap you cited is almost always dominated by cost-benefit effects. You _could_ go back and start over, tabula rasa, but it will be a long slog and dry spell before your new program can defeat your old program. The incentives for making this choice are rarely high. Especially when the programmers involved don't fundamentally believe that the legacy of human expertise is a particularly large problem; if tabula rasa Stockfish appeared likely to gain hundreds of rating points from the "purity" budget alone-fumigate all taint from those confounding narrow-minded human GMs-work along those lines would probably commence immediately.
One of the reasons that Stockfish is so much more powerful than similar programs is that Stockfish employs a very carefully tuned and narrower search: more acuity, less peripheral vision. We observed that in this game, where Stockfish completely fails to see the central pawn push. Narrow search is potent, but risky.
Of course, it's a huge win to have no human legacy at all (from either chess experts or chess programming experts). That's what Google sees in AlphaZero and the real reason AlphaZero exists in the first place.
To reiterate: Stockfish could easily be rebuilt tabula rasa with no human chess expertise at all, and with modern computers you'd expect this to work out better, not worse (microwave ovens have more computing power these days than chess computers from the 1970s). But the re-investment of chess _programming_ expertise would be non-trivial, and merely erasing the negative legacy of human chess-expert tunnel-vision would not pay enough dividends to support a ground-zero rework on this scale. At least, not unless adopting MCTS at the same time changes everything on that side of the chess world, as well (without any contribution from deep learning or a neural network).
Keep your eye on Komodo MCTS. Hybrid vigor could yet put traditional engines on a whole new track, all by itself.
@@afterthesmash Who wants to stick to black and white TV when they have seen colour?
It seems obvious to me that Alpha sees at least three times further than the Fish.
We humans, like Alpha, can see the long term benefits of early strategic decisions.
Yes The Fish has excellent peripheral vision and can't be easily surprised but is fatally handicapped by its event horizon as Alpha is demonstrating.
Its a shame Alpha can't explain its plans to us or how it comes up with them.
But it is providing the best learning material ever produced so far.
Hopefully our GMs can take it on board and get over the 3000 threshold.
You realize that traditional engines use AI right? Different algorithms and evaluation tools is all.
Insane. It was literally watching black squirm around until it felt it was the right moment to attack. This goes against everything I have ever learned about tempo.
This is late... but its playing chess like it plays Go. its moves are designed to increase its potential to win if it plays perfectly, if it can increase its odds of winning by giving up pieces, it will since greater odds of winning is better then pieces. as people can see, it reduces the moves available to its opponent in any way possible, increasing its odds of winning, eventually as it plays out this results in the enemy having no place to move, or essentially 'checking' as many of their pieces as possible, which is sort of the point of the game. After seeing it play just this one game, its quite clear that its the correct/best way to play, but also the most difficult as its literally playing on a tightrope, if it were to screw up and allow pieces to escape, it could ruin its pin and result in a loss. its like playing with absolute confidence that you wont lose with every move you play, monitoring more pieces then seems to be humanly possible. It would be interesting to see alpha zero play halfway through the game if a human screwed up its game by moving a piece badly, what would alpha zero do? would it be able to recover?
@@jessiejanson1528 It's just incredible that moves like Kb1, throwing away all the extra time/tempo it had gained, where actually the most optimal moves it could've made. It's like it's some kind of sadist that just likes watching the opponent miserably fail to activate their pieces.
you can not compare an A.I. to a human player. humans are bound to make mistakes along the countless calculations of possibilities opening up. so it is advisable to use advantages and exchange pieces since you will mess your advantage up by overlooking some detail at some point.
you can play like alphago zero if your game is without mistakes, flawless. but it isn't since you're human after all. try to play that way and get shattered.
alpha zero: "these pawns are restricting my inner beast!"
🤣😂🤣🤣🤣🤣🤣🤣🤣🤣
Why was this so funny
Lmaooo
Most underrated chess analyzer Anna Rudolf
Please cover more games of AlphaZero
Yes, it was sublime.
amazing restraint
thanks for the explaintionment
@@CandidDate Explanation is the right word, not explaintionment (that word doesn't exist)!
@@theUroshman haha
Was looking for a voice to voice Leela Chess Zero, ala Sierra Entertainment's 1996 PowerChess. Guess I found it.
I thought this was very well done. Would like to see more analysis like this Anna.
Anna more comments, please?
Greetings from Ecuador. Loved the explanation and even though I am new to chess, the way you presented the game made it understandable. You gained a new subscription. Keep up the great work.
Great video, Anna. I loved all your comments, they really highlighted and brought to life the ideas and approach AZ had in this wonderful game. (Side note: Looking good, Anna!).
I'm analyzing this match with stockfish 10, and it immediately found that d5 move. Plus, SF10 sees that AZ is better and disagrees with some of SF8 moves, which is logical. I really want to see a match with that new version of SF
Yeah, it didn't even take one second to find that move, I am so confused why they chose stockfish 8...
@@Rainnnny Stockfish 10 didn't exist when they were doing this match. Stockfish 9 didn't even exist. They did this near the beginning of 2018. It's a peer reviewed paper so it took awhile to be released.
Yes, thank you Anna for using Stockfish 10! We appreciate that you are showing the best our latest official engine has to offer.
Agreed, AMOF my SF showd more fitment with AZ's moves than these SF's moves. For a moment I started thinking the sides are switched!!!
www.chess.com/news/view/updated-alphazero-crushes-stockfish-in-new-1-000-game-match In this link they say it was Stockfish 9, can someone confirm?
It's like chess is so advanced that we've returned to the Romantic Era.
Stockfish: Life should be planned carefully to be most fulfilling
AlphaZero: IF YOU HAVEN'T SUFFERED OR FELT THE RAIN POUR AS YOU LAUGH, HAVE YOU TRULY LIVED?
Awesome commentary! We can hear the passion. I've bookmarked this video.
I love this video because of Anna's skillful understanding and in-depth explanation; also Alpha-0 astounded. Chess will never be the same. Bishops on side near the back rank with the Queen close at hand to help and the Rooks on the opposite side to the Bishops is what I try to emulate. And sacrifice Pawns to open files. Truly a concert of Chess.
something wonderful about chess is that despite knowing only the rules, it is still captivating to understand the thought process and understand it once explained. Harder to do in Go.
Welcome to Anna-lyzing with Anna
This has been the most charming presentation of chess I have ever personally experienced. Anna Rudolph makes chess fun...and fun to watch and analyze! Consider me a fan. Look me up when you find yourself in LA.
Prophylactic move. That's hilarious.
Jeffrey Gold She’s not using the word in its sexual sense, but its meaning as “a precautionary or preventive measure”! PS check out KWVE.com- excellent Christian teaching!
en.wikipedia.org/wiki/Prophylaxis_(chess)
22:20
Queen also defends pawn on C5 from there !!
Satisfactorily & interestingly instructive a tactical & gambit presentation. Please feel free to post more of those AlphaZero chess games' beneficial video demonstrations.
That was fun to watch. Your commentary was mesmerizing. And you never asked for compensation. I'm so grateful for this lesson. Thank you.
Alpha zero is basically Stalin
"These pawns are blocking my other pieces, I'll just sacrifice them to get them out of the way"
Absolute madlad
Isn't chess already a game about being Stalin?
@@milanstevic8424 I think Bobby Fischer said something to the tune that yes it's basically a game for egoists that want to prove something.
Unfortunately, Fischer didn't live long to see this...
oh, man, you just remembered me that fisher died, I'm just now discovering some of his games.
@@mtomazza Oof, he died 2008
I wonder how well this program would play Fischer Random?
billfinger most likely similar strength of trained in it
The trouble with computers is that they don't know when to resign-Bobby Fischer
After watching a bunch of these AI games, it is easy to find yourself starting to think like that: who cares about material! The most important thing is making sure your pieces have a lot of potential threats, and that your opponent's pieces are stuck in the corner. A bunch of opponent pieces huddled in the corner unable to move seems to be the AI engine trademark. We can all learn from this strategy :)
Can't know for sure what goes on in the AI's mind, but I think its game is more complex than just that. We humans, with our limited brain power, are only able to think of simple strategies - like "sacrifice material for position". I think AZ's model, even if encoded into its neural network and not being available as an actual mathematical model, is sort of numeric - its neural network has found a way to evaluate everything - material, position of its own pieces and of those of the opponent, and the relations between them on the board - as a single function. Its play is to just maximize that function.
@@a0flj0 Sure, but what we're talking about here is really the relative weights given to different types of advantage. AlphaZero and Leela seem to give more weight to closing off opponent pieces (especially bishops) vs. material than most humans, and thus also human-designed brute force heuristics.
Of course when I say "most humans" there are exceptions...I've often heard AI strategy compared to Tal, and it seems fitting, but Tal would often make such "wild" moves without being able to consider the possible outcomes should his opponent act positionally. In effect, his games seem to exploit the human bias towards material over preventing "really bad" looking positions. But the AIs are trained against themselves, so it makes sense that the AI makes these moves in a slower, more long term manner.
Similarly, the AI engines seem to value bishops over knights more than most humans would, except in extremely locked up positions. It does raise a valid question: why do we still teach beginners the old "knights and bishops are 3 points" dogma? Is that really a useful guide, or does that hurt more than help? Of course, advanced human players already evaluate the relative weight of material based on the potential action of each type of piece, to some extent, but the AI seems to take that to an extreme.
It also seems possible that programming in "traditional" material weights to a brute-force heuristic engine might hurt it more than it helps, when playing against other powerful engines. Are the AIs seeing positions that are ignored by the normal engine because the normal engine does not traverse that tree, due to "material imbalance"?
I don’t know chess much but I like her style. It’s very clear and understandable. Not like she assumes everyone is a chess master and she moves 10 pieces in 5 second.
This was an excellent analysis, and I really appreciated your point on AlphaZero being a fresh perspective in an age-old game! Thank you for the splendid content!
Stockfish can do cool Tricks, but what alphaZero does is pure Magic
Paul morphy would love these games, the style of alphazero.
Nah morphy only sacrificed because he gives them a free root so he's forced to save up for a big attack
Bobby Fischer had so much in common with Paul Morphy--both rose to the very top I'm meteoric fashion, then stopped playing competitively and died insane!
A0 Is like the Paul Morphy of chess machines!
In fact, Alphazero is said to play like Karpov, mobility of all major pieces as Anna mentioned
I THINK ACTUALLY ONLY ALPHA ZERO PLAYS LIKE ALPHAZERO@@NeutralDice
AO: Here, take this pawn. I don't really need it that much.
GM Ben Finegold: "Great job... Give up a piece because you don't need those." (Incredulous sarcastic 'what the heck are you thinking' look) "Stupid machine."
AO: Wins the game
*GM Ben Finegold has left the chat*
Tried stockfish 11 for this and I can see it has learned a lot from alpha zero. Now it recommends the moves no one saw coming.
Thanks Anna. of course we'd like you to cover more games - so informative, educational and well presented.