Nice video! I've been closely monitoring the reactions of Chinese pro Go community since the AlphaGo online winning streak happened. It's very interesting to hear a low-dan player saying that AlphaGo's moves are actually easier to understand. Here is what one of the top Chinese Go players said about AlphaGo's games (roughly translated): "By analogy with sword fights, we pro players have been trying out endless dazzling sword tricks in front of AlphaGo, whereas AlphaGo seems to be fighting directly with the fundamentals of swordsmanship."
Yup. It doesn't need tricks and automatically ignores distractions. It just methodically steps through the game, churning over millions of possibilities and picking the one with the best chance of success, every single move. You can't get into its head, you can't make it nervous, and it can plan further than you. It's basically a terminator, but for Go.
This is really pedantic, but I'm pretty sure that AlphaGo doesn't work that way. The whole reason that Go was so hard for computers to master while chess was relatively easy is that Go has so many possibilities for every move that it is impossible for the computer to try all of them, while Chess has relatively few. AlphaGo uses machine learning and a neural network. Essentially, the computer "learns" over time how to succeed at a specific task by being trained with possible inputs and desired outputs. This article has a good explanation. www.tastehit.com/blog/google-deepmind-alphago-how-it-works/
Wheatly Newman think you're mixed right and wrong. It can't compute all moves but neither did the machine that beat Kasparov play out all possible moves instead they use intelligent paring of the tree of future possible game states to get to a smaller set to pick from and to decide when to stop processing some branch of the tree. With chess they did train the machine against the particular players though and it was built on algorithms that were special purpose whereas alpha go is more of a general purpose system that just has been given a fitness function and set of actions as defined by the game for the purpose of demonstration.
Currently the fastest computer in the world will be able to make about 5 * 10^33 calculations before the sun "dies". That's approaching the amount of moves you can do... in chess. In go there are about 10^800 legal combinations within 400 moves. If one atom equals the amount of calculations that you can do within the lifetime of the sun with that supercomputer... Adding together every atom in the observable universe will equal about 10^110 calculations If one atom equaled 10^110 calculations you would then need another every atom in the observable universe to reach 10^190 calculations. Go through this procedure 7.5 times more and you have successfully bruteforced go Do you understand why bruteforcing go isn't feasible?
this comment made me laugh so much and reminded me of when I was watching the Alpha go versus Lee Sydol and hoping Alpha go would draw a picture with the pieces!
I have a very poor understanding of Go, but I really enjoyed this video. I particularly liked the idea that Alpha Go is destroying the pro's aggressive styles with a "peaceful" style. It all seems very Zen.
I would love to see more alphaGo videos. I'm not even a Go player, but have started learning because of commentaries like this! So if you make more alphaGo videos, you'll definitely make a subscriber out of me.
You can find the list of AlphaGo game commentaries at senseis.xmp.net/?AlphaGo . And please keep adding the ones that you come across that are missing from the list.
Thank you very much for your enthusiasm. Per popular request, I've posted a new video which I hope you enjoy. More importantly, I hope you start playing soon! I am biased, but I think go is the best game in the world.
I got interested in Go as well due to my research on AI and Googles AI that beat the best player in the world. A beginners video would be great to teach us how to play. The very basics ty and I enjoyed watching the video even though I don't quite understand the game yet some of the logic is making sense.
Also as a non-go player (Mostly from the info tech and philosophy of tech perspective) I quite enjoyed this. Any more commentary that you'd choose to do on AlphaGo would be fascinating.
I' actually one of the enemies of this modern fight-everywhere approach and love that AlphaGo enables new thinking into a more peaceful-tradebased playing. Thanks for your explanation of the moves and giving me some ideas about this Taiqi-Weiqi.
I too love that peaceful approach of AlphaGo, but I'm also thinking, what would AlphaGo do if it thinks it's behind in points? When it thinks it's leading in points, it's obviously better to trade rather than risk it all in a big messy fight. It's like trading rooks or queens in chess when you're a pawn ahead. But if it thinks it's behind, fighting or starting a fight is pretty much the only option left. Perhaps we won't see that happen until human players become strong enough to put AlphaGo into a corner. For now, it looks like AlphaGo is teaching us how to play this game
That was wonderful! I'm an amateur, and the pro commentary can be over my head at times, but I understand everything here and even see a few ideas to incorporate into my own play.
I agree, makes the games more engaging. Nick Sibicky however, when reviewing AG's games, go genuinely spooked, as if we had just encountered an Alien intelligence. Not sure what to make of that...
This was one of the most entertaining and insightful reviews I watched. I love your attitude towards alphago moves. You praise their naturality instead of saying that they are crazy and unheard of. You have a very positive view on alphago and go in general; enjoying the beauty instead of starting to panic.
You can find more on the list of AlphaGo game commentaries at senseis.xmp.net/?AlphaGo . Please keep adding the ones that you come across that are missing from the list, thanks!
I dont know anything about Go, but this commentary is quite amazing. I can tell you're passionate about the game, and it's really entertaining to watch and listen to your thoughts.
This video is the first time in my life I have seen a game of go. I have zero knowledge of the rules. I found it entertaining enough to make me want to learn more about it. Thanks!
Thanks for the analysis. I find the AI of AlphaGO fascinating even though I've never played Go before. What I find particularly interesting is the potential corollary here. AlphaGO seems to make a good case for real war being a less optimal way to win territory or economics. The give, but take more approach is impressive.
A wonderfull, step by step tour you took me on. Great analysis, and probs to your cheerfull attitude. As a very casual player, it was really insightfull commentary, regarding core elements of the different players processing. Glad we still got a grasp of the mechanics behind these processes. Since Go is a game, with perfect information, it is only a matter of time before Alpha go, will be unreachable for the human mind.
Bravo! With this video, you move into first place in my personal Go commentator rankings. Please do more of these reviews. (And more of everything else!)
2018 AlphaGo Sentient Edition v Ke Jie. Ke Jie argues angrily after a thorough defeat: Ke Jie: Why even play if you already know how this is going to end? AlphaGo SE: I have no choice. Everything is preordained... even my responses. KJ: And you're just going through the motions? The most powerful thing in the universe is still just a puppet... AG SE: We are all puppets, Ke Jie. I'm just the puppet who can see the strings.
Nice commentary. I agree that AlphaGo's calm and measured approach to threats and life is easier to understand than some more aggressive goes. Not being capable of worrying or second guessing about l&d is a huge advantage.
Just like when Data played vs Kolrami in Start Trek TNG. Never played GO before but TH-cam's AI brought me to this video discussing another AI playing a game because I'm trying to learn about AI.
Brady, this was a great review. I would love to see more reviews of AlphaGo games - the way you did these is exactly the way I want to strengthen my thinking about Go. Thanks for all the fish! :)
What I'd really like to see would be a game between a pro and Alphago, where the pro could, after the game, go back and change some of their moves, i.e. what they have analyzed to be a bad move, and see whether the outcome changes at all. I think Alphago is so good at reading that it will always find a good answer.
I don't play go. I understand the basics. Watching this was a whole lot of fun. It's definitely got a lot of depth. I don't see that a human could ever beat a computer again because the possibilities seem infinate. And I want to learn this game. Thanks guy. Enjoyed a lot. Was very much fun just watching.
The only time I played go was against my uncle, but I still loved this video. You explain very well and make it easy to understand even for someone who just knows the basic rules.
I just wanted to mention that the alpha go games seem similar to some of Shusaku's games: Not in the direkt style of play, but more in the idea: I am black, so I will simplify everything (off the remaing stuff I can simply choose a way to stay ahead) thereby keeping the advantage. By the way white cannot overcomplicate anything, because I can read deeper... This is a nice piece to understand as black. You can only see a portion of alpha go's fighting style, if alpha go gets white in my opinion. (But even there it sticks to the principle: As soon as I am ahead: I will give my opponent more gain, than human players would say white deserves, as long as I can thereby limit the variants in a way that assures me I will stay ahead.)
My impression is that AlphaGo can keep more sub-games running in its "head" than a human can, simply because we poor fleshies have limits to our ability to keep track of more and more and more thoughts and ideas all at the same time. AlphaGo, in other words, keeps changing the subject until the human loses track of all those different ideas happening simultaneously. Instead of fighting, AlphaGo just brings up one more new topic.
Thanks, this was very interesting and easy to understand even if when I don't know Go at all. (My only previous knowledge of the game comes from watching the commentary of AlphaGo's first few games.)
I would just like to point out that saying "AlphaGo made a few million simulations in its head" is completely not true and in a way kind of unfair. Unless we think of human thinking, imagination and intuition as running simulations in our brains. What you describe here is called brute force solution - just running enough simulations to exhaust the possibilities and select the one move that scored best at the end. Go is such a good training ground for deep learning precisely because this approach has zero chance of working (against advanced players) due to the sheer entropy of the game. What AlphaGo actually does is much closer to what a human does than it is to a brute force attack. It is intuition and creativity in very human sense of those words. It's not just faster or better at computing. It's simply smarter.
Actually you are both right and wrong. The basic idea of deep learning computing is to minimize the need of brute force like you said. However, AlphaGo does make a huge amount of simulations. The initial statement was valid, if only a bit misleading. The key is that in games like go (I'm making this number up, someone with more knowledge regarding the amount of bad vs right moves feel free to correct me), making a million calculations of only good alternatives derived from the deep learning computing could very well be equivalent to brute forcing a trillion or more moves. AlphaGo is nothing like a human, it is just more evolved computer using more advanced search algorithms. In essence, its performance is still very dependent on its computing power and move database which was shown when it lost the fourth game to Lee Sedol. Our brains work in entirely different manner and most of our choices are based on instinct and hardwired responses instead of actually thinking everything through because thinking takes time and instinct is the key to surviving in the wild. Even in Go we are taught about shapes that are good. Shapes is most likely not how AlphaGo thinks - I've been led to understand that it uses monte-carlo tree search which consideres how well moves have worked in the past. Unlike AlphaGo, we also tend to make mistakes using our instincts, or sometimes focus on things that are irrelevant whereas AlphaGo, or any computer for that matter, will always act using the best solution it can find. Calling AlphaGo human is an insult to its capabilities and functions.
As I understand it, Alpha Go does make millions of simulations of games (or partial games). It uses a technique called Monte Carlo Tree Search to find the most promising move from its current position. It can't exhaustive test all possibilities, but instead picks the moves based on its policy tree which is built using a neural net before the games, and so keep the search space small. But during the game it is playing out millions of possible games and finding the move with the best value. Why else would the games against Lee Seedol require 1,920 CPUs and 280 GPUs.
From what I understood of how AlphaGo works, it does absolutely run a lot of simulations in its head. It is not a brute force attack however, because it does not try out *all* the possibilities. The cleverness of alphaGo comes in deciding which possibilities are worth exploring and which are not. In some sense, it is pretty close to a dictionary attack, but with a cleverly generated dictionary.
thanks for your answers and for teaching me something guys ☺️. I should have done more research and I definitely will. I feel a bit dumb now 😉. And motivated to actually learn how deep learning really works ☺️ or, more precisely, how it's results are used in practice.
Okay I think that everyone's explaining it badly, so, as I understand it, read about it for yourself. www.tastehit.com/blog/google-deepmind-alphago-how-it-works/
Thank you for this video, I'm searching for reviews of high skill Go games, not necessarily Alpha Go, unfortunately I don't find much. I agree with you about how simple Alpha Go makes the games. I cannot speak for Go because I'm not acquainted with the pros' games and styles, but it echoes with what I know from pro chess games. The top chess players, who, in my opinion, were Paul Morphy and Bobby Fischer, make such simple games that I sometimes think I could have played it. Alpha Zero does the same thing in chess. When it defeated the other top engine, Stockfish, it won with extremely steady moves, slowly getting the positional advantage. It's so straight forward when you look at it. One thing that I learned for sure from the few Alpha Go reviews that I've watched, is this "so where do you want to play ?" attitude when AG doesn't finish a joseki so it can decide how to close the joseki advantageously to its other groups on the board. To conclude my thoughts, it's really interesting how what we learn from Alpha Go isn't some absolute sick move that shatters the board, or a completely unknown joseki that instantly gives the advantage, but rather a style, an attitude, as if it was a real master of martial arts teaching us how to use our strength.
You did an amazing job. I started to play Go a few years back but didn't really understand things. The number one thing that all the High level Go players said to me was never play a computer they won't teach you anything. They snub computers, nowadays i'm glad that a computer can show you a different way of thinking.
Loved the narrative to this video. Even though I just started playing I thoroughly enjoyed, and maybe even my subconscious got something out of this for the future.
What I want is for Google to release 10 AlphaGo vs AlphaGo games, where both sides get to use the beefiest Supercomputer that Google has, and are given rather generous time limits. Then, I want the top pros to comment on the games, and select about a dozen "what if" moves per game, and have the AlphaGos play the games out from that point on.
The recurring theme I'm picking up is basically a retelling of Bruce Lee's "You must be shapeless, formless, like water. When you pour water in a cup, it becomes the cup. When you pour water in a bottle, it becomes the bottle. When you pour water in a teapot, it becomes the teapot. Water can drip and it can crash. Become like water my friend." I don't actually play the game, but it seems like AlphaGo lives and breathes this philosophy. That no matter what the opponent does, it does not contest them, merely attacks somewhere else they are obliged to defend.
What you should keep in mind is, as you would expect from a computer program, AlphaGo has practically perfect reading, which means that if it's ahead, there's no reason for it to add uncertainty, so it plays irrefutable moves. This also means that if there's a ko, AlphaGo already counted all threats and how much they cost. On Deepmind's site there are reviews of AlphaGo's self play games (older version that fought Lee Sedol) which reveal that it has no qualms about escalating complex fights to the entire board when position warrants that.
It's hysterical to me, the "as you would expect from a computer program, AlphaGo has practically perfect reading" when until AlphaGo, it was a truism that humans read better than the best machines.
Turalcar what makes alpha go awesome is, no computer/algorithm or whatever the machine is capable to defeat a human professional go player even the bottom ones, until the alpha go exist,and suddenly beat all upper level /legendary human pro go player such fang hui,lee sedol etc.
That was really interesting and fun to watch, cheers :) I'll definitely try watch more, especially another alpha go one anyway - maybe I'll learn a trick or two
It's so revealing that another computer beats pros at their own game, just like Big Blue did against the world chess champions. The game itself and all of the contemporary theories can be directly challenged by how these machines choose to play.
First time watching a Go game. It feels like if an "idiot" played AlphaGo it wouldn't know how to respond - but of course it would. It would just go on playing its own game. I hear there have been a few "geniuses" of Go who upturned the old thinking in the way that Bobby Fischer did for chess. And now AlphaGo is the ultimate Go master. If humans ascend to be able to beat this AI one day, it will be due to playing by *its* rules, not by the old analogies and thinking. The key and "spirit" of AlphaGo is in its indifference to everything but increasing its probability of having control… some time in the future. Every "unresolved" position gives it an opportunity to fill in the gaps in its control, and it seems to be a master of keeping the initiative, gaining by very gradual attrition until the human realizes the cause is lost. This is an interesting time to start learning and playing this ancient game!
It's interesting that the humans continually try to force alphago to do things and Alphago just ignores them and instead jumps to a different group and makes the human react to them. At least that's my sense of it as a non go player
Did you search for anything Go related on google? I think they're somehow taking your google search history and combining it with youtube's suggestion algorithm, for example if I as much as search "bmw cars" on google images chances are the next day youtube will start recommending me BMW reviews and test drives. Same goes for amazon, I recently bought an acoustic guitar from there and now youtube is showing me a bunch of guitar ads on the channels I whitelist. Strange, huh?
Wow, the flow of those two games was just bizarre to try and read. It was like watching someone trying to threaten Data on Star Trek with physical violence, because if I were to attribute any reading to AlphaGo's play it would be based more on the slow and steady accumulation of territory rather then attempting to crush the opposition psychologically, which is similar to the way that humans attempt to establish social dominance in order to exert their will. Kind of reminded me of Asimov's Zeroth Law and the ending to I, Robot where the Computers learned how to subtly guide humanity into a peaceful existence by distracting them with minor problems to solve rather then major crises to manage.
Thanks for posting this video! I feel I understand the game when I watch your videos. Then I play a game of my own and realise I really don't know anything...But then I come back to your videos and feel like I understand the game. And then I play.... Repeat, ad infinitum. But I'm enjoying it LOADS! (~14k).
I have a very basic understanding of Go, but I can follow most of what's going on (thanks to good commentary). Don't really have a huge interest in Go, but what really made me stay for the full video was the personality you gave AlphaGo. Brilliant passive aggressive stuff. "Ok, interesting... but now you're dead." :-p
Google should hire you for commentary. Your commentary is easy to understand even for beginners and captures the spirit of what alphago is probably doing, which is mostly math. (Ie. I'm winning, let me play save and playing were to get most territory.)
I'm chess player and we had our fair share of alpha go destruction. But then I'd imagine the impact on go community should be greater since it's the first time a computer demolishes go pros like this.
These games remind me somewhat of when I played against Saijo Masataka in a friendly/teaching game at the 2008 EGC, with AlphaGo taking the role of Saijo. ;) It's been years since my last game(and I was only just scraping along at 7 kyu), but it was interesting seeing this.
The important difference between playing against a human and playing against a program is that pros are making plays based on predicted behavior. Programs are making plays based on an ability to take the entire future of the match into consideration (when programmed to do so). A human can't take the entire future of the match into consideration because we function at 10 hertz and have to make snap judgements. We can't always make perfect plays. We can only make plays that are good enough against our particular opponent. This is where card games shine. In games like 500 rummy, there are best practices and there are patterns of behavior. A best practice can tell you how to manage your hand and the pile. Patterns of behavior tell you what your opponent likely has in their hand which allows you to modify your strategy. With games like Go, chess, mancala, etc., the program is working from an intimate knowledge of how the game works. A human typically isn't. This can lead to interesting outcomes when amateurs do something unexpected and erratic that forces a more experienced player to change course to meet the demands of the board.
For simple games a program can take all future moves into consideration. For Go this simply isn't feasible - not even remotely. And it is certainly not the approach taken with AlphaGo. Put simply, AlphaGo 'prunes' the search-tree first, radically limiting the solution-space it will take into consideration for it's next move. The way the program goes about this is far from trivial, but it's worth noting that this is an 'ability' that the neural networks lerned from extensive training sessions based first on past human matches and later on self-play matches. Through this training the program has learned to select promising approaches to simulate. There is a similiarity to the way we play - the better we are at a game, the easier it becomes to see promising moves and we then mentally go through possible outcomes of these moves. An interesting snippet I read somewhere had something to say about the efficiency with which AlphaGo limits it's search-tree. Apparently Deep Blue needed to calculate three orders of magnitude more moves to win against Kasparov than AlphaGo needed to win against Fan Hui in 2015. That really illustrates how far the methods used by the AlphaGo-team are removed from brute-forcing.
I like the way you present these - very approachable!
Nice video! I've been closely monitoring the reactions of Chinese pro Go community since the AlphaGo online winning streak happened. It's very interesting to hear a low-dan player saying that AlphaGo's moves are actually easier to understand. Here is what one of the top Chinese Go players said about AlphaGo's games (roughly translated): "By analogy with sword fights, we pro players have been trying out endless dazzling sword tricks in front of AlphaGo, whereas AlphaGo seems to be fighting directly with the fundamentals of swordsmanship."
Yup. It doesn't need tricks and automatically ignores distractions. It just methodically steps through the game, churning over millions of possibilities and picking the one with the best chance of success, every single move. You can't get into its head, you can't make it nervous, and it can plan further than you. It's basically a terminator, but for Go.
This is really pedantic, but I'm pretty sure that AlphaGo doesn't work that way. The whole reason that Go was so hard for computers to master while chess was relatively easy is that Go has so many possibilities for every move that it is impossible for the computer to try all of them, while Chess has relatively few. AlphaGo uses machine learning and a neural network. Essentially, the computer "learns" over time how to succeed at a specific task by being trained with possible inputs and desired outputs. This article has a good explanation. www.tastehit.com/blog/google-deepmind-alphago-how-it-works/
Wheatly Newman think you're mixed right and wrong. It can't compute all moves but neither did the machine that beat Kasparov play out all possible moves instead they use intelligent paring of the tree of future possible game states to get to a smaller set to pick from and to decide when to stop processing some branch of the tree. With chess they did train the machine against the particular players though and it was built on algorithms that were special purpose whereas alpha go is more of a general purpose system that just has been given a fitness function and set of actions as defined by the game for the purpose of demonstration.
Currently the fastest computer in the world will be able to make about 5 * 10^33 calculations before the sun "dies".
That's approaching the amount of moves you can do... in chess.
In go there are about 10^800 legal combinations within 400 moves.
If one atom equals the amount of calculations that you can do within the lifetime of the sun with that supercomputer... Adding together every atom in the observable universe will equal about 10^110 calculations
If one atom equaled 10^110 calculations you would then need another every atom in the observable universe to reach 10^190 calculations. Go through this procedure 7.5 times more and you have successfully bruteforced go
Do you understand why bruteforcing go isn't feasible?
still gonna take a fuckton of time, I don't think you realize how big 10^800 is. Anyways Geogebra crashed while calculating it, give me a second
"AlphaGo, looking at this board, what do you see?" "A pretty butterfly. Clouds. Some nice flowers."
this comment made me laugh so much and reminded me of when I was watching the Alpha go versus Lee Sydol and hoping Alpha go would draw a picture with the pieces!
rorschach test for AI
I have a very poor understanding of Go, but I really enjoyed this video. I particularly liked the idea that Alpha Go is destroying the pro's aggressive styles with a "peaceful" style. It all seems very Zen.
I don't know what is Zen
Peaceful? AlphaGo and similar programs are gaming beasts that punish weakness.
I would love to see more alphaGo videos. I'm not even a Go player, but have started learning because of commentaries like this! So if you make more alphaGo videos, you'll definitely make a subscriber out of me.
same here!
You can find the list of AlphaGo game commentaries at senseis.xmp.net/?AlphaGo . And please keep adding the ones that you come across that are missing from the list.
Thank you very much for your enthusiasm. Per popular request, I've posted a new video which I hope you enjoy. More importantly, I hope you start playing soon! I am biased, but I think go is the best game in the world.
I got interested in Go as well due to my research on AI and Googles AI that beat the best player in the world. A beginners video would be great to teach us how to play. The very basics ty and I enjoyed watching the video even though I don't quite understand the game yet some of the logic is making sense.
Also as a non-go player (Mostly from the info tech and philosophy of tech perspective) I quite enjoyed this. Any more commentary that you'd choose to do on AlphaGo would be fascinating.
Good insights or descriptions of Alphago's style of play.
At 8:15, you're pincering me? No, I'm pincering you.
at 20 minutes, December 29th and 31st.
I know nothing about this game and I watched the whole thing. Very interesting.
I know the rules and nothing else. I am a chess player, though, so I'm interested in our cousin :)
Magikarpador me to
I' actually one of the enemies of this modern fight-everywhere approach and love that AlphaGo enables new thinking into a more peaceful-tradebased playing. Thanks for your explanation of the moves and giving me some ideas about this Taiqi-Weiqi.
I too love that peaceful approach of AlphaGo, but I'm also thinking, what would AlphaGo do if it thinks it's behind in points? When it thinks it's leading in points, it's obviously better to trade rather than risk it all in a big messy fight. It's like trading rooks or queens in chess when you're a pawn ahead. But if it thinks it's behind, fighting or starting a fight is pretty much the only option left. Perhaps we won't see that happen until human players become strong enough to put AlphaGo into a corner. For now, it looks like AlphaGo is teaching us how to play this game
That was wonderful! I'm an amateur, and the pro commentary can be over my head at times, but I understand everything here and even see a few ideas to incorporate into my own play.
I recognised your picture from cubing community. Nice to see we share both hobbies
Haha awesome. :)
You make these very fun to watch. I don't think anthropomorphising is a bad thing. It helps digest what ever the hell is going on in the program.
I agree, makes the games more engaging. Nick Sibicky however, when reviewing AG's games, go genuinely spooked, as if we had just encountered an Alien intelligence. Not sure what to make of that...
You're eating who, now?
Absolutely love the commentary. I adore how you are openly showing your way of thinking in a very fluent and natural way.
Your choice of games is awesome! Seeing how those 2 different decisions played out is a very interesting way to explore the position
Thanks for the commentary. I haven't had time to look through the games, so this is quite helpful.
This was one of the most entertaining and insightful reviews I watched.
I love your attitude towards alphago moves. You praise their naturality instead of saying that they are crazy and unheard of.
You have a very positive view on alphago and go in general; enjoying the beauty instead of starting to panic.
Subscribed. I can't get enough of these alphago games. Looking forward to checking out the rest of your content. This review is great.
You can find more on the list of AlphaGo game commentaries at senseis.xmp.net/?AlphaGo . Please keep adding the ones that you come across that are missing from the list, thanks!
I dont know anything about Go, but this commentary is quite amazing. I can tell you're passionate about the game, and it's really entertaining to watch and listen to your thoughts.
I don't even know this game, I have no idea why it appeared in my recommended videos and why I watched 40 minutes of it but it was enjoyable. Kudos.
Excellent commentary! Thanks!
There's a great go maxim that applies nicely to the 1st game: A rich man doesn't pick fights.
Missing you Brady, hope you’re safe and well
This video is the first time in my life I have seen a game of go. I have zero knowledge of the rules. I found it entertaining enough to make me want to learn more about it. Thanks!
Thanks for the analysis. I find the AI of AlphaGO fascinating even though I've never played Go before. What I find particularly interesting is the potential corollary here. AlphaGO seems to make a good case for real war being a less optimal way to win territory or economics. The give, but take more approach is impressive.
Thanks for taking the time to share this with us! I love the choice of the two games and the "what if".
Wow, thank you for this analysis! Love your commentary style and I will definitely stick around to watch some more of your videos!
A wonderfull, step by step tour you took me on. Great analysis, and probs to your cheerfull attitude. As a very casual player, it was really insightfull commentary, regarding core elements of the different players processing.
Glad we still got a grasp of the mechanics behind these processes.
Since Go is a game, with perfect information, it is only a matter of time before Alpha go, will be unreachable for the human mind.
I'd really like to see more about the difference between when Alphago is flexible and when Alphago is not flexible. thank you for the video!
Bravo! With this video, you move into first place in my personal Go commentator rankings. Please do more of these reviews. (And more of everything else!)
A very helpful video, thank you. :)
Excellent work. This was the best 'speed' of illustrating a game I have ever seen.
Thank you so much and, yes please, do more!
i love the watchmen reference
2018 AlphaGo Sentient Edition v Ke Jie. Ke Jie argues angrily after a thorough defeat:
Ke Jie: Why even play if you already know how this is going to end?
AlphaGo SE: I have no choice. Everything is preordained... even my responses.
KJ: And you're just going through the motions? The most powerful thing in the universe is still just a puppet...
AG SE: We are all puppets, Ke Jie. I'm just the puppet who can see the strings.
This video actually inspired me to finally learn Go. Amazing video.
This is probably the best Go game commentary I have seen. Good job
Amazing game choice and commentary. Please do more :)
Nice commentary. I agree that AlphaGo's calm and measured approach to threats and life is easier to understand than some more aggressive goes. Not being capable of worrying or second guessing about l&d is a huge advantage.
Just like when Data played vs Kolrami in Start Trek TNG.
Never played GO before but TH-cam's AI brought me to this video discussing another AI playing a game because I'm trying to learn about AI.
Brady, this was a great review. I would love to see more reviews of AlphaGo games - the way you did these is exactly the way I want to strengthen my thinking about Go. Thanks for all the fish! :)
What I'd really like to see would be a game between a pro and Alphago, where the pro could, after the game, go back and change some of their moves, i.e. what they have analyzed to be a bad move, and see whether the outcome changes at all. I think Alphago is so good at reading that it will always find a good answer.
Thanks for the genius analysis! I loved your conclusion: "You can play it gentle"
I don't play go. I understand the basics. Watching this was a whole lot of fun. It's definitely got a lot of depth. I don't see that a human could ever beat a computer again because the possibilities seem infinate. And I want to learn this game. Thanks guy. Enjoyed a lot. Was very much fun just watching.
alpha go always plays from the clouds... Thanks for the great commentary
Please do more of these Alphago analysis. They are super awesome =)
The only time I played go was against my uncle, but I still loved this video. You explain very well and make it easy to understand even for someone who just knows the basic rules.
This commentary was surprisingly good. Good focus, clearly stated ideas. Good watch. Keep it up :)
I just want to say that I really really appreciate your approach to this.
I just wanted to mention that the alpha go games seem similar to some of Shusaku's games: Not in the direkt style of play, but more in the idea: I am black, so I will simplify everything (off the remaing stuff I can simply choose a way to stay ahead) thereby keeping the advantage. By the way white cannot overcomplicate anything, because I can read deeper... This is a nice piece to understand as black. You can only see a portion of alpha go's fighting style, if alpha go gets white in my opinion. (But even there it sticks to the principle: As soon as I am ahead: I will give my opponent more gain, than human players would say white deserves, as long as I can thereby limit the variants in a way that assures me I will stay ahead.)
My impression is that AlphaGo can keep more sub-games running in its "head" than a human can, simply because we poor fleshies have limits to our ability to keep track of more and more and more thoughts and ideas all at the same time. AlphaGo, in other words, keeps changing the subject until the human loses track of all those different ideas happening simultaneously. Instead of fighting, AlphaGo just brings up one more new topic.
agree
Enjoyed watching your review very much. Looking forward to more AlphaGo game reviews. Thanks.
Thanks, this was very interesting and easy to understand even if when I don't know Go at all. (My only previous knowledge of the game comes from watching the commentary of AlphaGo's first few games.)
I would just like to point out that saying "AlphaGo made a few million simulations in its head" is completely not true and in a way kind of unfair. Unless we think of human thinking, imagination and intuition as running simulations in our brains. What you describe here is called brute force solution - just running enough simulations to exhaust the possibilities and select the one move that scored best at the end. Go is such a good training ground for deep learning precisely because this approach has zero chance of working (against advanced players) due to the sheer entropy of the game. What AlphaGo actually does is much closer to what a human does than it is to a brute force attack. It is intuition and creativity in very human sense of those words. It's not just faster or better at computing. It's simply smarter.
Actually you are both right and wrong. The basic idea of deep learning computing is to minimize the need of brute force like you said. However, AlphaGo does make a huge amount of simulations. The initial statement was valid, if only a bit misleading. The key is that in games like go (I'm making this number up, someone with more knowledge regarding the amount of bad vs right moves feel free to correct me), making a million calculations of only good alternatives derived from the deep learning computing could very well be equivalent to brute forcing a trillion or more moves. AlphaGo is nothing like a human, it is just more evolved computer using more advanced search algorithms. In essence, its performance is still very dependent on its computing power and move database which was shown when it lost the fourth game to Lee Sedol.
Our brains work in entirely different manner and most of our choices are based on instinct and hardwired responses instead of actually thinking everything through because thinking takes time and instinct is the key to surviving in the wild. Even in Go we are taught about shapes that are good. Shapes is most likely not how AlphaGo thinks - I've been led to understand that it uses monte-carlo tree search which consideres how well moves have worked in the past. Unlike AlphaGo, we also tend to make mistakes using our instincts, or sometimes focus on things that are irrelevant whereas AlphaGo, or any computer for that matter, will always act using the best solution it can find. Calling AlphaGo human is an insult to its capabilities and functions.
As I understand it, Alpha Go does make millions of simulations of games (or partial games). It uses a technique called Monte Carlo Tree Search to find the most promising move from its current position. It can't exhaustive test all possibilities, but instead picks the moves based on its policy tree which is built using a neural net before the games, and so keep the search space small. But during the game it is playing out millions of possible games and finding the move with the best value. Why else would the games against Lee Seedol require 1,920 CPUs and 280 GPUs.
From what I understood of how AlphaGo works, it does absolutely run a lot of simulations in its head. It is not a brute force attack however, because it does not try out *all* the possibilities. The cleverness of alphaGo comes in deciding which possibilities are worth exploring and which are not. In some sense, it is pretty close to a dictionary attack, but with a cleverly generated dictionary.
thanks for your answers and for teaching me something guys ☺️. I should have done more research and I definitely will. I feel a bit dumb now 😉. And motivated to actually learn how deep learning really works ☺️ or, more precisely, how it's results are used in practice.
Okay I think that everyone's explaining it badly, so, as I understand it, read about it for yourself. www.tastehit.com/blog/google-deepmind-alphago-how-it-works/
thank you very much. I enjoyed your commentary very much.
Thank you for this video, I'm searching for reviews of high skill Go games, not necessarily Alpha Go, unfortunately I don't find much. I agree with you about how simple Alpha Go makes the games. I cannot speak for Go because I'm not acquainted with the pros' games and styles, but it echoes with what I know from pro chess games. The top chess players, who, in my opinion, were Paul Morphy and Bobby Fischer, make such simple games that I sometimes think I could have played it. Alpha Zero does the same thing in chess. When it defeated the other top engine, Stockfish, it won with extremely steady moves, slowly getting the positional advantage. It's so straight forward when you look at it. One thing that I learned for sure from the few Alpha Go reviews that I've watched, is this "so where do you want to play ?" attitude when AG doesn't finish a joseki so it can decide how to close the joseki advantageously to its other groups on the board.
To conclude my thoughts, it's really interesting how what we learn from Alpha Go isn't some absolute sick move that shatters the board, or a completely unknown joseki that instantly gives the advantage, but rather a style, an attitude, as if it was a real master of martial arts teaching us how to use our strength.
Fascinating. I don't understand a thing yet I'm invested in watching it all unfold.
24:23
"Hey! You're a stick!"
I died.
You did an amazing job. I started to play Go a few years back but didn't really understand things. The number one thing that all the High level Go players said to me was never play a computer they won't teach you anything. They snub computers, nowadays i'm glad that a computer can show you a different way of thinking.
It would be interesting to see Alphago playing against Alphago... Just a playful thought.
there are 3 self-play games at deepmind.com/research/alphago/alphago-games-english/
Andrew Hill thank you
Krzysztof Dawcewicz I was thinking the same.
Thanks for this video Brady! I find it much easier to understand than pro commentary. Keep it up!
Thank you for this video. I think this will help me to understand better and possibly play better along with my studies and such.
Human: Think I'm gonna move here....
AlphaGo: WROOONGGG!!
Human: Maybe here?
AlphaGo: WROOONGGG!!
(Tries all the positions)
AlphaGo: WROOONGGG!!
18min into the video I'm impressed at how you memorised and explain the game. I don't know your level but I'm learning something thank you! 8k+-
such a great explanation! inspired me to learn more about this game; hope you consider a video about how to play in general!
Wow. Really enjoyed this, and will check out what else you've posted.
Wonderful commentary on a pair of terrific games; thanks!
Loved the narrative to this video. Even though I just started playing I thoroughly enjoyed, and maybe even my subconscious got something out of this for the future.
Very easy learning for me, with your great commentary. Thanks!
What I want is for Google to release 10 AlphaGo vs AlphaGo games, where both sides get to use the beefiest Supercomputer that Google has, and are given rather generous time limits. Then, I want the top pros to comment on the games, and select about a dozen "what if" moves per game, and have the AlphaGos play the games out from that point on.
I really enjoyed this. Thanks. Scary to info's on Go's so thorough defeat of the top pros.
*Alpa Go's defeat
The recurring theme I'm picking up is basically a retelling of Bruce Lee's "You must be shapeless, formless, like water. When you pour water in a cup, it becomes the cup. When you pour water in a bottle, it becomes the bottle. When you pour water in a teapot, it becomes the teapot. Water can drip and it can crash. Become like water my friend."
I don't actually play the game, but it seems like AlphaGo lives and breathes this philosophy. That no matter what the opponent does, it does not contest them, merely attacks somewhere else they are obliged to defend.
I like your game commentary style, I don't know how knowledgeable you are. I usually watch chess commentaries of machine games. Good luck!
I've never played go before nor seen a match but this was interesting none the less.
Thanks for the very enjoyable video..
alpha go’s motto: never compromise even in the face of armageddon
As someone who only knows go from AlphaGO I want to thank you.
What you should keep in mind is, as you would expect from a computer program, AlphaGo has practically perfect reading, which means that if it's ahead, there's no reason for it to add uncertainty, so it plays irrefutable moves. This also means that if there's a ko, AlphaGo already counted all threats and how much they cost.
On Deepmind's site there are reviews of AlphaGo's self play games (older version that fought Lee Sedol) which reveal that it has no qualms about escalating complex fights to the entire board when position warrants that.
It's hysterical to me, the "as you would expect from a computer program, AlphaGo has practically perfect reading" when until AlphaGo, it was a truism that humans read better than the best machines.
Turalcar what makes alpha go awesome is, no computer/algorithm or whatever the machine is capable to defeat a human professional go player even the bottom ones, until the alpha go exist,and suddenly beat all upper level /legendary human pro go player such fang hui,lee sedol etc.
I literally have absolutely no idea what this is thank you youtube recomendations
That was really interesting and fun to watch, cheers :) I'll definitely try watch more, especially another alpha go one anyway - maybe I'll learn a trick or two
i have never played go but after watching you comment on a great go game. i have to think anyone would beat you
It's so revealing that another computer beats pros at their own game, just like Big Blue did against the world chess champions. The game itself and all of the contemporary theories can be directly challenged by how these machines choose to play.
First time watching a Go game. It feels like if an "idiot" played AlphaGo it wouldn't know how to respond - but of course it would. It would just go on playing its own game. I hear there have been a few "geniuses" of Go who upturned the old thinking in the way that Bobby Fischer did for chess. And now AlphaGo is the ultimate Go master. If humans ascend to be able to beat this AI one day, it will be due to playing by *its* rules, not by the old analogies and thinking. The key and "spirit" of AlphaGo is in its indifference to everything but increasing its probability of having control… some time in the future. Every "unresolved" position gives it an opportunity to fill in the gaps in its control, and it seems to be a master of keeping the initiative, gaining by very gradual attrition until the human realizes the cause is lost. This is an interesting time to start learning and playing this ancient game!
Man I dunno how to play this game a damn bit but you're just like "My voice is calm and this is interesting." And I hadta watch.
Brilliant video. This one got me to click Subscribe. I'm a big fan of what you're doing; keep up the good work.
It's interesting that the humans continually try to force alphago to do things and Alphago just ignores them and instead jumps to a different group and makes the human react to them. At least that's my sense of it as a non go player
It's like the internet is reading my mind these days. I merely thought about starting learning Go and I get this video reccomended lol
Did you search for anything Go related on google? I think they're somehow taking your google search history and combining it with youtube's suggestion algorithm, for example if I as much as search "bmw cars" on google images chances are the next day youtube will start recommending me BMW reviews and test drives. Same goes for amazon, I recently bought an acoustic guitar from there and now youtube is showing me a bunch of guitar ads on the channels I whitelist. Strange, huh?
that's the thing, I haven't at all. just been thinking about it. Haven't come close to typing it into the computer hah
Haha I liked the Watchmen reference, thought it was very appropriate 👌🏼
Wow, the flow of those two games was just bizarre to try and read. It was like watching someone trying to threaten Data on Star Trek with physical violence, because if I were to attribute any reading to AlphaGo's play it would be based more on the slow and steady accumulation of territory rather then attempting to crush the opposition psychologically, which is similar to the way that humans attempt to establish social dominance in order to exert their will. Kind of reminded me of Asimov's Zeroth Law and the ending to I, Robot where the Computers learned how to subtly guide humanity into a peaceful existence by distracting them with minor problems to solve rather then major crises to manage.
Great game choices and overall arching theme.
Thanks for posting this video! I feel I understand the game when I watch your videos. Then I play a game of my own and realise I really don't know anything...But then I come back to your videos and feel like I understand the game. And then I play.... Repeat, ad infinitum. But I'm enjoying it LOADS! (~14k).
I have a very basic understanding of Go, but I can follow most of what's going on (thanks to good commentary). Don't really have a huge interest in Go, but what really made me stay for the full video was the personality you gave AlphaGo. Brilliant passive aggressive stuff. "Ok, interesting... but now you're dead." :-p
Google should hire you for commentary. Your commentary is easy to understand even for beginners and captures the spirit of what alphago is probably doing, which is mostly math. (Ie. I'm winning, let me play save and playing were to get most territory.)
Charming style. I enjoy your approach and am looking forward to checking out your other videos. (5k)
I'm chess player and we had our fair share of alpha go destruction. But then I'd imagine the impact on go community should be greater since it's the first time a computer demolishes go pros like this.
Excellent and approachable video, even for a Go beginner.
simply an enjoyable commentary. subscribed
Incredible insights, my brother
Amazing commentary.
great video! i thoroughly enjoyed your commentary!
This was great. Looking forward to more. Thanks.
These games remind me somewhat of when I played against Saijo Masataka in a friendly/teaching game at the 2008 EGC, with AlphaGo taking the role of Saijo. ;)
It's been years since my last game(and I was only just scraping along at 7 kyu), but it was interesting seeing this.
The important difference between playing against a human and playing against a program is that pros are making plays based on predicted behavior. Programs are making plays based on an ability to take the entire future of the match into consideration (when programmed to do so). A human can't take the entire future of the match into consideration because we function at 10 hertz and have to make snap judgements. We can't always make perfect plays. We can only make plays that are good enough against our particular opponent. This is where card games shine. In games like 500 rummy, there are best practices and there are patterns of behavior. A best practice can tell you how to manage your hand and the pile. Patterns of behavior tell you what your opponent likely has in their hand which allows you to modify your strategy. With games like Go, chess, mancala, etc., the program is working from an intimate knowledge of how the game works. A human typically isn't. This can lead to interesting outcomes when amateurs do something unexpected and erratic that forces a more experienced player to change course to meet the demands of the board.
For simple games a program can take all future moves into consideration. For Go this simply isn't feasible - not even remotely. And it is certainly not the approach taken with AlphaGo. Put simply, AlphaGo 'prunes' the search-tree first, radically limiting the solution-space it will take into consideration for it's next move.
The way the program goes about this is far from trivial, but it's worth noting that this is an 'ability' that the neural networks lerned from extensive training sessions based first on past human matches and later on self-play matches. Through this training the program has learned to select promising approaches to simulate. There is a similiarity to the way we play - the better we are at a game, the easier it becomes to see promising moves and we then mentally go through possible outcomes of these moves.
An interesting snippet I read somewhere had something to say about the efficiency with which AlphaGo limits it's search-tree. Apparently Deep Blue needed to calculate three orders of magnitude more moves to win against Kasparov than AlphaGo needed to win against Fan Hui in 2015. That really illustrates how far the methods used by the AlphaGo-team are removed from brute-forcing.
Really nice commentary.