Brady Daniels: Appreciating AlphaGo

Hi, welcome to Brady’s Bones, where I usually show you my mistakes, but today I’m going to continue with the Chronicles of AlphaGo and its recent 60 game series against the world’s top professionals. And I’ve got good news and bad news. So let’s start with the bad news. And that is Chinese professional Jenny Shen offered to do a video with me. And it was going to be this video. But it turned out when we started recording a couple nights ago that my audio visual skills are even worse than my go skills and we couldn’t make it work. And I need to install some new software and teach myself how to use it to get the sounds correct. So that’s the bad news. Jenny’s not here. Sorry. And sorry Jenny, I couldn’t make it work, but thank you so much for contributions. And that leads to the good news. The good news is Jenny and I still review the game. So in addition to my thoughts, we have her feedback as well, which is really cool. The other piece of good news is Jenny’s agreed to do a game review with me once I have my act together. So soon I hope. One of the major themes of my last video was that we amateurs can appreciate and learn from the games of AlphaGo. And I’m going to continue with that theme in this video and assuming I do another with that one as well. But there are a few comments made in various fora, YouTube comments read it, life in 1919, where people challenge that assertion a little bit. So the first warning is 100% legitimate and I agree with it completely. And that is a concern that some amateurs take away from AlphaGo, maybe you can avoid fighting and win, right? And because AlphaGo seems to avoid fighting and win. And that would be the wrong lesson. And if you avoid fighting, there will be a natural cap to your progress, right? If in a game that you fall in behind, if you refuse to create complications and fight, then you will just march your way to certain defeat. And if because you don’t like to fight, you’re not very good at it, then even when an opponent is well behind, you won’t have an answer when they choose to invade your network unresumably, make an overplay. But they’re better at fighting, so they easily live in your area where they shouldn’t have. So the take of way of AlphaGo shouldn’t be avoid fighting at all costs. Matter of fact, I’d make the opposite claim. One has to be a good fighter to get good at go. And we know AlphaGo plays some amazing fighting games because we’ve seen it in the three games that DeepBind provided us of AlphaGo versus AlphaGo. And in those games, AlphaGo, one of the versions of AlphaGo, always believed it was behind. And because it was behind, it created complications and we got these crazy, whole-board fights that we’re great fun to look at. So the takeaway is that AlphaGo can’t fight. The takeaway is basically in these 60 games, AlphaGo believed it was ahead and didn’t need to fight. Okay. The other challenge from the last video was that the assertion that we as amateurs can learn from AlphaGo. And one person said something to the effect of, if you can’t understand professional games because there’s so much stronger than you, then you can’t understand. There’s no way you can understand the games of AlphaGo who is so much stronger than them. And that’s sort of mistaking the point I was making. The reason I find contemporary professional go games difficult to understand is because they’re complicated because the fights break out so early that it’s actually very, very technical. And therefore I need pro guidance. And in these games with AlphaGo, it’s not like that. The games tend to be less fighting oriented than contemporary professional go games. And as a result, it’s just easier to understand. And I take it a step further. I actually feel like AlphaGo’s almost, AlphaGo almost created these teaching games. I know that’s totally anthropomorphizing AlphaGo and giving an intent that really wasn’t there, but there are a lot of moves that AlphaGo played that in hindsight, even for amateurs become clear. And there’s a good example in the game I’m going to show it today. Speaking of which, enough introduction. Let’s get into this game. And this game was played on December 30th. And AlphaGo was white and he was playing Leant Xia, who’s a really strong Chinese professional. According to GoRidins.org, he’s around the 10th strongest human being in the world. And he’s won multiple titles. And one funny thing that happened is almost a year ago, almost exactly a year ago, he famously said computers are a long way from being able to beat the strongest pros, which was just a wonderful instance of terrible timing because AlphaGo was introduced just after he said that. And this was his chance to take a shot at AlphaGo. All right. So again, Leant Xia is black. And he played here. Now this is an opening I haven’t really seen before. And if this black stone was over one, it would be a low Chinese opening, which is one very common. So my thinking as an amateur would be, well, if this gap is one space wider than it would be in the Chinese opening, then it must be more invatable. And it seems like a good idea to jump in right away. And I would have played here. And it turns out that’s exactly what AlphaGo played. So yay. However, the follow-up was completely different. And my intention would have been to play something like this, which I figured, you know, it isn’t great. I’d rather be able to expand one space further with that stone, but at least number one, I’ve got a base and number two, I’ve got outside access with these stones. So I should be OK. And when I explained that thinking to Jenny, she kind of laughed and said, I don’t think there’s many professionals who would play this. And the reason is, the shape’s not done. And because it’s not done, white’s going to get pushed around a little bit more. If you imagine a black stone here, white begins to get uncomfortable, whether black pushes here or here or here, white won’t be completely alive and is going to be forced to run. And black would continue to attack and profit from that attack. And if white goes ahead and plays here to make enough ice base, then you can see that, you know, whether black plays this right away or not, black is controlling the outside and seems much better in this situation. So the move that I would naturally play is, well, not ideal. And off we go, didn’t play it. And instead, off we go, went ahead and started an avalanche. And so this is a common joseki that we call the small avalanche. And off we go, went ahead and played it. Now it’s kind of curious because when off we go play this, it didn’t feel right to me. And when I learned this joseki, I learned that there’s one absolute requirement to play it as white, but also that there’s a general principle that you don’t want to play this as white. So I’ll explain both really quickly. So the absolute requirement to play this version of the small avalanche is the ladder has to be good for you as white. So if black was to play here and then white was to cut here, white has to be able to survive this ladder. And in this case, there’s a white stone right there, perfectly positioned to stop the ladder so white can play this. Okay, so as far as I know, that one is still absolute. But the other thing that I was taught is this is a bad choice of joseki when black has a stone around A. And that is because this joseki starts a fight. And the black has a stone around A that already helps black tremendously in the fight, making us a bad choice. Here, I’ll show you exactly what I mean. Black turned here, white cut, black played here, white connected, black played here, white turned and black captured. So right now black is totally safe in the corner, has the corner. But you can see that we’ve got these two white groups. Now this white group has a flexible shape. So it’s a little weak. It’s not really weak. It’s a little weak. It could be a target. But on the other hand, this white group looks really weak because it wants to run in this direction and in the direction it wants to run, there’s a black stone. So one of the principles we learn, one of the preconceived ideas if you will, is this joseki is bad if black already has the pincer stone in place. And AlphaGo doesn’t care about our preconceived notions. Okay, before I start, I don’t want to get into trouble here. And honestly, the strong pros are also beginning to doubt some of these preconceived notions. Not just in general, but even in this very specific joseki. And there are professionals who have played it, despite the preconceived notion being that this is bad. So I don’t want to make AlphaGo into a godlike figure that only AlphaGo would do this. But it’s kind of cool that AlphaGo does it. And I think we can really learn from it. Alright, so let me show what would be the outcome if I was to have played this. So as white, I would think, alright, now I have to get out with this group as fast as I can. And I would play a diagonal move like this. And you know, presumed follow up with something along these lines. And black could turn here, white would have to make shape, and then black would cap. And this seems like a really, really terrible result for white. All white stones are doing now is running. It’s not really making an attack on this black group because this black group is outside. So as white runs in this direction, white’s black’s going to build a lot of territory up here and continue to cover white and control the center. So this seems like a really bad outcome for white. And what would happen if I was to play this joseki. But what that actually shows is I lack the technical skills to be a really flexible player. All I’m trying, let’s go ahead and rewind and look at a couple of variations. Actually, white’s next move really surprised me. And it was here. Alright, so when white played this, I’m black played here, I thought, okay, what white is doing is sacrificing these three stones and white’s going to build some wall out here which works well with these two stones. And boy, white’s giving up a lot of points to do that, but maybe that’s okay. And needless to say, I was wrong. Probably the most important takeaway from this shape is don’t try to live with these three stones as white. Not straight up. Don’t make that clear and only intent. You could try something like this. Reality is black and double hane, but even let’s say that black couldn’t, right? And black kept playing like this. So you could, and let’s assume that you can live with this white shape here. But we have to understand is the result isn’t good. All of your moves are making one point at a time and meanwhile black is building a wonderful wall facing me outside. So this just isn’t the flexible enough way to play. Oh, and by the way, yeah, the double hane. You can see here that black can play this way and then A and B are mei, right? Can go ahead and make a safe extension from this group with B or why can help this group? But whichever one but black white helps, black is going to attack the other. And this would be an acceptable result for black. All right. Jenny said the first move professional would think about today in this position would be this, right? And the point that white is making is I don’t have to save those three stones. But if I can, that would be nice. But if you as black really, really want to eat those three stones, then you can, but the result won’t be good. All right. So for example, let’s say black, black subjective is just to make sure to kill these three stones. Black can do it, right? Black can play like this and definitely eat those three stones. But it’s not a good result for black. Because here’s white puncturing right through black’s area, making this stone look stupid. And black just spent a lot of energy just to capture three stones. And this would be a perfectly good result for white. So the objective for white shouldn’t be to save the three stones. And the purpose for black also shouldn’t be to capture the three stones. And that makes it really interesting. Complex. So if black agrees that capturing those three white stones as is is probably a bad idea, can he try to maintain good shape and capture everything at a larger scale? Well, it’s not so easy for him. It turns out that what white is trying to do is to cut with the sequence white, one, black, two, white, three. And then if white can capture either cutting stone, either this cutting stone or these two, then white’s going to be able to settle. And then it’s going to be able to get it himself pretty easily and get a reasonable result. So here, let me show you a real quick example. Something like this. So let’s say this was to happen. For certain, white can’t keep everything. Black can play here and then come back and capture the stone. But now white can capture these three and do it with some style. Now white’s got a living shape as well as a wall that’s facing the rest of the board. This would be an acceptable result to white. Remember, white doesn’t need everything. White started with one stone and black started with three in the area. White was a huge disadvantage. And with this result gets a reasonable result. And that’s all that white really wants. So the point of this is that a move like this is flexible. And whether white saves the three stones or controls the outside or hurts the left side, white doesn’t care. So long’s white gets something. Just some sort of result will be good enough. So what did AlphaGo play? AlphaGo played here, which to my knowledge hasn’t been played before. But it has the same intent as that move that I showed earlier. So black read that the honey doesn’t work and extended. So white extended too. And then black played here, helping this group. So why did black play this instead of hunting over and trying to cut these stones? So we don’t know for sure how AlphaGo would have answered in this situation. But it’s pretty clear that white’s able to get a reasonable result. And then something like this seems to work. And there’s more variations than I’m going to be able to go through. But in addition to that one, two, three combination, we looked at earlier. Now there’s a cutting point at A for black to worry about. But if black is stubborn and says, you know, I’m going to cut you off, then white can just connect. You can see that black has problems at both A and B. And if black fixes to shape this way, then white’s perfectly happy to just come out here and capture those four stones at a large scale. This would be a great result for white. If black tries to be tricky and hyper efficient and play a move like this and say, well, I’m not going to cover those cutting points, then white can still get a decent result by cutting. White cuts here, black pulls out this way, then white plays here. And again, black’s in a tricky situation. If you try to come out with these stones, then it’s no good because white can capture these three stones in a ladder, right? Black here, white here, black here, ladder. And again, white’s really happy. So this doesn’t work. But if black just extends again, then white is perfectly happy to eat these stones at a large scale. Again, we don’t know whether this is what I’ll have to go ahead in mind because it wasn’t played. But it seems to work, right? At any rate, I think that explains why Leanne’s show played here. And then I’ll go and play that harmony on the head of two stones, which if Alph goes human, I would say has to feel really good. At this moment, Leanne’s show bent around these stones. So if with this earlier movement, the head of the two black stones, then here he hits the butt of the two white stones and black blocks and white connects. And here’s where we get to a really interesting point because black plays here, helping his central stones. And it’s a really interesting question. We don’t know if Leanne’s show was offering a trade at this point or if he thought this move was sent to and that white would surely answer. But Alph will go look at the ball and said, hmm, this is a trade I’m willing to make and play it here. And that is would be a hard move for even professionals to play. Let me try to explain why. Well, let me show a little bit of the continuation first and then I’ll explain why. Black tried to fix his shape like this and Alph will go find this to Suji and it’s really quite cool. And my question for the viewers is, why can’t Black play here next? And Leanne’s show was in 32nd Bialiomi. This is effectively a blitz game, but he saw why he couldn’t play that and can you. And if you want to pause and read for a minute, go ahead and do so because it’s kind of a cool problem and you’ll like it. All right. I’m going to show the answer now. And that is it just plain doesn’t work, right? If White tries to connect underneath like this and Black blocks, now White can force here and then cover here and this Black group is captured and really cool now. So that means that this move here is Suji. So Black answered this way and then White connected all of his stunts. And in effect what White said with this is all this potential territory you have at the top is mine now. And furthermore, your shape is crap, right? Your connection with this stone doesn’t look good and you don’t have a base. And yeah, I mean, AlphaGo seems to get a really good position out of this. But the problem is Black is able to play this and eat all of those White stones at a really large scale. So why was this outcome not entirely predictable and what if two humans were playing this? And again, this feedback is from Jenny. So this isn’t my thinking on this one. And it’s that when you’re White and you play here, it’s hard to know what the final outcome is going to be. And therefore it’s hard for you to assess is White, oh, I should get X compensation. So you can see how much Black is going to get by capturing all of this. But it’s hard to figure out what White is going to get, what the final shape is going to be. And if you don’t know what that is, then it’s very hard to make the trade, right? After all, it’s White’s choice to make this trade, but it’s a difficult choice to make, at least for humans. And this gets to the greatest strength, I think, of AlphaGo, and that it’s positional judgment, right? For humans, we’re able to look at the upper right and say, OK, Black’s going to gain something like 40 points overall with this capture. But it’s very difficult to say exactly what White is going to gain this compensation, whereas AlphaGo has its value net and it’s able to make an estimate of what it’s going to get. And if it’s not doing its estimate at raw points, like I’m going to gain an equal number of points, it gains its estimate as a percentage increase in terms of winning. And clearly, when AlphaGo played that move, instead of answering, it thought, yes, this move gives me the best chance of winning, which means it took a good trade. But Jenny’s feeling is that most pros would play here because they’d prefer the certainty, right? And the idea being is White’s going to allow the fight to occur in the center, right, where the Black and the White groups both run together towards the center, and White will be good on the right side, because that’s something that we understand, right? A running battle between two running groups. And we can, whereas White just uncertain. Turns out it’s great, but this is what I mean by AlphaGo’s positional judgment is its greatest strength. For us, this looks really uncertain, but for AlphaGo, it’s able to say, yes, this is the move that’s going to hand me the win. So I’ve already talked about other things that AlphaGo is good at, and I’ll talk about more, like the flexibility, you know, with this attachment here. But that flexibility doesn’t feel inhuman, right? The strongest human-go players are really flexible, and they’re willing to switch groups and so on. I’m going to talk about technical skills soon, but spoiler. AlphaGo’s got great technical skills, but doesn’t seem like it reads better than humans. It’s, you know, in the same ballpark, anyway. Where AlphaGo seems to have the biggest advantage over humans is this positional judgment, this ability to know which move is biggest at any given time. Okay. And now it’s when we get to another really cool part of this game, and I told you that one of the challenges to my last video was if AlphaGo is so much smarter than humans, then we can’t learn from it, right? Well, certainly not you as an amateur can’t. And I don’t agree. I think that AlphaGo gives a certain teachable modes, you know, where it says this is the correct way to play. Let me show you. And this is one of them. And according to Jenny, a lot of pros wouldn’t want to play in this area as white, because they’d want to leave Ajie. And Ajie is the potential to maybe it’s to revive the group or do something with the group and it’s not clear how you’ll use that potential. So instead of spoiling the potential, they play, they would play elsewhere. Whereas AlphaGo looks at this and says number one, I think I’m winning. And number two, I want to simplify the game. So number three, how can I minimize black’s game? Right? Seems to do those things right now. And it says to black, do you really want to eat those white stones? Black says, I surely do. And it says, no, no, really, do you want to eat those black stones? And Leon Chow says, yep. And again, sure, yep, really sure? Yep. Okay, you can have them. And this is so elegant, isn’t it? It’s like I’m going to minimize the territory you gain from that capture. And at the same time that I’m minimizing what you get, oh, wow, look, I built the lower right. Pretty cool, huh? And so instead of worrying about IG Keshi and ruining its own potential, AlphaGo just goes ahead and simplifies the game as it marches towards victory. Okay, so the biggest area right now is the bottom. So black moved over here, white did a two-wheel answer, and then black played here. And AlphaGo really seems to like to pincer in this situation. And Leon Chow played on top. And oh, oh, oh, oh, let me back up. I want you one more cool thing. This is not the most common, just second, black’s choice here. Let me show you the most common. I’m sure you’ve all seen it. Usually black just takes the corner. But then white really wants a ladder to be good for it when it plays the preferred follow-up, which is this. But so is the ladder good. And why is the ladder good? Peep, peep, peep, peep, peep, peep, peep, peep, peep, peep, peep, peep, peep, peep, oh, it is good. Oh, it was that play that made black eat those stones that turns out to be perfectly positioned to make this ladder good and let white control the right side. Is that cool or what? Oh, well, black didn’t play that, so. But anyway, it’s really cool. Maybe that’s why black didn’t play that. Anyway. So black moved up to the center like this and white left the shape as it is. And now came up here. This is where white says, all right, I’m going to take advantage of your bad shape. Now, it’s kind of funny. When I looked at this game, I looked at sort of the expected moves and thought, oh, this doesn’t look so good for black. And Jenny said, actually, all the pros in the world would do exactly the same thing. And but whereas I play out the moves one by one in my head, she said, no, I’m close. We just kind of know what the finish shape looks like and decided it’s unacceptable. Let me show you what I mean. And so without playing the sequence in their minds, the pros would all see this sequence. And be able to say that even if white just pushes and pushes and pushes, this side is no longer interesting because of white stone here, right? So whatever black, whatever points black gets here, it’s going to be so great. But in the meantime, white is sucking up these five black stones at a larger and larger scale. And this would definitely lead to a lost game for black. So it would only take just glancing at the board after white plays here for black to say, oh no, that variation is just awful. So even in 32nd Blitz, where black can’t figure out and read out the full sequence that follows, it just feels like a moral imperative. The black has to push and cut and make do. I thought that was a pretty interesting way of looking at it. And this is where we get a little glimpse into AlphaGo’s technical competence, if you will. And what I mean by technical competence is it’s ability to find the best moves in the sequence when the stones come into contact. So it’s not just being able to read, but it’s also finding the best moves. And in a way, it’s understanding the wise. And let me show you exactly what I mean. And white’s got a real problem with this shape. And if I was playing this game, especially in Blitz, I think there’s only a 50% chance that I’d make the right first move as white. There’s a 50% chance I would play this and instantly lose. Because black’s got this threat. And if black plays this way, it can capture those stones in a ladder. And all of a sudden, what was a really bad corner for black becomes pretty good. So 50% chance I’d go wrong. I personally would go wrong already because of this move. And instead, this is what’s required. Because now you see that this push doesn’t work anymore because it attires the black stones. All right. But then when black played this, there’d be 100% chance I’d get the wrong move now. Just by reflex, I would play this. Black would connect. And I’d have to help my shape. So now we have the same problem. And you can see that these two black stones are of two white stones are doomed. So huh. And this is what we’re talking about when I’m talking about when I say the technical skills. White’s only play is to play here. Even though it looks like that stone is going to die, it turns out that it’s okay. And the reason this is important, it’s, again, it’s liberties on these two stones. And you want to keep those liberties at two. Black can’t really capture because white’s totally happy to give up that stone. Come on this way. And not even sure what the best shape is. Maybe something like this. So even if black lives in the corner with this, it’s really, really small. And you can see that white is still capturing all of these stones at scale. So this would be a perfectly good result for white. So that doesn’t work for black. So black can’t capture and play it here. So again, that’s threatening these two stones, right? So white plays here, getting liberties with this stone and making and threatening these. So black protects. White protects these two stones with this Atari. Black moves out. White covers by saving this stone. Black threatens these two stones. So white moves this way. And then black says fine, I’ll eat those two stones. And white says no, I’m going to make you eat those two stones. And then white played here. So it’s kind of funny. The pro’s natural reaction was to look at that variation where black came along this side and say no, no, no, that’s not good enough. I have to fight. But I’ll go had the technical skills to say, OK, I understand you want me to fight, but this results even better for me. And right now the game is really, really hard for black. OK. Black played a forcing move here. White connected. Black played here. Can’t let white have the whole left side. So white takes black’s base. Black lives. And then white counterpinsers. And then black plays here. And this is another thing that I like about AlphaGo. There’s no doubt in my mind where I would play right now. I would say that these two black stones are weak. And I can exploit the bad shape while still moving out right here. And maybe this would be fine. It probably would. But AlphaGo in its wisdom said, no, I don’t need to fight again. I don’t need to push those stones around. I’ve already got a one game. I’m just going to go over here, connect with these three stones and secure the corner. And this is a move we see a lot from AlphaGo where it secures the corner with two night moves instead of what we’re taught, which is this, right? This is the shape that secures the corner. And I think this is a shape that almost all humans would make. But I think AlphaGo knows that with the help of this stone and with the help of these three stones, that there’s nothing Black’s going to be able to do inside. So I’m just going to go ahead and play here and make a better connection with those other two stones. And in effect, make a declaration that I win. And this is where we get to the last feature, the last thing to appreciate about AlphaGo in this game, which is its willingness to punish over players. And Lance, Lance, yeah, said, okay, I’m going to try to live inside or at least get something out of this. After all, I’m one of the strongest professionals in the world. I should be able to do something. And AlphaGo said, I don’t think you can. And AlphaGo said, I’ll just eat this one. And then Lance, Lance, yeah, I made this really cool forcing move. And the reason it’s cool is it’s threatening to push through and then Atari and that stone and give these stones a short of your liberties. But it turns out not to matter that much. So Lance, yeah, connects AlphaGo comes out. Lance, yeah, comes out. And AlphaGo solves that problem that I just talked about by making the Tigers now. And then cover, not that way. This way. A movie that threatens to split the groups. AlphaGo says no. And then no again. And here we see a professional go player in his death throws because nothing’s going to work for Lance, yeah. At this point. At this point he resigns because White’s totally fine on this side. Totally fine on this side. This group is dead. AlphaGo’s rich. Black is poor. Game over. All right, to be honest with you, this was just an average AlphaGo game. It didn’t have the biggest kills. It didn’t have, you know, it wasn’t the closest. There wasn’t much in the way of new moves. That was just average. But to me, there was still a whole lot to get out of the game. So that’s why I think it’s kind of cool. I really liked AlphaGo’s move at A and the flexibility. My concern about trying to apply that flexibility is it requires really good reading, which means that’s what I need to upgrade. I love the trade that it offered at B. Again, you could argue that Lance, yeah, offered the trade first with the satari. But White certainly accepted the trade with the. And I think that’s where it shows it’s great is trying, which is positional judgment. I thought the moves where AlphaGo made Lance, yeah, pick up the stones in the upper right hand corner. The whole continuation, you know, with C, was really elegant and something that I hope I can apply to my games. I think one thing that we need to admire about AlphaGo is its technical skills. Again, I don’t know that it reads better than pros. As a matter of fact, I’ve done nothing to show us that it does. But it certainly reads well enough. It’s got the technical chops to play with the big boys. And then finally, when the game was when he was clearly behind Lance Xiao invaded in the corner and rather than compromising, rather than just saying, well, I’m going to win anyway and letting him let AlphaGo said, no, this isn’t overplay and I’m going to punish it. So even if this is just an average AlphaGo game, I really like it. And I hope you liked it too. Let me know what you think in the comments. Hopefully I can get Jenny here soon. But let me know if you want to see more AlphaGo videos. And even let me know specifically what you’d like to see in more AlphaGo videos. Anyway, thanks so much for watching. Feel free to comment, like, subscribe, whatever. Thanks a lot. See you soon. Bye bye.