All Things Go Podcast

Benjamin Teuber and AI Sensei

[00:21] Travis: After the birth of my second son, my family was able to take a trip to Brazil because that’s where my wife’s family is from. And while I was there, I was starting to play around with AI Sensei, which is a really great website for using AI to identify your best moves, your worst moves in a game, but also having all these other features around saving board states in your game to learn from them and train on them and improve on them. I really wanted to connect with one of the founders of AI Sensei and so I hopped on Discord. I reached out to Benjamin Teuber, who is one of the co founders, and he was more than happy to chat. What was great is that as I was preparing for the interview, I learned all these things about Benjamin. And so you’ll notice that our conversation is sort of split 50, 50 between his personal background with Go, which is really impressive and also has some great themes and takeaways, and also AI Sensei and where AI Sensei has come from with the sort of AI revolution and where it’s going to.

[01:28] Travis: So let’s jump right into the conversation, take a listen. So, Ben Teuber, you were nice enough to give me a little bit of a rundown of your background. You’re 41 years old and you live in Chemnitz with your fiance and your family there. And you are a 6-dan amateur player with the European Go Federation. You have your Master’s in Computer science. It was actually really interesting to learn about your background because I reached out to you really as the main person and one of the founders of AI Sensei, which is just an incredible tool that I think has a wonderful interface, but you have this whole background with go. You’re a two time German champion, multiple time German Youth champion, German Pair Go champion. So you have a lot of background there and you started Go back when you were 12 and became a dan-level player at the age of 14, which is really impressive. And you’ve done quite a few stints in Japan and Korea, one time going to the schools, being an insei for a little period of time.

[02:42] Travis: Can you just tell us a little bit about how you got into Go originally?

[02:47] Benjamin Teuber: Yeah, sure. Thank you. So I heard some theory that most players, they start playing Go on the second contact with that game and actually that’s true for me as well. So first contact was just randomly that my mom’s boyfriend, he liked to go to flea markets and he brought back this Go board with some, some small rule book and of course it was quite ugly, like with plastic stones and Everything. So we read that and tried to play it and played a couple games. But of course it was just about capturing stones and missing Ataris and on the 19 by 19 board it was kind of crazy. And so I, I got bored from this and stopped doing that and left that in the cupboard for a couple years. Then I just happened to. I guess it was also coincidence. Like we were having breakfast and some nice cafe on a Sunday and next door there was some. We call it a Burgerhaus. So it’s like a public government building where a lot of clubs are meeting and they had this children’s festival where you could play like the table kicker and whatnot.

[04:06] Benjamin Teuber: So I just dropped by and played a bit and there were some people playing this game I had seen before and I told them, ah, this is cool, I know about this. And they were like, oh really? Do you want to play with us? And I’m yeah, sure. And they started teaching me properly on a nine by nine board. And that’s when I got really hooked. So my mom had to take me to this Bürgerhaus twice a week in the evening and pick me up ever since.

[04:35] Travis: Okay, great. So you got started there. It sounds like you had a pretty, pretty fast improvement path because, you know, you became a dan-level player within two years or so. What I really like about some videos that you’ve produced on YouTube about how to get stronger at playing Go is you’ve talked a little bit about some of the challenges that you’ve had to do on your way to becoming German champion and these sorts of things. And I think it’s actually really helpful for a lot of us to hear because we all know what it feels like as Go players to have a losing streak or we have a period where we’re just overly focused in our rank or something like this. Am I correct in understanding that you said you were a nine time runner up to the German championship? Is that correct?

[05:26] Benjamin Teuber: Nine or ten? I think at some point I stopped counting.

[05:30] Travis: Yeah, right.

[05:31] Benjamin Teuber: But yeah, it took me a lot of attempts.

[05:34] Travis: Can you just tell us a little bit about that? You talk about in one of your videos how what that sort of did when you were playing and how you thought about overcoming that?

[05:45] Benjamin Teuber: Yeah, I guess I just wanted it so bad that somehow when I was close to that goal, my mind was just spinning around like, oh, now I can really do it. And where I really need to be careful now not to mess it up. And I was full of thoughts about this nearby possible championship that I couldn’t focus enough on the Game itself. Right. So actually I think the first time I actually succeeded much much later was when I kind of had given up at this year. Like I didn’t practice much for the tournament and I don’t know, my thoughts were with completely different things and was just okay, I’m gonna give my best, whatever and I focus on something else and somehow that worked out the best for me.

[06:35] Travis: You talk about in your video about how that is a lot to do with your mindset and psychologically you’re thinking about having breathing techniques and remembering that it’s just a game, you’re not a professional. This is not something you need to do for your primary income. And then I love this focus on the 10 second rule. You know, just kind of pausing and waiting before you play a move.

[06:59] Benjamin Teuber: Yeah, I mean it can still happen, right? Even no matter how experienced you are. Like sometimes you just see a move and you just want to respond immediately just to prove your point to show yourself opponent that he’s wrong and that you’re right. But yeah, you always have enough time to think at least for 10 seconds unless it’s like a lightning game. So you should always just do that. I want to say something else about the focus on ranks. I think it’s much worse today than it used to be. Like when I learned Go, there was I think maybe some ELO system somewhere but we didn’t pay much attention to that. So it was basically your self declared, your rank and when you decided you beat enough 5-kyus you would call yourself a 4-kyu. So I think it was much less pressure on that. And actually like when I visited a tournament I was, I guess I was thinking of my rank a bit but I was also just okay, I just want to play. And then I managed to win four or five games and then I was, oh yeah, now I’m 4-kyu.

[08:03] Benjamin Teuber: That’s great. That’s the best thing that ever happened to me. It wasn’t like okay, but yeah, in a month I want to be three ranks up and then that and then that and where do I end up? It was just, I feel like I was more focusing on just winning my next game.

[08:17] Travis: In a way it makes a lot of sense and I think I can agree with you that it’s very easy to get overly focused on rank in our day and age. So I appreciate you telling us that. So then you visited Japan many times starting at age 14 and you did become a student of the late Hans Pietsch and his teacher Kobayashi Chizu. If I’m pronouncing that correctly. And you said that Hans really became like a father figure to you. I’m so sorry to hear about his passing. Can you tell us a little bit about your relationship with him and how that formed?

[08:57] Benjamin Teuber: Yeah, so I think I first met him at my first go congress, age 13. It was in Albano Terme. And actually my mom was kind of busy buying a house, so she couldn’t really join me. So some nice woman from the Go club, she took me with her and took care of me for that week we were there. Yeah, I heard a lot about Hans Pietsch because I learned from his teacher in Bremen, where I was living at the time. And he told me, yeah, if you study really hard, maybe you can also be going to Japan if you want, like he did. And so I had never seen him, but I was. Was like a football star for me. Right then I heard that apparently Hans Pietsch is at the congress. I was like, super exciting. And at some point, yeah, I just happened to kind of run into him. Like there was this big blonde man and he was immediately talking to me. Ah, you must be Benjamin. I heard a lot of you and I was like, speechless. It’s like Cristiano Ronaldo talks to you, like.

[10:08] Travis: Oh, right, right.

[10:11] Benjamin Teuber: And so he commented me some. Some games and it was very motivating for me. A year later at the congress in Marseille, he was there again. And Kobayashi Chizu was also there. I played some teaching game with her and then she like officially invited me to come to Japan. And so the Easter vacation after and ever since, pretty much all the Easter vacations, I would always fly to Japan and study there for like three or four weeks.

[10:39] Travis: And what was that experience like from going there the first time to going there many times?

[10:46] Benjamin Teuber: It was crazy, like completely new world. Also, I guess the. The top pro players and Kobayashi Chizu was one of the top female pro players. Right. They’re living more a bit like higher society life with some fancy restaurants and clubs. And she had like a Go club where every member had to be a president of a company or something.

[11:10] Travis: Wow.

[11:12] Benjamin Teuber: But yeah, of course, also just the different culture and everything, it’s really like, I guess it’s a bit like Harry Potter must have felt in Hogwarts, like being in this completely new world and everybody was being nice to them. I guess not everybody for him, but.

[11:29] Travis: So you did that every Easter you were going to Japan and then you had an opportunity to go to Korea. You spent a month or so there. What was that like compared to going

[11:40] Benjamin Teuber: to Japan, in a way, it was. It was quite different. I think. I feel like the Koreans and also the Chinese, they were even more serious about that, right? So they would really study pretty much all day in that school and in Japan at least. Also, when I was in Seoul there, it was a bit more relaxed, let’s say. Or maybe not not relaxed. It was more like, let’s say it was self studying. So you had like official games once a week on. On Saturday and Sunday. And there was no longer a school where the children would live. Like they had that at some point in Japan. But when I was insei, it had already stopped for several reasons. So actually all the. the insei, they stayed at home with their parents or maybe sometimes in a teacher’s house or something. And they would join a study group once or twice a week, or if they’re really lucky, maybe three times. But that’s pretty much it, right? They had games on two days and like some study groups where they could show and analyze maybe one game each.

[12:50] Benjamin Teuber: And for all the rest, they had to just do by themselves. In Korea and also in China, it’s more really, you’re always in the school and you’re always playing with the others, and you’re always getting reviews from the teachers. So it’s, I guess, a bit more focused and controlled, but also tougher. Back then in the Korean culture, it was still more common to like do some sort of physical punishments, or when they didn’t behave, somebody would really pull their ears or give them some slap with a ruler. Was a bit strange for me because I never seen this growing up myself,

[13:33] Travis: nor have I. Yeah, they call that corporal punishment in the U.S. but yeah, I’ve never seen that either, so that’d be a surprise. One thing that you talk about on Sensei’s library, you have a really nice layout of how to become stronger at go. One thing that you do reference there and also in your videos is that having a Go teacher is really important. But you also emphasize that it’s really important to think for yourself and really go through things yourself and not be overly reliant on a Go teacher. Can you tell us a little bit more about that?

[14:15] Benjamin Teuber: Yeah, I think actually, I mean, there’s several layers to this. So one is during the game, right? And one is for analyzing during analysis. It’s really important to understand. And you can only understand if you question it. You shouldn’t just think, oh, okay, I’ll just copy that. But maybe you only understood part of that. And I still fall into this trap. I Think just a few days ago, I played some online games. The opponent played some joseki I didn’t really, really know, but I thought I saw that shape in some go problem. I did. Okay, that’s this atari I play from here. And I just thought, okay, I seen this, I just play it. And of course it was not correct in this situation. So I think also for myself, I used to do that much more, but I still sometimes do it that I just think, I know this, I play it and I copy it. But you should always use your head and question and think, is that really the best in this unique board situation you have here? And what will be the follow ups? And think it through. What are the variations you do lay

[15:27] Travis: out these key concepts. So number one, as you said, you have to just like play, play, play, right? You got to play games. That’s really the main way you get stronger. And then you talk about doing life-and-death problems and tsumego analyzing your game. Do more tsumego after that. And then maybe you can review pro games. But don’t try to just blindly play the moves. Really try to understand what’s happening. And then maybe after those sorts of things, you can go to things like books and things of that nature. Any other things that you would want to talk about? Perhaps this is a good spot for us to transition to your work with AI Sensei.

[16:07] Benjamin Teuber: Yeah, actually, I think due to the breakthroughs in AI, it’s now much easier to analyze your games. Right? So I can tell you one story. Like when Hans was already kind of my teacher, I had visited him many times in Japan. I played some tournament games, like some German championship qualifiers or whatever. And so I wanted to get a review for that game, right? And I was already 5-dan. And so there were not so many strong players around to comment on that. So what did we do? Like, I sent my kifu over to Japan via fax.

[16:47] Travis: Oh my God.

[16:48] Benjamin Teuber: Are you serious? Wow. Yeah. And I think my mom’s phone bill was like exploding and like all my manually written down game records and I sent these to Hans and then he would review that and send me back some, some text with his comment on my moves, right? And so I got a pro game review for like, I don’t know, 30 euros of telephone, both sides. Wow. And that was how I sometimes was lucky enough to get really good high quality game reviews, right? And now you just upload a game to AI Sensei and you’re done. Of course, you cannot really compare it, right? Because of course Hans did all these comments explaining the why of it. And of course AI Sensei won’t do that, or unfortunately, we’re trying to work on that, but it’s not an easy task. But still, of course, it’s very useful to know where your mistakes were in the game and what you should have done instead. And many times you understand that yourself, and sometimes you don’t. Right? And I guess the weaker you are,

[17:59] Benjamin Teuber: the less you will understand why. You can maybe accept it and try to copy it, but it’s still helpful to have a teacher explain why this move is better. If the teacher can even understand. Sometimes with AI, maybe that’s not possible either.

[18:17] Travis: So you started working on AI Sensei back in 2018, and you said that first you were working with a gentleman named Eric, but you were doing a lot of the work yourself. And then later there were two other co founders that joined. And how did it start? Was there one moment where you said, oh, it’d be really great to have a better interface to use AI to review it? How did that process happen?

[18:44] Benjamin Teuber: Yeah, I mean, of course it started with AlphaGo in a way, right? Getting up very early to just watch that game with friends. And yeah, I guess at that point the normal Go AIs you could get, they also started getting stronger, Right? They, of course, were nowhere near professional level, but I think not so long after this AlphaGo match, I first managed to lose against an engine I had, which was, I think it was Leela. Like, not Leela 0, but the previous Leela. And the second I lost to that, I thought, oh, that would be nice to have as a teaching tool. Right. And so actually we did some experiments already back then, so it must have been a bit sooner. Maybe 2016 or 17, we wrote down a prototype where you could like upload an SGF file and then we would pipe that into Leela. It was already in the cloud, so in a way it was already kind of the architecture of AI Sensei.

[19:52] Benjamin Teuber: And we would then somehow use its output to generate another SGF file you could download with some explanations. Like, this move loses that many percent, I guess it was back then, and you should have played this variation instead. But at that point it was not really like, we tried it and showed it to a couple of people and we got some positive feedback, but I wasn’t completely sold because it messed up so many elementary life-and-death problems and things. So even though I sometimes lost against it, it didn’t feel like this is something I want to let loose on the Go community. So we kind of stopped that for now. And it was just maybe like two years later when Leela Zero was finally getting strong enough to be at pro level or at superhuman level. And then we somehow took the old code from the drawer and tried it again. And people liked that a lot. I also liked it a lot. And we kept working on that. I was still like freelancing for a company in Hamburg where I was living back then after studying there.

[21:05] Benjamin Teuber: And I took three months off to just focus on AI Sensei and to build a prototype before the Go Congress in Pisa. And yeah, we still didn’t make a lot of money with that, but still we got so much positive reception and we liked it so much ourselves. And at some point, even after coming back, I started working again in that company. But somehow it didn’t feel as exciting as working on my Go project. I guess that’s just where my love is or where I was always inspired all my life. And so I just, at some point I just quit completely, said, okay, now I want to focus on AI Sensei. I guess that was possible also because I was. We didn’t talk about that. I’m still working for Pandanet since many years now as a freelancer and so I get some basic level of income for just doing maintenance and sometimes new features for them. But yeah, that’s not a full time thing. So actually I could manage to focus on AI Sensei and I really like that. I mean,

[22:10] Benjamin Teuber: the first couple of years we made like no money and it was a bit hard, but some financial reserves. But I guess we’re getting to a point where I can live from this and this is pretty cool and very satisfying. And also I feel like in a way I’m kind of meant to do this even. I mean, I mean also the thing with Hans and with Japan, like where it was at some point I. I asked Hans how I can ever thank him for all he’s taught me and done for me, and he basically just said, yeah, don’t do anything for me, but try to pass on the knowledge to the next generation of go. I mean, I’m sometimes also teaching myself, but I think for me, because I also love programming, it feels much more natural to build a tool that can teach thousands instead of just teaching some stuff to a handful of students.

[23:07] Travis: Yeah, it absolutely has accomplished that goal. As one person, I can just speak to how incredibly thoughtful the user interface is. So I’m not a coder, I don’t have a coding background. But I know that it takes a lot of work to get all of these bugs out. And to have a really nice user interface where you can play against an AI, but particularly just the way that you have developed the actual review section. I think one thing that was really nice to see was how there is this training feature that you have. So not only can you make quizzes out of your own board states and spots that you want to improve, but that you can actually now benefit from everybody else’s board states has really been fun.

[23:53] Benjamin Teuber: Yeah, I mean, we had this idea for a while, but I guess we’ve been busy doing other things. But I guess the biggest inspiration was really just I got a bit into chess, I have to admit. I mean, not that much. At some point I installed light chess or got an account, and the thing I liked most was actually not playing, but doing their real game problems where you get a rating and after the problem, you would vote if you liked it or not. And so we thought, okay, we should do that too. And also it kind of made sense, right? Because you can already do your own problem collection on AI Sensei and practice from these. And I think it’s really great if you want to remember what you learned from a particular game even a year after, and try to not repeat your own mistakes. But of course, a lot of people when starting, they also want to do some problems, and they don’t have so many games themselves. And it’s kind of natural to offer problems for other situations too.

[24:58] Benjamin Teuber: I guess we could even consider adding some life-and-death or just isolated local problems at some point. But I guess in a way, it’s also fun to have this unique whole board real game Go problem thing going, which is, I guess, quite different from traditional tsumego.

[25:17] Travis: It’s interesting you say it, because personally, my Go teacher, that’s what he does, right? He’ll take snapshots of whole board situations in my game or from his other students and get that to me as homework to review. But to have it in an actual coalesced place on AI Sensei and not only be able to review your own, but have a very easy way to click the Remember button and it does that or shuffle the games where you do have those saved is really helpful because to your point, the local life-and-death problems are very important for becoming stronger at go, but it’s also always in the context of what’s happening on the board, right? So as you know very well, if there is a life-and-death situation in the corner, but let’s say I get hyper focused on wanting to live in the corner, that might feel Good. But I might say, hey, I know you want to live in that corner, but it’s actually not as valuable as this other thing that’s happening somewhere else in the board.

[26:18] Travis: So you also have that tool where you can actually restrict the zone and the area on the board to say, hey, AI, can you just look at this area? Which I think is incredibly powerful. Are there any new features or any new things that you all are working on right now?

[26:38] Benjamin Teuber: Yeah, actually there’s a couple of things, like one major thing which maybe is not so interesting for the users because it won’t give shiny new features, but it’s really just restructuring our internal database and how everything is done because it turns out like the tools you’re using for a new fresh website with only a couple of users can be quite different from the ones you need when you have like thousands of users and uploading, I don’t know, tens of thousands of games every day. We also focused a lot on Google infrastructure, which I guess is also has been nice to get started with, but it also gets starting to get a bit expensive and also it’s not that flexible. Like we consider trying to do something in the Chinese market or so. Right. But of course that’s impossible with a Google firewall, so we would kind of need to get rid of that completely,

[27:35] Benjamin Teuber: which is I guess a lot of work. But we’re still trying to move to a more traditional database where we at least have the option to be a bit less coupled to Google and I guess some other advantages we hope from that, like reducing costs or gaining performance. But yeah, we also recently got a bit more into doing something for beginners, actually. I guess traditionally you would say that AI tools is not really the best for beginner, right? Because we said the weaker you are, the more you need a human teacher. So we never really focused on beginners and that, okay, maybe it makes sense from 10-kyu upwards, but I guess there are just a lot of beginners that want to learn the game and play against the machine and learn something. Right. Even if you can’t explain all of that in a way, it just happened by, I would say by accident, like many things.

[28:31] Benjamin Teuber: Like we started working on an Android app for AI Sensei, and it’s not that sophisticated. It’s mostly like just the web browser in an app doing the same website does with some extra stuff that makes it easier to like upload games from other programs on your phone or some extra handling. The thing is that ever since we released that app to the App Store, even though we didn’t advertise it at all to our own users. Somehow we got a quite a big stream of registrations and I have no idea why exactly it happens because we see that they are not coming from the app itself, but somehow it’s connected to the app being there. So maybe that’s like Google magic or something that when you have an app they will promote you more in the web search 2 or I’m not

[29:25] Travis: sure, right, in their search engine optimization or something.

[29:28] Benjamin Teuber: Yeah, something. But the point is that ever since we uploaded the app to the Play Store, we’re having like 50 new additional registrations every day compared to what we had before. And they all seem to be beginners, right. If I look into some of their games or what they’re doing, they’re playing with some 15-kyu bot and just clicking randomly and not understanding the rules. Also since then I like got a couple of mails already. Yeah, that’s a nice website, but how do you play the game? Interesting. I never seen a message like that before. Right. So it really must be related. So we thought, okay, somehow we seem to have a lot of beginners coming to the website so we should do something for them. And so I started working on a Go tutorial and it’s also a bit of interactive hands on. So we try to have not so much text and really just. Just have problems to have them experience the thing themselves. Kind of throwing them into the water and.

[30:29] Benjamin Teuber: But I like it a lot. It has like eight chapters and each has like eight to ten steps with problems and a few explanations. So that’s there now for new players I can recommend that. I guess we kind of want to follow up on that. And because now they, if they know how to play the game and are like 25-kyu then of course they’re not ready to play a 15-kyu bot on 19 by 19 and just take the move recommendations without further explanations. So I guess we should do something that we have some ranked 9 by 9 where they can level up from 25-kyu to let’s say 15-kyu. And maybe we should also try to work on more explanations like at least some patterns, right? That when you make a move that makes your group die, we can write a sentence, we can detect that. I mean for now we do have something. We have some sad smileys if you’re dying.

[31:25] Travis: Yes, very helpful.

[31:27] Benjamin Teuber: But I guess we could just put some text and say okay, now this group dies and this move would save them instead. Or we could teach them some basic shapes, right? Like if Something is an empty triangle, or if a good move is this and that shape, we could explain that. So I guess for me, the next important thing is to have more basic explanations for weak players that they can still somewhat study by themselves, even though I still recommend a human teacher.

[31:58] Travis: In addition, even for somebody like myself, where I’m lower ranked but I’m still not a beginner, it’s been really nice to have some of that functionality around the visualization. So the smiley faces and the happy faces and the sad faces, as basic as that is, it really helps because sometimes I’m not completely aware of what’s happening to the strength or weakness of a group that’s on the board. And particularly the loss functionality is also really nice. Sometimes I’m not aware that a group on the board has actually just gotten very weak or is close to being lost. And so that’s really helpful compared to using the online Go server or the OGS board. It does show you some sort of visualization of what areas you have, but that concept of what you might be losing was really helpful. When you use the AI review portion of the site, which is really 99% of what AI Sensei does so well, you have that nice bar where you can change how many mistakes you see in the game.

[33:01] Travis: And so you can slide that dial from being beginner all the way to being a high dan player. And it seems to correlate to how big of a point loss in moves there are in the game. How did you think about tying those point loss moves to a certain ranking? Was that something that had to be thought out very carefully or was it something where you just said, hey, we have to have some way to start somewhere and end somewhere. And so we’ll say that if you’re a high dan-level player, you should only really be having a 1 to 2 point loss or suboptimal move from the AI?

[33:45] Benjamin Teuber: I guess something in between. Right. In the first version we just had a slider that said where you could set the point threshold yourself actually, but that was not really instructive for people. Right. I’ve seen a lot of double-digit kyu players setting that slider to like one point or half a point or, or whatever because they just wanted to know. And so we thought it might be more helpful to make this more abstract and also give some concrete recommendations. Yeah, of course we try to figure out what makes sense by looking at sample games from all, all the ranks. But also you should take this with a grain of salt because after all, it’s just Some arbitrary thresholds and point loss isn’t really a good measure of strength. I mean, it can help, but it’s possible that a strong professional makes a misread in a very complicated life-and-death situation and loses 50 points by a move. Right? It happens. And so this is not a 30-kyu mistake,

[34:50] Benjamin Teuber: even though he did it. And maybe you can make a very basic shape mistake that usually a 10-kyu should know and only lose two points with it. And yeah, so this is not perfect and we’re thinking of better ways to actually do that. You said the analysis is 99%, but actually it’s more like 80%, I would say, because a lot of people also just enjoy playing against the bots on the website too. I mean, I guess they also analyze most of the games, but maybe not all. So about of the 1/3 of the analysis we have is like the number of games we have with this AI. So it’s also quite a big driver nowadays because we also were the first to come up with these human like bots, we called it. That was actually fun because it was my first own artificial intelligence research project, so to say. Because with KataGo or also before with Leela, right. I’m just building on top of others, right? And they do the smart AI and I just build the interface and the cloud and the database and everything around it.

[36:00] Benjamin Teuber: But yeah, we had this idea. Oh, we also had the demand from users, like they want to have bots that are more natural to play against and not just this superhuman KataGo that you need a lot of handicapped stones to get started playing with. So we actually started trying to train a neural network on 5-kyu Games of users. And I think we found some smart shortcut that made them much more realistic. Because we kind of used KataGo too, right? Because if you train yourself just the visual patterns, it’s really hard that they still understand the whole board situation. Sure, we know that KataGo already has that understanding. So what we did was kind of plug these networks to the output of KataGo. The knowledge of the go thing was kind of baked into the input. So we could get away with fairly small networks that we could train on average hardware in a very short period of time. And people like them a lot. So like this is one of the feature that gets the most praise nowadays actually.

[37:07] Benjamin Teuber: So now KataGo, I think also has this human like thing. So I think that’s one thing they actually copied from us and not the other way around. That’s great. And I think they’re doing it much more Sophisticated with way more games and be more independent of the other KataGo thing. But yeah, so this thing not only does give nice bots, but you could also imagine using that for the review, right? Because you could say this mistake is very likely for 5-kyu bots, that it would make it 2. So maybe it’s fine if you make that as an 8-kyu, but this mistake is even avoidable for a 15-kyu bot. So maybe you shouldn’t make that one as a 5-kyu. And so we could put more emphasis on that one.

[37:55] Travis: So there’s almost a filtering back and forth, right? There’s a filtering of, okay, let’s train the AI bot on different levels of rank of players of human players so that they become more human. Like which is something that you pioneered, which is really cool. And then there’s when a human reviews a game, whether it’s a human game or not, this could be kind of used to inform, oh, that move you played is similar to this level of human bot or human like bot. That’s a great concept.

[38:27] Benjamin Teuber: Could also be used for suggestions, I guess. Right? Like if the AI wants to play the super crazy hard to understand move, but actually a very easy move that a human could find is only 0.1 points inferior on average, then maybe that’s the move we should recommend to the human because it’s the understandable clear move that’s also about as good as the perfect AI move, if you can even call it that.

[38:54] Travis: That does feel like a critical moment for users that are not dan-level players. Because I know for myself as a double-digit kyu single digit kyu player that there’s a lot of moves that the AI plays where I say I simply don’t understand what it’s trying to accomplish for the points you just mentioned. So that would be really powerful. I notice when I upload some of my games, it’s interesting sometimes to your point about, hey, you’re always trying to find a way to make adjustments to the system, but sometimes I’ll find that I’ll play a game and if I set the dial for how many mistakes there were, there might be not many huge mistakes. Right? There might not be a 10 point loss move or a 15 point loss move, but there are just a lot, a lot, a lot of four point loss moves. And so maybe if I set my dial to the same spot every time, that game might not show me a lot of mistakes. So I’d have to increase the dial to a higher ranking. How do you think about that?

[40:03] Benjamin Teuber: Yeah, I mean, I guess it’s not so easy, Right? Of course, we could just say no matter how the game went, we just show the top five mistakes.

[40:13] Travis: Right. No matter what. Right.

[40:16] Benjamin Teuber: But it also doesn’t feel right to me in a way. Also, I think it’s. It can be very motivating. Right? You upload your game and you set your slider to 5-kyu and there was not a single mistake you did. Right.

[40:30] Travis: And you can be really. That’s true.

[40:31] Benjamin Teuber: And you should be. Yeah, that’s a really good point. You could still say, okay, but if I want to be shodan, how are my mistakes then? And what can I still improve from that game? But of course, the other way around, it’s a bit annoying, right? If there’s like 50 mistakes or there’s also this kind of thing where both players just miss to play in one corner, right? Yes, yes. Ling is all over the game. And so every move is a mistake. Same. And it’s a bit ugly. But I’m not sure how to solve it. Because just pretending that they’re not mistakes, although they objectively are, feels wrong as well. And I’m not really sure what’s the best way. Maybe somebody can come up with a smart suggestion.

[41:16] Travis: It’s a good point. I think you already sort of bring up some of the natural thought processes from that. Like you said, if you’re playing that game and you said, wow, there’s no bigger double digit point loss moves, maybe I should infer from that things actually went pretty well that game and the opponent was playing such moves that I only had smaller losses. So that’s a good thing. Do you think that it’s ever possible? I don’t know how to think about this. The AI finds a move in the game and says there’s a 15 point loss. Right? Sometimes I do find that a human Go teacher will point to a move that might not have a dramatic point loss in that one move. Can you relate to that as sort of the creator of AI Sensei, but also as a Go teacher, that sometimes the move you want a player to look at isn’t always the, you know, the 15 point loss move. It might be something before that.

[42:15] Benjamin Teuber: Yeah, I think that makes a lot of sense. Because sometimes you just have the wrong idea, right? You think this area is important and you neglect to defend in the other spot. That would be urgent. And so maybe the first move you played away was still not objectively losing points because it was still kind of forcing to answer or something. But in a way, that’s the moment where you made the logical decision to play for that area and then the rest, the actual mistake is just a logical continuation. It only makes sense to follow if you had the original idea. And so in a way, sometimes the first move shows the bad idea, but actually the AI says no. If you just abandon your completely and change your mind and just throw away the stone you just played, then you still just lost a point or so. But nobody will ever do that because they wouldn’t have played that move if they were to change their mind about it. Right, Got it.

[43:13] Benjamin Teuber: But yeah, that’s also very hard to automate in a way. So I think, hey, of course now something that we should look out as human teachers instead.

[43:22] Travis: Maybe before we end our time here together. What final thoughts do you have about either AI Sensei or about your personal journey in go? Anything else that you would like to leave us with?

[43:35] Benjamin Teuber: Yeah, I think it would be that we already talked about this over. Focus on ranking and rating. And I think for a couple of professionals, winning might be important for their income, right? So maybe it’s understandable that they put so much pressure on themselves, but for everybody else, we’re amateurs and this is a game. And I think the main focus should be on enjoying the game. Of course, it’s also about learning, because this is like a fascinating universe of things, but I think it’s easier said than done. But whenever you lose a game, actually you should be glad for an opportunity to learn more and understand more about the game and try to enjoy that instead of just focusing on improving and on winning. Of course, during the game you should try to win, right? Because all this game complexity and beauty just unfolds and both people try to win. But after the game, you should just enjoy being involved in this beautiful game and learning more about it and continuing your journey.

[44:40] Benjamin Teuber: And in a way, if you try to learn as much as you can, and if you enjoy learning, then it will just happen that you become stronger and then you will win more. So you won’t lose out on that, but you still shouldn’t focus on it. I think that’s the wrong attitude.

[44:58] Travis: Benjamin Teuber, thank you so much for your time. I really appreciate it and it’s been fascinating learning about your story in AI Sensei. So thank you again.

[45:06] Benjamin Teuber: Yeah, thank you. It was a pleasure.

Sponsor and Support Message

[45:30] Travis: Hey everybody. Just taking a quick pause to thank our individual contributors. It really helps, so I really appreciate it. Also, I want to talk about the sponsors for this episode. SmartGo One is our first sponsor. They’re your complete one stop shop for playing the game of Go, practicing, learning from the professional games that they have in their massive database, and also reading Go books that have been digitized and are much more functional and convenient to use. SmartGo One is available in the App Store on iPhone, iPad, Mac and the Go books are actually available on EPUB file as well so they can be used on other devices. Go Magic is also a sponsor of this episode. They are a modern way to study Go like never before with an online platform to learn the game of Go, study different parts of your game, get tutorials from some of the best teachers out there. You can go to gomagic.org to learn more. The final sponsor of our episode is BadukPop.

[46:25] Travis: This is an app for Go players of any level to practice Go problems, do interactive Go lessons and play against other players as well as play against a variety of AI opponents with different styles and strengths. You can download the app at your App Store or go to badukpop.com that’s b a d u k p o p dot com with that we will go on to the rest of the episode. Shifting gears a little bit from the conversation around AI, I got to talk with Gazza White again and ask him about one of the top Chinese professional players, Ding Hao, because he was actually in a documentary that I had come across where he was talking about the role of AI in his training but also in his play and how he thought about who were the best players in the world based on how they used AI. So Gazza is going to give us some thoughts about that. Let’s take a listen.

Ding Hao, AI, and Professional Style with Gazza White

[47:24] Travis: There was A I think 15 minute documentary clip that I saw on online about Ding Hao, the Chinese player. I don’t know if you had a chance to listen to it, but he’s 24 years old, he’s a 9-dan professional and I think he’s in the top five or so in the world rankings, maybe top 10. You would know better than I would. But he had some really interesting answers to his thoughts on like AI. He trains with AI, but that he also thinks that the newer generation of top pros like what you’ve mentioned before Gazza, that it’s created a somewhat homogeneous style of play and that he feels like to then really rise and be the tip tip top. You actually need to have in his words or in the translation, some sense of style or something a little bit different to really capitalize. Otherwise everybody is really good but they’re very, very close in ranking to each other. And I think you referenced this before about Chinese players as well, can you speak to any of that or do you have reactions to that?

[48:31] Gazza White: I haven’t seen the documentary on Ding Hao but the sentiment that AI has introduced, sort of a cause players to have a similar style to each other is definitely one that’s been around, definitely one that I agree with and it has been around for quite a few years. It is worth noting that apart from Shin Jin-seo there hasn’t really been many players in the past five years that are consistently putting out outstanding results at international level. Maybe Ding Hao is the most notable. He won two tournaments in 2023 and he won a tournament in 2024. But in the years prior to 2023 he was struggling just to even qualify for international tournament. And you see other players like Byun Sang-il. He was in the LG Cup this year. He also won the Chunlan Cup, but other than that, he’ll sometimes lose in the second round of tournaments. Another example is Li Xuanhao who was on a big rise in I think 2022. He made the final of the Chunlan Cup and then he ended up winning the MLily Cup. But recently he hasn’t continued that form.

[49:53] Gazza White: So it does seem like it’s almost 50, 50 sometimes as to whether a player will win or lose against, you know, another player that’s learning under AI. Like all of these Chinese players, it seems to be a revolving door as to who gets picked for the Nongshim Cup for instance. But for all that, we’ve still got Shin Jin-seo who is well known to study under AI. And yet he still a clear step above everyone else in the world. So there must be something that he’s doing that’s causing him to stand out among others. It’s definitely hard to say exactly the reason behind Shin Jin-seo’s dominance or what players like Ding Hao need to do to stand out from everyone else. Yeah, I’m really not sure.

Losing Streaks and Winning with TelegraphGo

[50:52] Travis: So we all have them losing streaks. I wanted to ask TelegraphGo what he thought about losing streaks because I know he’s given it a lot of thought and also how important winning is. We’re going to just jump right on in.

[51:10] TelegraphGo: So I actually went on a very massive losing streak recently. One of my forays on Fox 9-dan ended in about 20 straight losses or something like that. And I would say that was largely due to my mental state aside from Go. Even if I am meditating before my game and completely locking in and getting ready. If you don’t have the right sort of patient but active, strong but Careful, optimistic, but not overly optimistic. Really precise mindset that you can really only get into when you have a sort of positive attitude yourself. The Go will naturally follow that. Like if something goes wrong, it’ll hurt more if you yourself aren’t happy. So I think there’s a good reason why the top pros are generally taking care of themselves in good shape. And if they are, you know, not doing so well, then mentally, then they describe that they don’t do so well on the board. And it’s the same for me and the same for many people.

[52:10] TelegraphGo: So I think honestly, if you get in a losing streak, it’s more important to take care of yourself in real life and let Go be a boon for you. So a lot of the time I go to tsumego to love Go without any pressure of actually being good at it. Any way which you can do that is very successful. I mean, a 20 game losing streak, obviously there’s something wrong if you’re just on a short losing streak a lot of the time. It’s also okay to keep playing through it, not to play through it all in one day. If you go on the same streak on one day, then maybe the next day it’ll be better, you’ll be in a better mindset. But yeah, I mean, this is a game so it’s supposed to be fun. A lot of the time the losing streaks come from a mindset where you weren’t making it very fun. And then when you do make it fun, you can win again. You can definitely enjoy a game very much that you lose. At least I can.

[53:04] TelegraphGo: And you can also not enjoy a game that you win. I’m sure there are games that come to mind for most Go players when you think about something like that, like, yeah, I won, but oh my God, the captain baiting oh my God, like,

[53:17] Travis: or it was such a letdown, you’re like, oh, I was hoping for a tight battle and then it doesn’t happen and you’re just like, ah, suddenly you

[53:24] TelegraphGo: win on time in a position which you were just about to struggle for your victory or yes. So yeah, I don’t think winning is that important because of those things. Still, you can aim to win and the more that you extract a joy from the journey because you’re really looking forward to that end goal of winning, the better. But the main point is still the journey for sure, the enjoyment of the journey along the route to winning or not winning. What I’ve realized recently is that being part of the difficulty of Go is actually a good thing. To respect and understand and use as your weapon as well to try to win the game that you can purposefully try to play a very long game with many opportunities for the opponent to make a mistake as your strategy, as your style that you extend your fights so that you prefer situations where your opponent has a weak group which is unsettled and will not be settled for the next 50 moves, even if it’s not that much weaker than a,

[54:27] TelegraphGo: you know, another situation where they could say, well, let’s Keeping this sort of latent pressure over many, many, many moves just expands the chance that you will score. You can find a chance to gain. I guess it’s just part of style that some people will prefer to take situations where they have opponent weaknesses and immediately create a master stroke. But another style is to wait for the master stroke and wait and wait and wait. And to set up opportunities where you can just continue to wait and wait and wait. This sort of extended pressure is a style in itself.

[55:05] Travis: That about covers it for the episode. Thank you so much and catch you next time. This is All Things Go,

[55:18] TelegraphGo: Sam.