Fujisawa on AI

[John Fairbairn]

I have remarked several times before, how little AI is mentioned in pro talk, even though they obviously use it. When they do mentioned, therefore, I pay special attention. One such case is a recent interview with Fujisawa Kazunari 8-dan. His main role nowadays is as a teacher, and he is one of the most successful. Apart from his daughter Rina, he has taught her rival Ueno Asami and her sister Risa, and his protege Hirose Yuichi, just turned 7-dan has been making his mark in the China Weiqi League. Motoki Katsuya and Seki Kotaro have done enough better domestically. Seki, of course, has won the Tengen.

So when Fujisawa says, "Jump" we say, as the interviewer did, "What do you think of AI?" as a way of jumping higher.

Fujisawa said his father (the legendary Hideyuki, also misknown as Shuko) used to teach him that there is no need to learn joseki, and he has always passed that advice on to his own pupils. However, AI research has reached a new depth through which it is possible to be at a disadvantage in fuseki. Therefore he now also advises pupils to look for simple plans and ways of avoiding new-fangled variations.

In other words, he is not specially encouraging them to go down the AI rabbit hole themselves. They reason is that, while that may seem fine for young people who have great reading skills and can concentrate intensely, "they stop thinking for themselves. They may feel like they have been studying but in fact their thinking is slowing down." There is also just too much to memorise, it seems.

The interviewer remarked that they have invested no flesh and blood in the process. Fujisawa agreed, adding that AI is great to use but hard to use correctly, and if you get it wrong to are in trouble. He therefore says that he tells the children in his school to watch NHK games (which have now introduced AI charts) to watch for the first time by hiding the winning percentages and candidate moves.

That's what he did say, and it's interesting enough. But what he didn't say is also interesting. He made no mention of the middle-game or endgame, or even tactics in general. Which is the case in general, I might add. AI is just an ideas bank, it seems, for most pros. They are assumed to have optimal reading and counting skills anyway.

In a separate part of the interview, he revealed that he doesn't give much guidance to Rina any more. He joked that they now just meet for dinner sometimes, because whenever they do meet he nags her about when she is getting married and she doesn't like that! She's 24, which in Japan is traditionally the sort of age where parents start panicking. (Kazunari is 58 BTW.)

[RobertJasiek]

"He made no mention of the middle-game or endgame, or even tactics in general. Which is the case in general"

Is not mentioning them the general / usual case only for Fujisawa or professional players?

[Elom0]

I have remarked several times before, how little AI is mentioned in pro talk, even though they obviously use it. When they do mentioned, therefore, I pay special attention. One such case is a recent interview with Fujisawa Kazunari 8-dan. His main role nowadays is as a teacher, and he is one of the most successful. Apart from his daughter Rina, he has taught her rival Ueno Asami and her sister Risa, and his protege Hirose Yuichi, just turned 7-dan has been making his mark in the China Weiqi League. Motoki Katsuya and Seki Kotaro have done enough better domestically. Seki, of course, has won the Tengen.

So when Fujisawa says, "Jump" we say, as the interviewer did, "What do you think of AI?" as a way of jumping higher.

Wow! I had no idea Hirose Yuichi is doing well in the China League, it must be in the second division

Fujisawa said his father (the legendary Hideyuki, also misknown as Shuko) used to teach him that there is no need to learn joseki, and he has always passed that advice on to his own pupils. However, AI research has reached a new depth through which it is possible to be at a disadvantage in fuseki. Therefore he now also advises pupils to look for simple plans and ways of avoiding new-fangled variations.

In other words, he is not specially encouraging them to go down the AI rabbit hole themselves. They reason is that, while that may seem fine for young people who have great reading skills and can concentrate intensely, "they stop thinking for themselves. They may feel like they have been studying but in fact their thinking is slowing down." There is also just too much to memorise, it seems.

This is very poignant point for me, I had been wondering indeed whether AI has now made the margin of error in the opening smaller. To get around the concept of joseki, however, I'd rather think in fuseki terms!

The interviewer remarked that they have invested no flesh and blood in the process. Fujisawa agreed, adding that AI is great to use but hard to use correctly, and if you get it wrong to are in trouble. He therefore says that he tells the children in his school to watch NHK games (which have now introduced AI charts) to watch for the first time by hiding the winning percentages and candidate moves.

That's what he did say, and it's interesting enough. But what he didn't say is also interesting. He made no mention of the middle-game or endgame, or even tactics in general. Which is the case in general, I might add. AI is just an ideas bank, it seems, for most pros. They are assumed to have optimal reading and counting skills anyway.

In a separate part of the interview, he revealed that he doesn't give much guidance to Rina any more. He joked that they now just meet for dinner sometimes, because whenever they do meet he nags her about when she is getting married and she doesn't like that! She's 24, which in Japan is traditionally the sort of age where parents start panicking. (Kazunari is 58 BTW.)

The idea of an optimal reading skill seems strange to me. Surely reading is always lacking to some extent. I would say that if you study AI ideas in the opening, it's suggestion in the middle game are simply an extension of the same intuition so is not really necessary to study it in depth in later parts of the game since the middle game is just reading guided by intuition. Perhaps in that sense,go is like a person with a personality we must study. With people it's really good to just understand in detail how they react in a specific situation, and use that as a DNA strand to predict how they will react to other situations. Cho Hyeyeon in her late 20s described herself as being married to the Baduk board, having never had a boyfriend. She only got married in her mind thirties, which is normal these days and is likely usually the most logical time. She's doing well, and now I myself who was a fanboy of her from her live journal I read around 2015 discovered her YouTube channel last year, were they seem to be doing great as a couple, so from what I've seen, Fujisawa's dad need not be worried!

[RobertJasiek]

the middle game is just reading guided by intuition

Calling for discussion? :) The middle game is a combination of all skill and knowledge, where analysis should override subconscious thinking.

[gowan]

I understand the advice to avoid the complicated "AI Joseki". Something like how pros teaching back in the 1970's would advise weaker players to avoid complicated variations in the large avalanche and taisha joseki. Speaking of the avalanche joseki and the taisha, they seem to have disappeared from play.

[RobertJasiek]

Usually if an avalanche starts, I choose a simple territory vs. thickness / influence variation. Taisha can also be driven to simplicity.

[Kirby]

Fujisawa's point about the importance of thinking is intuitive. Thinking for oneself is clearly essential for the same reason that memorizing a joseki book as a beginner won't make you 9-dan. Training one's ability to think for oneself is probably one of the most important areas of study.

However, inspiration from acquired knowledge is helpful for learners. A beginning player who has seen a few professional games, for example, might recognize that people often start playing on the 4-4 and/or 3-4 points. They might, consciously or subconsciously, try it out in their own games. Through game experience and thinking on their own, they develop their own intuition about what does and doesn't work. They might be able to develop the same intuition through pure trial and error, but it'll take a lot more time.

AI can be used in the same way. We can see examples from AI and get inspiration. We might even get new feelings about what constitutes good or bad. Case in point: early 3-3 invasions can feel a lot better as the invading player, now that we've seen them used in high-level play.

This is not to say that it's a good idea to stop thinking.

Rather, it's this combination of: thinking for oneself + getting ideas from stronger entities (AI included) + trying out new ideas + reviewing how that went, which lets us improve. Thinking for oneself is a key ingredient in this recipe, but inspiration from stronger sources is something we all do to some degree - and it's helpful for learning.

[pajaro]

Whenever I read a discussion about AI, I remember Redmond's 60 videos commenting Master's victory over top players. Undefeated, Master was not Sai but a version of Alpha Go, more advanced than the one that beat Lee Sedol for the first time.

Of course, in many cases Master played a better sequence than the human, or a new variation of a joseki that was still unknown. The AI would get an early lead and keep it to the end. Now, we know that it's how things usually go.

But in many cases, Redmond said "we used to think this was bad, but Master does this and gets a good position". Or the opposite. We used to think this was good, and...

AI has changed the game, of course. It has brought many new ideas to the fuseki, with new joseki and making others obsolete. But for the middle game, I think you can't just memorize moves and use them. That's completely unrealistic, because the same situation is not going to happen again. So, what's the proper use of AI then? I think it's useful to review your own games, find situations where a certain tesuji could have played, or when you missed the direction of play. This all adds to the player's intuition. But you can't stop thinking by yourself. You still have to read and calculate. The difference, maybe, is that you can consider moves that you would reject (because the former experience says it doesn't work), or you can get new ideas.

But in the end, during a game, you can use only the resources that you have inside.

[John Fairbairn]

AI has changed the game, of course. It has brought many new ideas to the fuseki

Trying to simplify this, I see the process as follows. The standard fuseki advice in the golden age of human go was "corners, sides, centre. AI has changed this into "corners, corners, corners."

If you dig a little deeper, you could maybe argue for "corners, corners, centre" but as the centre moves are really to do with the corners (see below), "corners, corners, corners" serves better.

However, to get full value out of this, you need to redefine your concept of corners a little. I think this re-definition comes under three headings:

1. You have to think of a corner nowadays as a place not just IN which but also FROM which or AROUND which you can make moves. The old CSC principle simply focused on making moves IN a corner. The new principle is CCC + IFA.

2. You have to forget the idea of thickness versus profit. It leads to wrong evaluations. Also, if you think of thickness as building a castle, it's too easy, as AI shows, to make it end up as being in the wrong place.

3. Corners nowadays extend into the centre. I covered this idea on L19 under the concept of barmkins, which I won't repeat here except to say that modern AI fuseki puts a lot of stress on the diagonals connecting corners, and the area in the centre just outside a corner (which is the fortified keep or pele tower) is the loosely defended barmkin where you keep your cattle and sheep. AI moves such as shoulder hits put a lot of focus on these areas.

The attached AI self-play game seems typical of AI fuseki to me. In the first 50 moves, the number of plays on the side which can be said to be more or less independent of the corner is always very low - here two, I'd say. (And even when they exist, they seem to be more often on the fourth line, not the third.) In contrast, the number of plays on the fifth line or higher can be very high - 20% as here is not unusual. These ultra-high plays seem to be about keeping diagonally-opposite corners in contact, if possible (i.e. via the barmkins - here in the SE and NW corners for White). This can be seen as a modern version of the old Chinese zhaoying (call & response) moves, though these tend to be more focused on the middle game than the fuseki. AI plays them throughout the game.

Modern AI-influenced play by pros tends to look rather different from AI self-play. Humans tend to prefer moves they think they understand better, and so side moves are more common there. Since AI often labels these moves as only trivial mistakes, pros are tempted to play them. However, I'd hazard a guess that such moves may make the fuseki easier for humans, but they also make the later part of the game (i.e. the play in the centre) harder for them. In other words, an AI bot that first builds its pele towers and barmkins properly, on the right high ground, is going to have a much easier time coping when the reivers arrive.

[Ferran]

But in the end, during a game, you can use only the resources that you have inside.

Hmm... I know a guy with a scar on his cranium about 8" long... Do you think his surgeon read 'The peace war'(*)?

Take care.

(*) Vernor Vinge. Published in Spanish a couple decades ago. One of the sports is AI-assisted chess.

[CDavis7M]

He therefore says that he tells the children in his school to watch NHK games (which have now introduced AI charts) to watch for the first time by hiding the winning percentages and candidate moves.

This is an interesting way to study. And while I don't know if I would benefit from trying to analyze NHK games at their level of play, I could adapt this for my studies of my own games. I often just hop straight to the AI to see the failures but maybe I should find my failures by myself first.

[Elom0]

AI has changed the game, of course. It has brought many new ideas to the fuseki

Trying to simplify this, I see the process as follows. The standard fuseki advice in the golden age of human go was "corners, sides, centre. AI has changed this into "corners, corners, corners."

If you dig a little deeper, you could maybe argue for "corners, corners, centre" but as the centre moves are really to do with the corners (see below), "corners, corners, corners" serves better.

However, to get full value out of this, you need to redefine your concept of corners a little. I think this re-definition comes under three headings:

1. You have to think of a corner nowadays as a place not just IN which but also FROM which or AROUND which you can make moves. The old CSC principle simply focused on making moves IN a corner. The new principle is CCC + IFA.

2. You have to forget the idea of thickness versus profit. It leads to wrong evaluations. Also, if you think of thickness as building a castle, it's too easy, as AI shows, to make it end up as being in the wrong place.

3. Corners nowadays extend into the centre. I covered this idea on L19 under the concept of barmkins, which I won't repeat here except to say that modern AI fuseki puts a lot of stress on the diagonals connecting corners, and the area in the centre just outside a corner (which is the fortified keep or pele tower) is the loosely defended barmkin where you keep your cattle and sheep. AI moves such as shoulder hits put a lot of focus on these areas.

The attached AI self-play game seems typical of AI fuseki to me. In the first 50 moves, the number of plays on the side which can be said to be more or less independent of the corner is always very low - here two, I'd say. (And even when they exist, they seem to be more often on the fourth line, not the third.) In contrast, the number of plays on the fifth line or higher can be very high - 20% as here is not unusual. These ultra-high plays seem to be about keeping diagonally-opposite corners in contact, if possible (i.e. via the barmkins - here in the SE and NW corners for White). This can be seen as a modern version of the old Chinese zhaoying (call & response) moves, though these tend to be more focused on the middle game than the fuseki. AI plays them throughout the game.

Modern AI-influenced play by pros tends to look rather different from AI self-play. Humans tend to prefer moves they think they understand better, and so side moves are more common there. Since AI often labels these moves as only trivial mistakes, pros are tempted to play them. However, I'd hazard a guess that such moves may make the fuseki easier for humans, but they also make the later part of the game (i.e. the play in the centre) harder for them. In other words, an AI bot that first builds its pele towers and barmkins properly, on the right high ground, is going to have a much easier time coping when the reivers arrive.

At the risk of commiting treason to my species, I must say this new AI way of seeing things to me is just far more intuitive than the old way. I still remember when AI justified moving into the 3-3 early which always seemed fine to me but not so by pros, so I didn't play it. Perhaps humans have an urge to have to 'do something' with their corner by extanding to the side.