Introducing SensAI Go

[John Fairbairn]

At the risk of distracting from the main point, this Western kyu-level player has learned that the tortoise-shell shape formed by capturing at B is incredibly strong ("the tortoise shell is worth 60 points" is in John Power's proverbs book) and my eye is drawn to it even more than a ponnuki.

It does distract from the main point, I think - or at least one of MY main points. Calling it a tortoise shell (maybe turtle shell is better?) is focusing on the shape - a major western fault, in my view. Go Seigen focuses on the pon effect produced by the nuki - dynamic over static, suji over katachi. You seem also to be trying to drag the conversation back to the shape aspects by making a distinction between a tortoise shell and a ponnuki. How many more times do I have to say that the move that produces a tortoise shell IS a ponnuki. It's a nuki (capture) that creates the effect pon (plop, wow, etc). The biggest ponnuki I have seen mentioned as such (I think it was an example by Kitani) was a 9-stone capture.

The tortoise shell proverb certainly exists in Japanese but I don't think I've ever seen it outside of a proverbs book or a dictionary of terms. Commentaries use the "X-stone ponnuki" version. To stress the point, Hayashi Yutaka's go encyclopaedia describes the tortoise shell as follows: "A shape produced by a ponnuki capture of two stones in the form of a tortoise shell." He also adds that if there are extra stones of the capturing side at the corners, "we do not call it a kame no kou." (Hayashi was the editor of Kido among other things.)

The Japanese proverb is usually in the form "ponnuki wa 30 moku, kame no kou 60 moku" which may seem as if ponnuki is being contrasted with kame no kou. But I think you'll find that the effect is additive, not contrastive (i.e. the full meaning is "a single ponnuki is worth 30 points but if you can capture two stones in a ponnuki it is worth 60 points" - there's no contrastive "wa" before 60).

[John Fairbairn]

I have been doing the tabula rasa approach to new vocabulary. I do not see how AI (pure moves and their statistics) provides any vocabulary - please suggest a related approach!

I have not been doing any work on this myself directly, but because of my own interests and professional background I analyse language as a matter of course. I have therefore noticed certain things. I'll mention some of them, but in a higgledy-piggledy fashion in the hope of starting off a brainstorming exercise. I'll number the points, but that is not to imply any kind of order. It's just so that other people can refer back to the individual points if they want to.

1. I have been struck by how similar old Chinese go is to AI go. It has often been remarked that Go Seigen's go looks very like AI go. He was heavily influenced by old Chinese go. There may be a connection.

2. Old Chinese go was heavily influenced by group tax, which has far more impact on ALL phases of the game than you might expect. Among these effects is an emphasis on group connections, even very early in the game. This leads to a great use of call & response (zhaoying) moves early on. It is not necessarily an immediate connection. I like to call them lighthouse moves - where the move stands out as a beacon offering comfort to ships (groups) that may fall into distress on the vast ocean of the board. I am 100% certain this is a feature of AI go.

3. Another feature of old Chinese go is that moyos are rare. This seems to be true also of AI go, too. I suspect the lighthouse moves are there to provide connectability and NOT to create moyo frameworks - this seems to be borne out by the near disappearance of the term "moyos" in commentaries on AI-inspired pro go.

4. There is great reliance on pressuring (jin) as opposed to attacking moves in old Chinese go. This seems matched in modern AI-inspired go where the Japanese are now increasingly referring to semaru instead of semeru.

5. Although I have not seen anyone else use the spider analogy in modern go (it does exist in old Chinese, though), it seems to me that AI go strongly resembles the formation of a spider's web. A spider starts by spinning a sticky thread and lets the wind blow it where it will until it sticks on an object some distance away. It then marches across this thread to add an extra, strengthening strand. This process can be repeated. It seems to me that AI go is similar. It is not specially concerned with bases but with connections across the vast expanse - the empyrean - of the centre of the board. I see the interference presented by the opponent as the effect of the wind. You have to be prepared to follow it - go with the flow.

6. One thing that follows from the web-making analogy is that you have to be flexible. I think that in commentaries this becomes most noticeable in the frequent use of trades (furikawari). I personally don't see this as a technical term - it can't really be defined, and so is more of a descriptive term. But trades do seem to be more frequent in AI-inspired go, and (if that is true) I can see that it would follow from the spider's web style of play.

7. When AI first impacted pro go, there was a lot of mention of AI creating overconcentration of the opponent's shapes early on. That seems to have disappeared. I infer that the reason is that pros have learned to avoid being overconcentrated even more than they did before, and since it rarely occurs in their games now, there is no need to use the term in commenting on their games. But I do think it is still a major element in amateur play, and so needs to be talked about, not just as something to avoid, but as something to do to the opponent. Since you can see it quite easily, it has very useful function as an evaluative term. More subtly, lack of it implies a positive evaluation, but that in itself is nothing new in pro go. I see nothing wrong with using the term "overconcentration" in English.

8. There seems to be great stress now on "efficiency" in commentaries but it somehow seems different. It's certainly not the old kind of tewari efficiency, yet it also seems to be more than avoiding overconcentration. As best as I can judge, it seems to be more to do with stones ending up in a good POSITION (nb position, not shape). And my gut feeling is that good position has a lot to do with stones acting as lighthouse stones (or, if you prefer, as parts of a web).

9. The word 'rhythm' seems to come up as lot. This may just be fashionable use of an English term in place of choushi, but it may also betoken a sense of a different feeling to the concept - something more related to creating a web in the centre.

10. Boundary play, even early in the game, seems to be a big feature of AI go. This mirrors the use of shou in old Chinese go. But it is seems tied in with pressurising moves (jin/semaru). The point is that if you simply pressurise instead of attacking to kill, you do not end up overstretching and leaving gaps behind in your position. With slower pressuring you form natural boundaries in a defect-free way. Of course, if the opponent blunders you can still kill him - your defect-free positions also make that a lot easier, too.

11. The 9-dan Nakane Naoyuki is currently playing a 12-game uchikomi match against Katago. The uchikomi kicks in after each game, so he started at even and lost, then switched to -B- (and lost). Presumably the next game will be at two stones. Interestingly, the time limits are set at 20 seconds a move. I'm guessing that he thinks AI is more likely to make mistakes at that rate, but it also means he is relying largely on his own intuition. But even more interesting is his commentary. It may just be journalistic flourishes but he consistently avoids strategic terms and instead uses similes and metaphors to get his points over. As examples that come to mind, he referred to moves acting like an air-freshener (wafting over the entire board, presumably - lighthouse moves, I'd say). Other moves he has acting in unison like a school choir. In different vein he refers to one move being like a firecracker, and another as a "semaru shocker" (using the English word), both times denoting his consternation at overlooking the AI move.

12. Going back to lighthouse moves, I was sensitised to looking for miai moves by writing my book on Shuei's games. He was famous for his use of miai, and this was often associated with his use of L-shapes facing the centre. I therefore now naturally look for such moves, and what I'm convinced I've seen a lot of in AI play is what I call triai and quadai (i.e. moves that leave 3 or 4 alternative follow-ups instead of just two. That would certainly tie in with lighthouse moves.

13. When I used to actually play go, there was one thing in particular that used to dismay me - and I know I'm far from alone in this among amateurs. It was how often I'd get to the end of the middle game and find that I had lots of open skirts or the like, and if I didn't have sente, the Barbary ape opponents would tease me with monkey jump after monkey jump, snatching all my doughnuts from my sticky fingers. This didn't seem to happen very often with pros, but I couldn't work out why. It happens even less in AI go and at last I think I can see why. AI seems to have a habit of making a lot of early moves on the second line (just like Go Seigen, I'd say) but not to make bases (bases don't seem to mean much to AI). AI does it so that it doesn't have to worry about making these moves later. If you want a term for this, I will grit my teeth and say "early doors." That is a a term I actually hate - it puts my teeth on edge in fact. I've never really understood it, and it's (in my mind) a Cockney or Southern phrase. It was certainly not used where I lived. Some people say it refers to the practice of men queuing up outside pubs just before opening time when licensing laws were strict about opening times - I did see such lines, very long ones, as a kid, but that may be folk etymology. But it may fit in go in another sense: locking the doors on the open sides ultra early.

14. I strongly suspect that the Five Lands theory of Huang Longshi, in which regions of the board are described by five of Sun Zi's nine categories of landscape an army might find itself in, will prove fertile for talking about AI go. You can check out Chapter 11 of the Art of War if that interests you.

Here are a couple of interesting points fron Nakane's latest game with Katago. Katago played 10 next at 'a'. Nakane described this as a probe. Not the first term that would come to mind in traditional go! This is something I have seen elsewhere.

$$B
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . 5 . . . . |
$$ | . . . 2 . . . . . , . . . 8 9 , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . a . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 6 . . |
$$ | . . . 4 . . . . . , . . . . 7 , . . . |
$$ | . . . . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|

Click Here To Show Diagram Code

[go]$$B
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . 5 . . . . |
$$ | . . . 2 . . . . . , . . . 8 9 , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . a . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 6 . . |
$$ | . . . 4 . . . . . , . . . . 7 , . . . |
$$ | . . . . . . . . . . . . . . . 3 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|[/go]

Later in the game, Nakane described White 1 (actually 58) as the semaru shocker. Again, semaru is not the first term that would come to mind traditionally, but it served the purpose of semaru (as opposed to semeru) in that White almost the entire lower-left quadrant as a result of pressure on Black's lower-centre group in return for sacrificing the entire group that included White 1 (a trade, in other words - large sacrifices and trades being a feature of AI go)

$$W
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | X O . . O O X . . . X . . . X . . . . |
$$ | . X O O O X X . X . X O . O X . . . . |
$$ | . X X O . O X X . O O . . O X , X . . |
$$ | . . X X O O O O O . . . . . . . . . . |
$$ | . . X O O . . . . . . . . . . . . . . |
$$ | . . . X X . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . X . 1 . . . . O . . |
$$ | . . . O . . O . . , . . . . X , X . . |
$$ | . . O . . X . O . X . O . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|

Click Here To Show Diagram Code

[go]$$W
$$ -----------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | X O . . O O X . . . X . . . X . . . . |
$$ | . X O O O X X . X . X O . O X . . . . |
$$ | . X X O . O X X . O O . . O X , X . . |
$$ | . . X X O O O O O . . . . . . . . . . |
$$ | . . X O O . . . . . . . . . . . . . . |
$$ | . . . X X . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . X . 1 . . . . O . . |
$$ | . . . O . . O . . , . . . . X , X . . |
$$ | . . O . . X . O . X . O . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ----------------------------------------|[/go]

[RobertJasiek]

Thank you for your ideas of some aspects worth looking for! Some of them might deserve terms while some others are better described as strategies. Part of these ideas I have not, or not fully, had so far. A reason might be that I have concentrated on the opening while your observations might be more relevant during the middle game. After half a year of AI opening study, I have just double-checked my related new go theory:

- 0 new terms from AI study. (My tabula rasa terminology of the previous 28 years has been sufficient so far.)

- 2 terms inferred from statistics and descriptions by AI programmers: empirical score (as shown in AI GUIs per move candidate); confidence (level) (interpreted due to the number of visits of a move candidate; for convenience, classified as levels for ranges of such numbers).

- 7 new principles of go theory.

- 4 of my principles of go theory got a higher relevance than I had assessed before. (I have not reevaluated extensively yet which of my many other principles and methods persist or need to go.)

[kvasir]

It is sometimes hard to tell what people mean when they talk about learning Go. To me that means improving at playing the game from start to finish and to the best of my knowledge you need to play games, from start to finish, without any external aid if you wish to improve at this. Everything else that you do is support for the main activity of playing games. However, there are other views.

I agree with your first statement: the goal is to improve at playing the game from start to finish without external aid. Let's call this "performance". The only way to know if you have improved your performance, is to perform more (play) and measure the outcome (review).
Your second statement is probably equivalent to my previous sentence but the categorization of everything else as tertiary activities suggests that all training which is NOT by means of performance is of a much lower importance. Here I'm not so sure.

I think you disagree. You are saying that you think playing games is performance and you appear to have interpreted what I wrote and then raised questions about, or even taken issues with, that interpretation. It is when you write "Your second statement is probably equivalent to my previous sentence..." that you appear to completely remove what I wrote from consideration. The thing is that playing games isn't only performance. Most of the time it is just practice or exercise. Surely you agree that there is a need to practice, to exercise and to repeat the mental process that goes into finding a move until it is second nature?

Then I'm not sure why you interpret what I wrote to mean that every activity other than playing is of much lower importance. Instead it means that the tertiary activity depends on the primary and secondary activity, if you don't do the first two there is little purpose in the the third. It is not that the third level is of low importance, it is that it supports the first two.

I do have a harder time to make similar constructions in Go. The common "isolated training" consists of joseki, life and death, and endgame. However, the impact on one's performance (winning games) of such training is not as strong as in the abovementioned skills. There are even proverbs saying (rote) learning joseki makes you weaker, suggesting the whole (board) always defies the sum of the parts.

I think there are two reasons for that:
1. Sometimes this kind of training strengthens skills that are rarely needed, sometimes you just won't be able to detect the change until much later.
2. It can be difficult during training to get into the same mindset that you are in when you need to use the skills.

Of course you can try to overcome these limitations, for example you could have better designed exercises or work on getting into the right mindset. Another way is to play games, one thing that Go has over many other activities is that is it rather easy to play and review.

Computer programs are of course great for exploring shapes and testing your reasoning about some positions, their main drawbacks is that they still tend to be fairly opinionated and they often prefer going down the rabbit hole over the simple solutions that often work better in practice. Their evaluation often isn't objective for human vs. human game and it gets worse the farther you are from their level of play.

http://www.neuralnetgoproblems.com is doing a good job at countering your argument.

Now, I don't know what argument that is or how it is countered. My limited experience with those problems is that they aren't that great. Does it really have any relevance to what I wrote?

[kvasir]

As far as I know such signaling of how an exercise is going isn't considered conductive for learning in general.

Do you have some source for that? I'm genuinely interested.

[...] remembering what you were doing must be crucial for improving on it, right?

That's exactly why I think training with immediate feedback is worth pursuing.
From my own limited experience using the system I can say that the instant feedback for some huge mistakes and my natural emotional reaction of "Damn, I just f***ed up" has really "burned" those shapes into my mind.
Of course, the review afterwards was still necessary to find out how I should've played.

It is what the education professionals in my Go club tell me. I don't think it needs much discussion that if you interrupt an exercise that it is likely to have a negative effect. As far as I see your suggestion it is interrupting the task, it is a distraction from the task, but with the nuance that there is some information being conveyed that is useful for completing the task. Yet the question is not if the task can be performed better, the question is if something is learned faster this way. I'm not sure what that something would be, is it the feedback itself?

In your video you show a game in progress on an online platform and some audio feedback at the same time. Now you talk about your experience playing while having such feedback. Who did you play?

[Knotwilg]

It is sometimes hard to tell what people mean when they talk about learning Go. To me that means improving at playing the game from start to finish and to the best of my knowledge you need to play games, from start to finish, without any external aid if you wish to improve at this. Everything else that you do is support for the main activity of playing games. However, there are other views.

I agree with your first statement: the goal is to improve at playing the game from start to finish without external aid. Let's call this "performance". The only way to know if you have improved your performance, is to perform more (play) and measure the outcome (review).
Your second statement is probably equivalent to my previous sentence but the categorization of everything else as tertiary activities suggests that all training which is NOT by means of performance is of a much lower importance. Here I'm not so sure.

I think you disagree. You are saying that you think playing games is performance and you appear to have interpreted what I wrote and then raised questions about, or even taken issues with, that interpretation. It is when you write "Your second statement is probably equivalent to my previous sentence..." that you appear to completely remove what I wrote from consideration. The thing is that playing games isn't only performance. Most of the time it is just practice or exercise. Surely you agree that there is a need to practice, to exercise and to repeat the mental process that goes into finding a move until it is second nature?

Then I'm not sure why you interpret what I wrote to mean that every activity other than playing is of much lower importance. Instead it means that the tertiary activity depends on the primary and secondary activity, if you don't do the first two there is little purpose in the the third. It is not that the third level is of low importance, it is that it supports the first two.

I do have a harder time to make similar constructions in Go. The common "isolated training" consists of joseki, life and death, and endgame. However, the impact on one's performance (winning games) of such training is not as strong as in the abovementioned skills. There are even proverbs saying (rote) learning joseki makes you weaker, suggesting the whole (board) always defies the sum of the parts.

I think there are two reasons for that:
1. Sometimes this kind of training strengthens skills that are rarely needed, sometimes you just won't be able to detect the change until much later.
2. It can be difficult during training to get into the same mindset that you are in when you need to use the skills.

Of course you can try to overcome these limitations, for example you could have better designed exercises or work on getting into the right mindset. Another way is to play games, one thing that Go has over many other activities is that is it rather easy to play and review.

Computer programs are of course great for exploring shapes and testing your reasoning about some positions, their main drawbacks is that they still tend to be fairly opinionated and they often prefer going down the rabbit hole over the simple solutions that often work better in practice. Their evaluation often isn't objective for human vs. human game and it gets worse the farther you are from their level of play.

http://www.neuralnetgoproblems.com is doing a good job at countering your argument.

Now, I don't know what argument that is or how it is countered. My limited experience with those problems is that they aren't that great. Does it really have any relevance to what I wrote?

Let me not try then to agree with or rephrase your statements, although I don't know how to claim relevance to your thoughts without referring to them.

1. I prefer distinguishing "performance = games" from "training = any other activity, which I think you labelled secondary or tertiary", even though every game is indeed a learning opportunity and every training triggers our desire to perform well. I claim that playing many games as away of improvement is inherently inefficient, because of the fuzzy learning you get out of each "repetition". Now I do understand that Go might be comparatively more complex in terms of skill acquisition than e.g. table tennis, where the skills involved can be more easily isolated and re-integrated, and the relative impact of each skill on the performance is more easily identified. E.g. serve & return are skills with great impact, since each point starts like that and most points are rather short, so serve & return take up a big proportion of the overall value. In Go we can't say the same about the opening moves, nor any aspect of the game to have such impact on overall performance. My best shots are "emotional balance", "time management" and "liberty awareness", since lack of these skills can make us easily throw away an otherwise good game, and all can be practiced on every move. OTOH, they mostly help us performing at our current level of understanding, but don't enhance it.

2. You said - if I got that right - that computers don't take into account the human capacity to absorb their advice. I offered https://neuralnetgoproblems.com/ as a good example of how AI combines its knowledge with the human learner's level to provide well targeted exercises and solutions. We seem to disagree - if I got that right - on the quality of that resource.

[pwaldron]

9. The word 'rhythm' seems to come up as lot. This may just be fashionable use of an English term in place of choushi, but it may also betoken a sense of a different feeling to the concept - something more related to creating a web in the centre.

Would rhythm be different than flow or development?

To my Western vocabulary, rhythm implies some level of dynamic development, something you have commented about in the past. I could image also overlap some with the haengma that was so popular years ago.

[John Fairbairn]

Would rhythm be different than flow or development?

As I indicated, I think we'd have to see inside Nakane's head to know his intentions. I suspect it might be equivalent to choushi, because that is also a musical term that denotes the beat. But the go usage is hardly musical, so I prefer the term 'momentum' for that. You might say choushi in go resembles conducting an ensemble from the keyboard.

I think development is an overloaded term, and flow is as perhaps as wishy-washy as it sounds. I usually reserve flow to talk about suji, but since I assert that haengma is suji + katachi, it would be wrong top call hanegma either 'flow' or 'rhythm.'

In short, I wot not.

[kvasir]

1. I prefer distinguishing "performance = games" from "training = any other activity, which I think you labelled secondary or tertiary", even though every game is indeed a learning opportunity and every training triggers our desire to perform well. I claim that playing many games as away of improvement is inherently inefficient, because of the fuzzy learning you get out of each "repetition". Now I do understand that Go might be comparatively more complex in terms of skill acquisition than e.g. table tennis, where the skills involved can be more easily isolated and re-integrated, and the relative impact of each skill on the performance is more easily identified. E.g. serve & return are skills with great impact, since each point starts like that and most points are rather short, so serve & return take up a big proportion of the overall value. In Go we can't say the same about the opening moves, nor any aspect of the game to have such impact on overall performance. My best shots are "emotional balance", "time management" and "liberty awareness", since lack of these skills can make us easily throw away an otherwise good game, and all can be practiced on every move. OTOH, they mostly help us performing at our current level of understanding, but don't enhance it.

You are of course leaving out practice and exercise when you say that you prefer to distinguish between performance and training. Playing games is practice, exercise and sometimes training. I didn’t mention that it was training, is this the issue, that playing games isn’t supposed to be good training? That seems nuanced or even pedantic, we do after all play games to improve, not only to practice what we can already do. You prefer to say games are “performance”, that seems like a mislabel to me, they are usually practice, exercise and sometimes training. Now, it doesn’t make sense to me to say activities such as reading a Go book are “training”, even if you read Go books with a “no pain, no gain” attitude; it is another mislabel from my perspective.

I’m not sure if you only wish to talk about training, since that is one of your two categories, I mean I talked about more general activities not only training. You gave some examples of skill acquisition in table tennis and Go. It strikes me that the skills that you identify for table tennis all involve fitness and coordination, and the skills that you identify for Go don’t. The table tennis skills that you mention are immersive tasks which totally encompass the athletes’ experience for the duration of that task. I think they are meant as examples of training. When it comes to Go you don’t mention tasks that I think should be immersive and I’m not sure why they should be comparable to the table tennis tasks. Yet, playing Go also involves a series of highly immersive tasks that require a high degree of mental fitness and coordination of different thoughts, not unlike how table tennis or any physical sport rely on physical fitness and coordination of the movement of the body. I don’t think it should be difficult to come up with examples of training for Go.

2. You said - if I got that right - that computers don't take into account the human capacity to absorb their advice. I offered https://neuralnetgoproblems.com/ as a good example of how AI combines its knowledge with the human learner's level to provide well targeted exercises and solutions. We seem to disagree - if I got that right - on the quality of that resource.

That is not at all what I said. I don’t know what the argument is about here, it is as if I write one thing and you then read something entirely different. Sure, I haven’t found the problems on that site to be great. I have read your praise of them, but even then, I don’t understand what the connection is. You don’t appear to take any issue with what I did write.

[Knotwilg]

1. I prefer distinguishing "performance = games" from "training = any other activity, which I think you labelled secondary or tertiary", even though every game is indeed a learning opportunity and every training triggers our desire to perform well. I claim that playing many games as away of improvement is inherently inefficient, because of the fuzzy learning you get out of each "repetition". Now I do understand that Go might be comparatively more complex in terms of skill acquisition than e.g. table tennis, where the skills involved can be more easily isolated and re-integrated, and the relative impact of each skill on the performance is more easily identified. E.g. serve & return are skills with great impact, since each point starts like that and most points are rather short, so serve & return take up a big proportion of the overall value. In Go we can't say the same about the opening moves, nor any aspect of the game to have such impact on overall performance. My best shots are "emotional balance", "time management" and "liberty awareness", since lack of these skills can make us easily throw away an otherwise good game, and all can be practiced on every move. OTOH, they mostly help us performing at our current level of understanding, but don't enhance it.

You are of course leaving out practice and exercise when you say that you prefer to distinguish between performance and training. Playing games is practice, exercise and sometimes training. I didn’t mention that it was training, is this the issue, that playing games isn’t supposed to be good training? That seems nuanced or even pedantic, we do after all play games to improve, not only to practice what we can already do. You prefer to say games are “performance”, that seems like a mislabel to me, they are usually practice, exercise and sometimes training. Now, it doesn’t make sense to me to say activities such as reading a Go book are “training”, even if you read Go books with a “no pain, no gain” attitude; it is another mislabel from my perspective.

I’m not sure if you only wish to talk about training, since that is one of your two categories, I mean I talked about more general activities not only training. You gave some examples of skill acquisition in table tennis and Go. It strikes me that the skills that you identify for table tennis all involve fitness and coordination, and the skills that you identify for Go don’t. The table tennis skills that you mention are immersive tasks which totally encompass the athletes’ experience for the duration of that task. I think they are meant as examples of training. When it comes to Go you don’t mention tasks that I think should be immersive and I’m not sure why they should be comparable to the table tennis tasks. Yet, playing Go also involves a series of highly immersive tasks that require a high degree of mental fitness and coordination of different thoughts, not unlike how table tennis or any physical sport rely on physical fitness and coordination of the movement of the body. I don’t think it should be difficult to come up with examples of training for Go.

2. You said - if I got that right - that computers don't take into account the human capacity to absorb their advice. I offered https://neuralnetgoproblems.com/ as a good example of how AI combines its knowledge with the human learner's level to provide well targeted exercises and solutions. We seem to disagree - if I got that right - on the quality of that resource.

That is not at all what I said. I don’t know what the argument is about here, it is as if I write one thing and you then read something entirely different. Sure, I haven’t found the problems on that site to be great. I have read your praise of them, but even then, I don’t understand what the connection is. You don’t appear to take any issue with what I did write.

OK, let's put it to rest then. I'm rather tired of misreading and misquoting you. I'm not doing it on purpose but neither do I question my ability to read English.

[Javaness2]

It's an interesting idea, but I'm not sure that's the best medium with which to deliver feedback during the same.
I think if I was playing I would just find the additional noise offputting. Did you try the same idea but with a purely visual feedback for the key ideas, urgency & state.

When I review a games I probably focus on a few things
- If there was a fight I totally misread, I revisit that.
- If I wanted to play something but didn't out of fear, I revisit that.
- Look for places that I felt I went wrong

If I have an AI handy I would additionally (after doing the above)
- Look to see where I totally messed up, according to the AI
- Look at any move the AI might suggest which is just unexpected for me
- Where I had no idea what to do, check the AI's suggestion

[John Fairbairn]

One thing I've noticed over the years in many fields is that mathematicians, programmers, scientists and the like cleave religiously to the scientific method, logic, repeated testing and the like, and the aim is always to end up understanding rigorously how things work, or what is going on, and how to avoid mistakes. In their own fields, I think that is a wonderful thing. I'd certainly prefer to be treated by a doctor who knows what he is doing and why, for example.

But there is a risk among such people of déformation professionelle -- looking at fields in which they are not experts through the prism of their own major work. Like applying logic to go, for example.

There are problems even within scientific fields. All the fancy maths, logic, scientific theories and philosophising that went into "experts" telling us that the world began with a Big Bang has been shot down in an instant (if - big if - I can believe Youtube) by the Webb telescope. That telescope is itself a masterpiece of human ingenuity based on maths, logic and scientific method, of course. But looking at the rather wider problem of the universe, their methods have failed.

Indeed, these methods have failed on the slightly smaller universe of the go board. AlphaGo and its spawn produced bots that beat the best humans, yes. But have they explained how go works, why some moves are better than others, etc etc? No.

Yet in this thread and others on L19 we are repeatedly seeing mathematicians, programmers, scientists and the like trying to apply their professional procedures to go - and getting nowhere.

But along comes a little girl with no special interest in maths, science or programming - and certainly no expertise, and no time to acquire it yet - and she reaches a level (1-dan pro) at age 11 that far exceeds any of the mathematicians, programmers, scientists and the like. And she's not that unusual. Lots of young kids in the main go-playing countries are playing at pro level.

I don't believe for one moment that Nakamura Sumire and her ilk indulge in the navel gazing of trying to understand their moves or mistakes in the various fashions described here. They just copy the best moves they see. They learn by mimicking. Indulging oneself in trying to unravel mistakes wastes time that could be better spent in mimicry of good moves - it could even lead to mimicking bad moves by playing over one's mistakes and reinforcing them in one's subconscious.

Take an utterly trivial analogy for comparison. A little child grabs a toy. A parent says, "Don't do that, say please." The child says please next time, mimicking polite and successful people, and gets the toy without a clip round the ear. Some parents may philosophise about politeness and morality to the child, but what little child listens to that? They don't even understand what you are saying. It merely makes the adult feel good (been there, done that ). But the odd thing is that the child doesn't actually understand the word 'please' either. It just mimics the word and gets the toy. Magic. Template for future behaviour!

People like Sumire have discovered a way to spot the go equivalents of saying 'please' - and again without understanding why they work. Their brains "understand" in more or less the same way as AlphGo, of course, through the intuition produced by neural networking, but it is their subconscious brain that is doing all the clever work.

On a bigger scale, this is how we all - including mathematicians, programmers, scientists and the like - learn language. Logic doesn't work there. A child says "I breaked it", you say "I broke it." You both understand each other. One form is not mathematically, logically or scientifically better than the other. We just agree to treat one version as socially more correct. That is, we don't really need to understand (though, of course, a neurolinguist might get great personal satisfaction from try to fathom it).

It is my contention, therefore, that those who wish to improve significantly at go (1 amateur dan in two years or something like that is far too slow to count in that respect) must treat the learning process in the same way as learning a language, or in the same way as Sumire learns go. That's what pros have been telling us for decades, after all. Play over lots and lots and lots and lots and lots of pro games.

But we now have a new tool: AI. OK, but don't use it to do scientific analysis of mistakes - if you want to improve significantly, that is. If you just want the glow of exercising your professional skills, it's fine for that, but that's all it gives. The correct approach seems to be, instead, simply copying what you think are AI's good moves. If they work in your real games, you will note that and so your intuition will also note that and prompt you to play them in future. If they don't work, your intuition will note your disappointment and not prompt you. Of course, if you override your intuition and keep playing what are bad moves, you deserve what you get.

One common argument is that adults can't spend the time that young people spend on building up intuition, and so have to try to find a quick fix in other ways. Broadly, I accept that, but the point here is that AI may (MAY) be a way to streamline the mimicry process even for adults. We won't know unless enough people try it, but it has to be said that mimicry of AI without proper understanding has worked for many top pros (in chess as well as go).

I hope it goes without saying that other skills as diverse as reading or controlling one's nerves are necessary for true mastery, but without filling the databank with data, those other skills have nothing substantial to work on.

I hope it also goes without saying that improving to 1-dan pro or whatever is not necessary to enjoy go. People like me - and I suspect I'm in a very big majority - just like adding snippets of knowledge continuously as a way of better appreciating the best go players. We are go fans. Our needs may therefore differ. We can probably get by with no AI at all.

[RobertJasiek]

John tells a tale and covers it with alibies, such as mentioning the word reading.

[Knotwilg]

You do realize, John, that mathematics is also a kind of language, with a syntax of symbols and semantics that try explaining the world in a certain way. And although we have acquired that understanding in a rigorous, structural way in the curriculum of the past decades, it has always been helpful and way more interesting to follow the history of scientific progress. Education has realized that and is now swinging away from the axiomatic approach, back to the experimental approach.

The dichotomy between the cold, purely logical, theoretical, non-empirical approach in science, and the warm, intuitive, concrete and imitative acquisition of language, is false.

In both your examples of the origins of the universe and the game of Go, I think science has actually been very successful. But man can't grasp the notion of "nothingness" in space and time. Surely there must be something outside the borders of the universe. Surely there must have been something before the big bang. Similarly, AI is now playing Go on a superhuman level. "It" understands perfectly well what it does. We don't, but that's not AI's fault. I think BTW that, if we give AI the challenge of explaining what it does, rather than playing well, we might get there very soon.

[dust]

(snip)
The thesis is "Longterm training with live AI feedback (in addition to "classical" training) has the potential to improve evaluation, judgement, intuition and playing strength in general". A counter-example would be if someone had actually studied this way for a long time but seen no improvement. And even one counter-example wouldn't be a refutation since I'm not making all-or-nothing claims here.
The only scientifically valid way to approach this would be to have a group of players (big enough to be statistically relevant) train longterm (months/years) with this method and compare their improvement with a control group that followed the exact same training regimen, only leaving out the live feedback during games.

Anyway my point is not to devalue any existing methods for improvement (play, study, tsumego etc.), on the contrary, they are absolutely necessary. I'm just proposing that we can add something more to get a little extra boost.

The human brain is an all-purpose learning system, and it might be able to incorporate AI feedback signals as well in order to improve it's internal model of the game of Go (or other domains).

That is my theory and only time will tell us if there's anything to it.

It'll be interesting to see this theory tested.

It's also good to have a new idea to shake up the rather familiar discussions on L19 on 'how to improve' featuring a handful of ageing regulars - each with their own idiosyncratic strong opinions, and none of whom are probably improving significantly (I include myself fully in that description)