I've already ordered the book, thank you.

This made me smile:



When I was working on computer shogi, I attended lots of major chess events (as a journalist) and met lots of top players. My basic spiel was always "why don't you try shogi" of course. Conversations about the differences in shogi would often ensue, and one of them is that it is very common to push the edge pawns early in shogi - moves than can seem mysterious to weaker players. I can't remember who it was, but one grandmaster leapt on that with pleasure because he'd noticed early pushes of the rook's pawns was a distinctive feature of Bent Larsen's play that had intrigued him, and he wondered where it had come from. Another person said he knew that Larsen had at least looked at shogi.

Another shogi feature that excited grandmasters was the shogi technique of castling in which the king goes on a march. This became a fairly common tactic/strategy in chess around then but I have no idea if it was borrowed from shogi.

I remember standing next to Karpov as he analysed games in the press room (i.e. seeing the game from the same angle as him) and being struck both in his play and comments by a resemblance with pro shogi talk (of which I had the good fortune to get a lot).

Chess people told me that many of them had looked at shogi, found it fascinating and had probably borrowed ideas from it. Larsen may have done this, but he wasn't around for me to ask. I did ask various players why they didn't give shogi a try professionally - the money's much better in Japan - and I think Walter Browne summed the general reaction best: "I've spent a lifetime getting this far in chess. Why should I give all that up?"

Later on, there were a couple of cases where shogi and chess pros did try the other game at a high level. Larry Kaufman was then not quite the chess grandmaster he is now but he was almost certainly the strongest western shogi player, yet he found it impossible to get beyond being a strong amateur. The other way round, Habu Yoshiharu, a shogi sensation (think Sin Chin-seo in go), tried out pro chess and got to FIDE Master level. I'm not sure what represents but it's clearly not Grandmaster level. So my earlier assumptions that top pros in one chess variant could transfer their skills to another were, in practice, probably too optimistic.

There is a lot of talking to the effect we can't learn well from AI because they can't teach by explaining things the way humans do. There is a common teaching technique that is used by masters teaching disciples by telling them that what they are doing is wrong without explaining why. Apparently the idea is that the pupil is required to figure out for him or her self what is "right". This could be frustrating for the student but if one figures things out for one's self it will be well understood and could even lead to creativity.

In the end it is always the student that learns. Reminds me of Meno's paradox. Also if pro knowledge can be overwritten by AI knowledge is this then a sign that this was all only true believe?

Adding to that, I don't know how one could advocate "replay pro games" as a well tested method to improve while warning against "learning from AI", which is essentially watching a pro playing live.

Also, pros learn from and with AI. What's good for the pros is not always the best diet at the recreational level, so I'd agree that as a beginner or whatever digit kyu or low dan it's best to learn from a guide, who can articulate things and can pace out the learning curve, but as one grows stronger, there is definitely a case to be made for learning more like pros do. I'm not watching over the pro's shoulders as they learn, so I accept they will often revert to old masters games, as a way of finding fresh ideas, or calibrating the AI "truth" for its fundamental or specific aspect. But AI is a fundamental part of their study, so we can't be too far off if we do that ourselves, not farther off than anything else we do as mere amateurs.

I've taken a few lessons with pros and high dan amateurs, and use AI on a regular basis. I felt I learnt with both. Not because pros can't show the same things as AI, but because AI talks to me much more often. However I only spend a few minutes to review each game. Go is just a hobby and I prefer to spend most of my time playing, without caring too much to improve. The AI is just there to fill my need for explanations: what was the joseki move? Was there a way to kill the black group at move 124? How to defend at move 146? What happens if Black cuts at move 157?

Dieter, it’s not just replaying pro games, it’s doing that in a very specific manner, namely from a single diagram (and not clicking through an sgf file). The idea is to turn every move into a next-move problem Doing that enhances the experience (apparently - I’ve never really tried it against the clock, though I have done a lot of transcribing.

Also it is misleading to talk about a warning against AI. As far as I am concerned, the danger I is throwing the baby out with the bath water as nd forgetting the valuable human pro resource. This is especially likely to happen when people say to themselves AI is free but pro resources cost money.

OK - thanks for reminding and restating that challenge. I might give it a try.



Also fair.

I have taken lessons with Guo Juan, who could articulate pro understanding to a certain extent. I eventually opted out of it because I didn't get what I expected from it - which is my responsibility not hers. The best articulated pro level understanding I've ever come across was that by Kang Minue which took a hybrid form of KGS lessons, sgf outtakes and articles on SL.

As I've probably already recalled too many times, my most tangible progress came from self review, 20 years ago. Now that could be a combination of relatively young age (30) or the fact that 2k-2d is easier to cross than 2d-3d. But indeed it could be due to my approach at the time, which was void of any influence and using the physical goban for review and replaying from memory, rather than letting the sgf take care of that and browse through it.

With your repeated hinting at the dangers of AI-only based progress attempts, I may actually return to the slow, goban based analysis, without permanent supervision by AI which both inflates the impression of good review and degrades the actual work the brain must do.

Even more than philosopho-strategic advancement AI has brought about in Chess, of which I already am very interested, I wonder about AI contribution to Elephant Chess and Shogi all three games are likely to now be played by AI at a high level and with correct philosophy, so we can safely take it's opinions as divine ordinance. Large-size gameboards bigger than modern shogi is not 'solved' by AI to the level amateurs tend to behave.

As far as I know, while ways have been found to program pure neural nets to have some competency on ladders without it being hardcoded, they still fail on the fancy ones. This shows certain situations covering the breadth of the board are still beyond the scope of AI, and another example which shows this is Igo-Hatsuyron 120.

It's simply the case that human pros have not discovered these limitations partly because it would be hard, too, for human pros to discover it. But if it is they'd be able to reduce their handicap to the current level of AI form two-and-a-half stones to one-and-a-half-stones.

The Irony is that Neural Nets were employed to play pro-level go and they did do that, but for the games of Chess and Shogi, a new level beyond pro-level play has been achieved; philosphical perfection, you can take bots at there chess-playing word. Go is still yet to be 'cracked' in that regard, so ironically still remains an AI enigma compared to other games.

Well, I tried out the guessing game on the most recent KG net (took me many hours). I wrote down my top 5 suggestions (often only 1 when tired or thought it was obvious).

kata1-b60c320-s642... rating game 758309

Unfortunately for the stats, the latter half of middlegame was a ko, for which I could guess the top move most of the time.

Of 262 moves, here is the rank of my move, where ? indicates a move outside of my top 5.
1:154
2:27
3:9
4:5
5:3
?:64

As you can see the AI openings are comparatively fixed, so the first 50 moves can be easy. Arguably fighting can be easier because there is often just one move to save/kill a group. The most difficult was the opening-middlegame transition and then the middle-late endgame (where there are lots of smaller moves).

Sometimes there is lian2guan4si1lu14 (connected plans), so knowing the AI move tells you the next few AI moves, but sometimes it is misleading towards the end of the game when the loser starts messing around.

In the chart, I plotted moves outside my top 5 as a score of 7 (the lower the better match I have with AI).

Assuming I'm right in thinking what was going on, my basic response to that is that although this version of the guessing game is valid (in fact GoGoD introduced at a London Go Congress many years ago) and popular - Bill Spight was a big fan - this is not quite what I am talking about here. And so not what the Japanese pros are talking about.

What is meant here is taking a single diagram (which will average about 250 moves) and recreating the game move by move on a board (or making an sgf record, of course) as quickly as possible. No real thinking, just intuition. There is no need to work out a list of candidate moves or to assess them relative to each other. The idea is simply to find the next move as quickly as possible, to place it on the board, and go straight on to the next one.

Mark Hall could often do this in about 20 minutes per game, which is a stellar performance according to the scoring criteria given in Japanese magazines, where I think 30 minutes is the top level for amateurs. There are pros who can do this at high speed just playing the game over in their heads and not on a board, AND understand it enough to make comments such as "White 142 missed a chance to kill Black." But actually Mark could do that sort of commentary as well (he just couldn't do the in-head stuff). In contrast, I could usually transcribe a game in well under an hour but I would very often fall asleep in the middle. That is not a self-deprecating joke: I really did find it soporific. Mark could do six hours at time, but four hours was usual, which is why the GoGoD database (and others based on it ) is so large.

I'll try and find an example from Go Monthly and post it here, but it won't be for a little while. Too many new dances to learn this week.

Incidentally, one other thing that Mark found, which illustrates it was his intuition that was being improved rather than his go "knowledge", was that on the rare occasions he transcribed an amateur game, it took him more like an hour and drove him nuts. Even though these were high-level amateur games (e.g. Amateur Honinbo final) the amateurs were not making the moves in the places he (so used to pro play by then) expected to find them.

Pros recommended a lot of things when people ask about getting stronger. Sure, replaying pro games is one of them. But life & death, playing games, and reviewing are also popular recommendations.

Practicing what you actually want to get good at has got to have great practical value.

Pros also have some incentive to recommend methods of study that require the existence of pros. After all, their fans are what keep them in business.

_________________
be immersed

Michael Redmond has Go lectures on YouTube. He is recently doing English lessons on proverbs which are sort of a modern refresh of his old NHK lecture series on proverbs.

My favorite parts of these lectures are when he mentions the moves that AI would play and how he does not recommend such moves to amateurs.

Splitting the side: AI will not play it in the opening (maybe some exceptions), but it's a valid play even at the pro level.

That's great. Because I can understand and use that move.

Good point. When you look at what the AI there's usually not much difference between any of a number of plausible moves--less than a point by KataGo's estimate. I'd happily give up a point or two to get a game that I'm comfortable with. After all, I'm going to lose 50 points through the middle game.

There's also something to be said regarding the type of game you are playing. Even before AI, sometimes I would see some joseki pattern and not totally understand it.

I'll actually use a pattern that became popular after AI as an example:

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X X . O . . . . . . . . . . . . . |
$$ | . . O X . . . . . , . . . . . , . . . |
$$ | . . O a . . . . . . . . . . . . . . . |
$$ | . . O X X . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |[/go]

I saw this pattern for awhile, always feeling uncomfortable about the cutting point at 'a'. I didn't spend enough time thinking about/studying it, and played the pattern in a tournament. Given that I was in a tournament, I started worrying about that cut again, and played something weird, because I didn't understand the shape well. I lost the game, and regret not studying the pattern before.

Why did I play a pattern that I knew I didn't understand well??

On the other hand, if I am playing a non-tournament game for learning, pushing the boundary of what I understand and trying to play something that I know is probably correct - even if I am not an expert on it yet - it would seem to be a good way to learn. My opponent may play in a way that I didn't think of, and it's a good opportunity to learn and understand *why* it's a good shape.

In the same way, I believe that there are a lot of good AI moves that I don't understand, yet. Trying them out and experimenting with them during my games is a good way to get a feeling for how that pattern develops and/or what kind of shapes result.

But if it's a tournament game that I want to win, it's probably best to stay in comfortable territory.

_________________
be immersed

Qualitatively the graph looks different than a professional's. John noted in his "First Teenaged Meijin" that in the last game, the AI largely agreed with the pros. The agreement tended to be "clumpy"--they played sequences that the AI agreed with, but then would make a mistake in direction. Once the new direction was selected, though, they seemed to do quite well again.

I am curious what the longer-term effects of the AI on professional play with be. Many of us remember the new-fangled idea of playing online go, but it was about 10-15 years before those players who grew up on online go became professionals. Pretty soon we'll see a new generation who have always had an AI available to them.

I saw a bit of a match, once, between a resurfaced Navratilova and a young promise in the top... 5? Graf? Can't recall. Thing is, Navratilova's precision and reading had the younger player an inch from throwing her guts, teenage power and all.

Take care

_________________
一碁一会

What you are saying is that "direction A is OBJECTIVELY better than direction B". And you need to look at the percentages at the end of each best play line, not just after the first move.

But the pro might NOT be making a "mistake" because a real game between two particular pros is not purely objective. In the real game between two particular pros there is a matter of the current state of the game (who is ahead) and whether the direction A or the direction B best fits the style and strengths of each player.

Thus if player 1 is slightly behind, if direction A results in a simple best sequence but direction B leads to a complex sequence where there are lots of alternatives, hard to pick out the best sequence, more chance to go wrong, then B might be the best PRACTICAL choice. Or perhaps player 1 is a "fighter", player 2 is not, and direction A is "quiet" compared to direction B.

When rating the two directions, AI is assuming it is playing against an opponent equal to itself and is not allowing for "which direction offers most chances for a mistake". .

Found this topic again because I was thinking more about how well human pros do in matching AI plays/recommendations, and I was also thinking about a (non-pro) human commentary technique I've seen, which I sort of don't like, and then wondering how often I see pros do this...

I don't like when commentary points to a "mistake" according to the AI. I do like when a mistake is pointed out and an explanation is given as to how it was refuted or why the game result started to turn sour because of that play, but all that could be done without saying things likes like "this move lost a few points." It's just that this type of description doesn't tell me much about the game itself (as others can have described better here on L19 and elsewhere).

And this seems to be a shift in the perspective of commentary. The player made a mistake according to the AI analysis. Not that their opponent took the initiative. Of course, game commentaries always liked to point out moves that were slow, or which were the losing move, but I feel like this tendency to phrase the move in the context of a "mistake" seems to be growing. Most mistakes hardly seem like an actual "mistake" in judgment.

I've already said that I don't prefer this commentary style. But I don't think it's a bad thing and I wonder if others like it. It is good to have an "expert" opinion on the moves. But I just appreciate the discussion of how the other player seized the opportunity, not identifying how many points the mistake cost.

I think your instincts are spot on here. I have thought a lot about the art of commentary recently because it is what is at the heart of my Museum of Go Theory project. One thing that struck me about the latest work in this series, the Yi Mo of 1662 (not yet issued), is that the commentator (Ji Xinxue, a top master) makes sure he praises both sides if possible, and his praise is frequent. And the same goes for criticism of mistakes. But more importantly, he strives to let the reader know what is going on in the game.

I find this is typical of pro commentaries in general. It is the amateur commentator who harps on mistakes. Probably this is because mistakes are easier to spot (especially now, if you have a bot handy).

I agree that we need to know when a mistake is a mistake, but it should be like seasoning. We need salt on our food, but a plateful of salt is no substitute for a plate of roast beef, Yorkshire pudding, two veg and gravy.

Yet another way of looking at it is journalism, which is after all just a form of commenting in the world. A typical news story tells us what, who, when and where about an event. Valuable information - but how do we process it? Probably just through our existing prejudices. We may be more knowledgeable, but no wiser. A discursive feature, however, adds "why" and "how" to the mix. That's far, far more valuable.

I think the message is: stick with the pros.

This is obviously the best case possible.

If possible, it's always desirable to have a comment from a human, even if this human is not a pro. Usually, just a stronger player than us will do, because it's what most players always have. But I think that this is stating the obvious. Only a very strong player can benefit only from AI suggestions, with no more explanations.

So, what to do with "mistakes according to AI"? Use with caution, I'd say. In my case, I look at the AI evaluation, and I try to see if it matches my opinion. Or when things go wrong for somebody, I look for the critical move and try to see why.

As long as you don't use AI to say "look, Iyama made a blunder, what a noob" or other hooligan-ish comments like that...