TMark's promotion record is reasonably simple to find, especially if you know where to look for it


I'm not sure why AI would make us any more lazy than a joseki dictionary would do. It's simply another source of knowledge you can refer to.

_________________
North Lecale

Off to a tangent of tangents here: I'm probably a numbers guy but while I'm borderline savant on (birth) years, I'm disastrous at (birth) days and months. The only explanation I found for that is that years are linear, while day+month is circular. I tend to assign a kind of (synesthetic) identity to a number. That works for a year but it doesn't for a date of any year. I guess ...

I can also relate to frantically searching for X while X is in my pocket/hand/... I have changed my idea about this trait from being absent minded to being caught in a preconception. The idea that I lost my wallet overrules the sense of holding it in my pocket.

A lot of interesting comments!

Now that the curious case of T Mark Hall has been mentioned again, I wanted to bring up something that I've recently been wondering about the advice of "replay pro games to get stronger", especially if the claim is that doing so is actually the fastest way to improve.

Before I go on, I want to say that would love for this to be true. Studying pro's games (Go Seigen in particular) is one of my favourite ways to spend time in front of the Go board, and I have in fact replayed about 200 of Go Seigen's games so far in the same way that T Mark Hall did, which is by replaying them from a single diagram that contains all the moves of the game.

The main thing I noticed was that my intuition became a lot better, but this lasts about as long as the opponent plays somewhat "proper" moves. Take the following example:

Click Here To Show Diagram Code

[go]A common shape
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . . .
$$ . . . . . . a . . . .
$$ . . O . . O . B . . .
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

This marked shape comes up very often in one's games, and so naturally also occurs often in Go Seigen's games. However, play this shape in a DDK (or even SDK) game, and it's quite likely that your opponent will play at 'a'. How do you handle it? This might be a bit of an easier example, but my point is that (I believe) simply playing through Go Seigen's games will not really teach you how to handle the move at 'a', because the black player (either Go or his opponent) will not play this shape to begin with if the move at 'a' is a good move for the opponent. So when you're training your neural network by playing over 1000 pro games, you will not really encounter the types of suboptimal moves that your opponent plays against you. This could foster the type of player who manages to win against a 2 dan but then loses to a 5kyu, but I think most people here would agree that this type of player doesn't truly "understand" the moves he's playing then. That brings me to the point of overplays and punishing mistakes, and that's where AI comes in again. Even if we simply compare the two methods of "studying pro games" and "reviewing with AI", and abstract all the rest away, I think AI is capable of giving more practical advice on your own game, because it can and will show you how to mercilessly punish the opponent when he oversteps his limits.

Then there are even skills that neither AI nor pro games can really teach you. One of them would be "visualisation" (as an aspect of "reading"). In-Seong sometimes tells us students to "train our go eye" by trying to read out the solution to a tsumego again after having solved it together. It's there that I noticed the astounding but obvious gap in "visualisation" capabilities between me and In-Seong. Sure, replaying pro games would help a lot with shape recognition and pruning away branches, but without the ability to hold a 30-move variation in your mind and evaluate the result, the shape knowledge will end up coming short.

Anyway, I'm deviating from the topic at hand, so I'll better stop rambling before I deviate even further.

No, no, no, no! You are neither rambling nor deviating. Some topic headings are very specific, and so it's not always polite to derail them, but others are just starting points for a good old natter. As here. After all, if we start a conversation in the pub with "Chilly today, isn't it?" we wouldn't expect to talk about the weather all evening.

I found what you said interesting but I do disagree with some of it. For example, in the magari pattern you showed, you may never see a pro play at A and so get flummoxed when your amateur opponent plays it. But your brain still learns something from the fact you never see it, and so it will tend not throw it up as a candidate move (which may be a good thing).

I would suggest that what is supposed to be happening when you are playing over pro games is that you are learning suji. How to make your stones flow, how to get the right (long-term) shapes, and how to pick out the best candidate moves. What you are referring to in amateur games is something quite different: reading (yomi). You don't learn reading by playing over pro games. For that you do tsumego, tesuji problems or visualisation reps.

You need both suji and yomi, but suji is often the best one to start with because its knack of throwing up good candidate plays enables you to prune the dense trees that pure reading generates.

The drawback with suji is that you can become very LAZY (see how easy it is to get back on topic!). We amateurs are too apt to play by shape alone and not read out the consequences.

A problem with using AI in this sort of situation is that you get a "you can have any colour you like so long as it's blue" response, and also a response that is very game-specific. I'd suggest using suji and kombilo instead. Put in a shape as above and you'll get the full range of moves ever played, and in order of frequency, along with the pro games that show you various contexts and also the follow-ups to the very end of the game. I just put in the pattern you showed into kombilo and got 63 pro-game hits for the kosumi (so it does occur after all - never say never!). In reply, the two possible connections are the main moves and are split about 50-50 relative to each other, but there are other possible replies, some very surprising. These too are precisely the types of move that can improve your suji no end by showing you new stuff you can experiment with.

Some comments on the topic of using AI for studying and laziness. I understand how Mark Hall could improve his playing strength through transcribing pro games. Pros also recommend laying out the moves of a game as a method of studying. Yoda Norimoto pro 9-dan wrote a book titled Yoda Ryuu--naraberu dake de tsuyoku naru--Kogo Meikyoku Shuu (Yoda-style--Become stronger just by laying out the moves--A Collection of Old Go Master Games). It is common knowledge that there is a correlation between playing strength and speed of finding moves in game records.

Click Here To Show Diagram Code

[go]A common shape
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . . .
$$ . . . . . . a . . . .
$$ . . O . . O . B . . .
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

Analogies rarely fit perfectly, but this example brings to mind shooting free throws in basketball. If you're not a basketball player, you can look up the "proper form" to shoot a free throw on the internet: position your body, bend your knees, shift your weight, line up your shooting arm, yada yada yada.

Practice all of the textbook stuff, and you can have the proper form to shoot a basketball.

But it's not the same as learning how to... properly shoot a basketball. For that, you need to practice and gain experience. Play some real games. Get used to shooting when there's pressure. The list goes on. But the proper form will help you in the end. Combine proper form and experience, and you should see a good result. Why? Because "proper form" has been developed by players over the years, and with enough practice, we know that people can apply it to make it work.

Back to AI. Reviewing a game with AI can give you ideas on the "proper form" to play in a given position. Same thing with joseki and/or learning from pros. But it's necessary to combine that knowledge with experience.

Someone with bad form in both basketball and in go may be able to win games by sheer experience and strength. But if you're going to be practicing anyway, why not practice with good form?

_________________
be immersed

As for laziness, what about playing through game records by using a computer app rather than reading a game record and putting stones on a board? Using a computer app tends to encourage just clicking through, the moves hardly registering. Playing through in person with a board and stones offers us more opportunity to experience the moves. First you have to find the move in the diagram, second you have to pick up a stone and put it in the proper place on the board. Both of these activities allow us to "see" the move and its relationship to the whole board position. Teachers know that students who take handwritten notes in class tend to learn more that just listening to the lecture and even reviewing the lecture if it was recorded.

For those who are not familiar with the implications of that black shape, I'd like to share some shape knowledge about it (as we are allowed and even encouraged to wander somewhat off-topic here).

The stability of that black shape relies on this tesuji.

Click Here To Show Diagram Code

[go]$$W Refuting the double peep ("Elephant's eye") and cut
$$ . . . . . . . . . . .
$$ . . . . . . 4 5 . . .
$$ . . . . . B 3 B . O .
$$ . . . . . . 1 2 . . .
$$ . . . . 6 O . B . O .
$$ . . . . 7 8 . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

So white's cut is not working well. White needs an extra stone at any of the marked spots to invalidate black's tesuji.

Click Here To Show Diagram Code

[go]$$W White's threats
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . O .
$$ . . . . M M a . . . .
$$ . . . M M O M B . O .
$$ . . . . M M M . . . .
$$ . . . . . M . . . . .
$$ -------------------------
$$[/go]

Note that this does not include a white 2-space extension or knight move to the left, so those moves would not really prompt black to defend against a.

So a common way for white to activate the a cut, is a one space jump to the left (a double purpose move, defending and attacking).

Click Here To Show Diagram Code

[go]$$W Black needs to defend
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . O .
$$ . . . . . . a . . . .
$$ . . . 1 . O . B . O .
$$ . . . . b . . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

A clever black defence is the probe at b, making miai of A...

Click Here To Show Diagram Code

[go]$$W Sabaki for black
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . O .
$$ . . . . . 4 . . . . .
$$ . . . 1 5 O . B . O .
$$ . . . . 2 3 . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

and B...

Click Here To Show Diagram Code

[go]$$W Sabaki for black
$$ . . . . . . . . . . .
$$ . . . . . . . . . . .
$$ . . . . . B . B . O .
$$ . . . . . . . . . . .
$$ . . . 1 3 O 4 B . O .
$$ . . . . 2 5 . . . . .
$$ . . . . . . . . . . .
$$ -------------------------
$$[/go]

I probably aqcuired this knowledge from books and lessons I got from stronger players. I suspect it would have been difficult to discover this kind of shape knowledge totally by myself, even with the help of AI.

Thanks for sharing. I have often seen the shape but I don't know if I ever recognized the tesuji, instead just focusing on the cuts.

Without AI or study, my understanding of this shape comes from the following position in a professional game, supposedly great for White. White has the 2-stone wall 2 spaces away and bumped against the cap before playing the peep.

Click Here To Show Diagram Code

[go]$$W
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . 5 . . . . . |
$$ | . . . O . . . 8 6 . 2 . . . . . X . . |
$$ | . . O , . O . . 7 9 . . . . . 4 . . . |
$$ | . . . . X O . . O . X . 1 . 3 X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------------------
$$[/go]

Surrounding positions matter. With the support of the nearby 2 stone wall, a one space jump would be very inefficient shape, so naturally, white didn't play there.
But still the elephant's eye is not working because of black's refuting crosscut tesuji, so white bumped to urge black to defend. But black didn't want to follow white's orders, so he resisted, allowing white to play the elephant's eye that is now working (move 80 of Akaboshi vs Jowa, "The blood vomiting game").

That white bump up (uppercut?) seems to be quite rare in pro games though. So I'd say that it looks like a very circumstantial move, not like the "book" move against this black shape.

So if you only play the uppercut against black's shape (because of that game), then maybe you also want to check out some of the many pro games where white plays the much more standard one space jump.

While I'm at it: that black shape may also work nicely in situations like these:

Click Here To Show Diagram Code

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

invades black's area, threatening to link up with a or escape around b.
Black may be at a loss how to prevent both threats with one move.

Click Here To Show Diagram Code

[go]$$W
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . X . 2 . . . X . . |
$$ | . . O , . O . . . . . . . . . . . . . |
$$ | . . . . X O . . O . X . 1 . . X . . . |
$$ | . . . . . . . . . . a . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------------------
$$[/go]

But defends against both threats. Even linking-up underneath with a is not working anymore:

Click Here To Show Diagram Code

[go]$$W
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . X . 2 . . . X . . |
$$ | . . O , . O . . . a 7 9 0 . . . . . . |
$$ | . . . . X O . . O 5 X 8 1 . . X . . . |
$$ | . . . . . . . . . 4 3 6 . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------------------
$$[/go]

White needs to defend against black a now, so the invasion is failing.

What about

Click Here To Show Diagram Code

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

Of course, other things are possible.

Next post: Gennan's tips and tricks making me lazy?

Click Here To Show Diagram Code

[go]$$W
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . Y 4 2 . . . X . . |
$$ | . . O , . O . . . . . . 3 . . . . . . |
$$ | . . . . X O . . O . X . 1 . . X . . . |
$$ | . . . . . . . . . . 5 6 . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ -------------------------
$$[/go]

for looks like a really bad exchange. I find it very hard to imagine any sequence that could work better with + than without these two moves.

I think it is a really good advise to learn how to play normal moves rather than study the exceptional. For one thing it is because you need some baseline to compare with. One also needs something to fallback on in time trouble and when confused. Black didn't have the marked stone in the original position, in this case is not really normal and this position can quickly become very complicated if white tries to justify the invasion.

It was cheeky, hence the above. I did look with Leela already (making me less motivated?) and playing is not "good" but it's not really "bad" either (by my standards). Also, Leela did suggest that black extend right rather than connect at .

I agree is the goal is improvement. But if the goal is entertainment, then it is really good advise to review the exceptional. I didn't actually study any of this. I reviewed it for entertainment. And this game's position was just the first thing that I recalled when seeing that shape. Next time, my first recall might be what Gennan said.

This sort of goes back to the original question in the first post: "Ryan Li’s recommendation is to generally use AI less, because knowing what the correct move is in a situation does not help furthering your understanding of the game." For me, I'll use Ghost Intelligence less, but only after Halloween.

The same could be said about josekis. Maybe we would benefit by deviating more often from josekis.

Click Here To Show Diagram Code

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

Before AI we played this joseki, up to pro level.

Click Here To Show Diagram Code

[go]$$W Invasion, interception
$$ ------------------
$$ | . . . . . . . . .
$$ | . . 7 5 6 8 . . .
$$ | . . 1 2 a O . B .
$$ | . . 3 X . . . . .
$$ | . . d 4 . . . . .
$$ | . . 9 b . . . . .
$$ | . . . . c . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]

AI played this move and so we have investigated why this would be better. The general understanding is that the aji of the cut at A here is less cumbersome than the aji of White dropping down at A in the previous diagram.

AI has showed us the right move and through studying variations we have improved our understanding of the game.

We can of course agree on the tautology "using AI in a lazy way will make us lazy".

I beg to differ, not with you but with the "general understanders".

I'm totally prepared to accept the crawl as better if for no other reason that, when a beginner, I thought the crawl was the obvious move and couldn't ever completely understand why the connection was favoured.

The bit I have difficulty in accepting is the alleged modern reasoning. "Less cumbersome" is rather abstract and meaningless. It is insufficient. It is a "lazy" explanation. Which means AI has so far taught us zilch about this position (except to play A not B on the monkey see, monkey do principle). I think it is reasonable for us to expect a more concrete explanation.

FWIW I think a passable concrete explanation can be that the crawl effectively creates a Go Seigen group on the side if (here) White cuts. I came to this conclusion after seeing someone (I think it was Sumire) willingly accept the cut and give up the three cut-off stones. We already know from Go Seigen that offering to give up three stones is very "cumbersome" for the opponent. It seems to me that bots and the best modern players seem increasingly willing to give up even bigger sacrifices. If I'm right about that, three stones must seem even more trivial now than in Kamakura days. And going on from there, AI is therefore not teaching us anything about the tactics of one joseki at all really - it is teaching us about whole-game strategy.

So,



Yes. Absolutely.

That's at least a conceptually richer explanation than merely describing the aji as more or less cumbersome. Let's say that was just me describing the advance in understanding somewhat lazily.



I agree. In the Suzuki thread, if I was able to learn from that position at all by using KataGo, it's that again and again, AI is looking for sente because with every local move, temperature may suddenly drop, or otherwise stated, chances for temperature to drop are higher than to rise, so you should always be on the lookout for sente. That matters more than the actual local result. My words/understanding.




And the same applies to human (pro) advice. The main difference with human advice is the reasoning behind the move. I'm not convinced that getting the best move + reasoning would make us less lazy than merely getting the best move.

I don't use AI or look at AI-AI games, though I do look at lots of games by pros who are following AI. That probably gives me a skewed version of what is going on in the AI 'mind', though I do think it's easier for us amateurs to learn from such pros than from AI direct.

However, reality is that my go environment is also heavily affected by western trains of thought, notably what I read here. Yet I often feel a strong disconnect between that and what Japanese pros say. Sente is one of the main areas of polarisation.

I imagine western go players include many chess or ex-chess players and so we are influenced by chess thinking and chess terms. As such, we take very much to heart Bronstein's famous dictum "The most powerful weapon in chess is to have the next move.” Add to that the fact that westerners adore the lilt of the word sente, we end up using it far, far more than you ever see it in Japanese (or Chinese) texts. In fact, we use it as a kind of portmanteau word, blending two or three meanings that are usually differentiated in the oriental languages. That, I think, makes it difficult for us to have conversations in terms such as your quote above. (That's leaving aside losing people like me once you start talking about temperature )

My instinct ("trained" by seeing other relevant and very common words such as ichidanraku in Japanese) is to make a firm distinction between having the next move, having a forcing move and having the initiative. Having the next move means at most controlling the next move. Having the initiative means controlling the game. Controlling the game means you often don't need to worry about having next move.

The concept is not at all strange to us - we have just chosen, perhaps for historical reasons as well as chess influence, to ignore it in go.

To give an example from western sport, NFL football shows it most starkly. One side is literally given the ball to make the next move. After they have their turn, the ball is given to the other side. They then have the next move. It is such a stark turnaround that different squads of players are brought on (defense and offense) in each phase. In the early part of the game, each side is trying to score or to prevent the other side from scoring. But as the game proceeds, one side may pull ahead with touchdowns and so have a lead. They then have the initiative in the game.

Imagine a situation where the Patriots are leading 45-3 with two minutes left on the clock. The Jets have the ball. They have the next move. The Patriots have the initiative. They can just let the clock run down.

But that matters a lot more than being a more description. Imagine a similar situation but with the score at 21-14 in favour of the Patriots. The Jets have the ball and so can still tie the game with a converted touchdown, but they have no timeouts. The Patriots have the initiative (the lead), but more than that they even control the next move - they don't have it but they control it. That's because the Jets, with no timeouts to stop the clock, have to choose plays that end up running out of bounds (that stops the clock), and they can't make do with a field goal. They are effectively playing under a handicap. They control the ball but not the game.

It is my belief that AI bots (or at least pros who follow bots) play to have the initiative, and don't really care who has next move. They want sente = control, not sente = next move. Control in go is probably definable not as who has next move, but where the next move will be played, by either side. Until we unravel that distinction in English, I don't think we will make much progress in fathoming what AI is up to, but I'm even more certain that we won't be fully in synch with what oriental pros are saying about the topic.

Click Here To Show Diagram Code

[go]$$W
$$ ------------------
$$ | . . . . . . . . .
$$ | . . O O X a . . .
$$ | . . O X X O . X .
$$ | . . O X . . . . .
$$ | . . . X . b . . .
$$ | . . O . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]

My understanding is that the computer dislikes this positions because white can play a or b as forcing moves and there is little downside to this even when white doesn't follow up with anything. Both of these moves were explored by human players for years but at least a was considered (or taught) to be a negative and b seems to have been rarer.

Click Here To Show Diagram Code

[go]$$W
$$ ------------------
$$ | . . . . . . . . .
$$ | . . O O X X . . .
$$ | . . O X . O . X .
$$ | . . O X . . . . .
$$ | . . . X . 1 . 2 .
$$ | . . O . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]

For comparison in this case if then because it is not so easy to cut.

The difference between the connection and the connection under is rather small when you check with the computer and I have seen the computer recommend the connection in situations when the forcing moves don't make sense. What is more surprising is that there have been people playing the connect under move for years but not so much on the top pro level. Even more surprising is that the decision often doesn't seem to affect the next many moves much. Is this ultimately how AI makes us lazy? We seem to chance one move and magically have better percentage without really changing how we continue from that point.