Life In 19x19

The following is a tiny sample of AI charts of complete recent games. The y axis is the win rate and the x axis the number of moves. White is blue and black is black.
The first chart is what I regard as very typical of pro play - perhaps the most typical shape. This is a posh tuning fork: a smooth handle, a decorated Celtic-knot centre-piece, then the actual fork tines.

Example 2 is another common shape. It still shares the tuning fork characteristics.

Example 3 is a rarer shape, and is usually associated with the very best players.

All the shapes, including those not shown, share the characteristics that White starts off with a lead, and the shape of the two graph lines in pro play tends to be parallel tram lines for up to 50 or 60 moves at least. At some point, a crossing of the lines usually occurs, representing a bad move by one player. In fast games, multiple such crossings can occur. In top-quality games, often just one or two crossings may occur in the first 100 moves or so, but sometimes, and especially in fast games, we get a real Celtic-knot formation and the lead can change several times. My impression is that the range of such crossings/mistakes in such games is about 5 to 15 per game.

Then at some point, usually late in the middle game a major crossing/mistake occurs and the tuning fork tines emerge. At this stage, it seems highly characteristic of pros that the tines remain crystal clear: pros know how to win won games.

What can we infer from this?

My inferences are that pros already have a very good handle on fuseki play in the AI style. Their weak point is the middle game. No surprise there, but two points emerge. One is that they don't actually make very many mistakes in the middle game, and the number of major mistakes is tiny. The second point is that a move is, in practical terms, only a mistake if the opponent punishes it. So, when a crossing occurs, one pro has made a mistake but the other pro has recognised the mistake. For pro-dom as a whole, therefore, the "knowledge" of best play is there. We can also infer that endgame play is close to perfect. When major mistakes occur at this stage, we can probably reasonably posit tiredness or time trouble as the first explanations to occur. Complacency may also come into it, but lack of the requisite knowledge would seem to be the least likely factor.

In short, it seems that pros are already very, very, very, very good. And have been for a long time. AI charts of old Edo players, even where skewed a bit because of no komi, have so far shown a high level of play overall. Quite a few of the famous moves such as the ear-reddening move or Jowa's ghost moves have been called into question, but the general standard of play has been shown to be very high. Of course bots are superior to humans as regards results. The nature of the beast is such that the pro only needs to make one mistake against a bot to lose. But we must not overlook the other 99% of his moves where he held his own.

Since human pros can therefore presumably tell other humans what is going on 99% of the time and bots can only tell us 0% of the time, the obvious conclusion seems to be that it's more beneficial for amateurs to read human commentaries than it is to mess around with bots.

I happen to believe that, but for perhaps a surprising reason. I'm not sure that human commentaries tell you all that much directly. I think that the only way to improve significantly is to put in the hard work of playing over masses of games. In other words, to build up intuition. So what we need first of all is motivation. Fan-oriented commentaries, or even stories about pros, can provide that, as can other modes of study of course, but is no guarantee. Personal drive is still required.

But commentaries/books do provide important elements to those who do have the drive to study. Above all, they can provide clues about what to look for and how to avoid bad habits (or bad intuition). I think the most obvious bad habit among amateurs is playing by shape. This covers a lot of areas: katachi, honte, thickness and sabaki among them. Too many amateurs preen themselves after making a good-shape move or a honte as if they are playing like pros. Which, in a way they are, but it's posturing and such posturing was already being derided in Japanese senryu in Edo times: この味がなどと上手の口を真似 "Look at this aji, he says, imitating the way strong players talk."

What a pro can do, in a book or a commentary, is to point out that there are "fake hontes" (as Hane Naoki and/or Takao Shinji have done in recent books). Or they can get you, as in Korean books, to change the posturing "this is good shape" to "it's good shape but is it good haengma?" You can even get certain people to tell you that sabaki is not really about "light and flexible shape .

What we are really lacking, of course, is a regular supply of pro commentaries on modern games in English, or technique-oriented books such as those Bob Terry used to like to do. It is understandable in these circumstances that many people turn to AI bots (although I do think that there is also a lot of bling involved). But that doesn't, I think, obviate my over-riding point that human pros do know what they doing virtually all the time, yet are being consistently underestimated.

(Note: the above is based on my impressions, not on any statistical analysis.)

I think you are right that the point is how to win won games. This is winning percentage. However, looking at points lost, you might not be able to tell.

Shin Jinseo is on another level, but even in a relatively peaceful game between top players in the Chinese A league, total points lost per game often goes over 50 points. But perhaps it isn't a good judge as they often fight. For myself, losing 100 points is a good game.

I think that AI games often repeat the same openings. It is good for studying Go logic, but the mistakes humans make lead to more interesting variations to test yourself against.

As for what winning a won game means in practice, I think the key point is to focus on clear territory and defending your own weak groups even at the cost of bigger potential with centre attacks. Sometimes this leads to terrible shape, or takes unnecessary gote, but normally such loses at most 3 points compared to best since you focus on prevent any aji for the opponent inside your territory and may tempt them into playing too close to your strong shape. Even if the AI wants to get something in the centre or attacking the opponent, perhaps that requires more reading to make work which means more potential for misreading. When you secure your own groups, often you can do so with smaller moves but very clear life. When trying to kill, you must read all around if you are strong enough, including not letting the opponent sacrifice. The focus is on playing moves that are closer to secure territory rather than potential (double attacks etc.) that takes more moves to secure even if a large area could be obtained if successful.

I think the number one principle for the difference in play when you are winning (other than the more important normal ones) is:
Minimise what your opponent can get from playing two moves in a row.

This is just in case they get a double attack or you misplay or forget how to play.

If I recall right, even Go Seigen once stated that he believed he could defeat God with a 2-stone handicap. In other words, he estimated that pros weren't so far from perfect play. Not him, at least.

This has been proven, more or less, in this time of AI, in which pros always lose to AI by small margins (the thing is designed to win, not to brag), humans seem to be behind but not too much. The point, IMHO, is not that AI are a lot stronger, but a lot more consistent. That's why a single mistake can mean defeat for humans.

About why pros seem to be understimated (I don't think they are), let me offer another comparison: cycling. Tour de France, Giro di Italia, Vuelta a España. These races take 3 weeks, 4000 km and 100 hours. Rough numbers. The winner makes just a few minutes less than the runner-up. Minutes after lots of hours. If any of us were to do this, in the end we would say "we were together". But in a competition, those few minutes matter. And they matter because there is no way to make bigger differences. Because in the end, runners are so even that this couple of minutes mean difference between victory and defeat.

In human (professional) go, it's the same. Many games are determined by small margins, even 0.5 points. Komi had to be created because playing the first move was a big advantage, and it had to be adjunted from 5.5 to 6.5 points because that point made a difference. What I am trying to say is that, although pros have a very high level of play (even the weakest pros are very strong), there is something else, some non-human players that can play still better, and get a few extra points. And it's not that the gap between humans and AI is very wide. It's that it's very hard (impossible) to jump. And like in any competition, only the winner matters.

The last post's analogy backfires in my mind. It says that small margins are not indicative of small differences in strength. Anyway I don't think cycling can serve as an analogy for the question at hand.

Haven't pros played AI with handicaps yet? Didn't those matches indicate a difference between 2 and 3 stones?

Human pros are extremely strong. Stronger than amateurs can fathom. Borrowing from another sports, I recently saw a top tennis coach, who attained professional level himself, address the question whether a high ranked amateur would be able to beat John McEnroe, given his old age and dito style. And the coach said: you have no idea how strong pros are. Mac's style may look odd today but he has served a million times in his career and struck the ball a few million times more.

So even if by today's style, which has been changed considerably by AI, especially in the opening, old pro play looks odd, it certainly doesn't mean that any AI bred amateur of today would stand even the smallest chance against them.

Knotwilg wrote: . . . Haven't pros played AI with handicaps yet? Didn't those matches indicate a difference between 2 and 3 stones?

Human pros are extremely strong. Stronger than amateurs can fathom. Borrowing from another sports, I recently saw a top tennis coach, who attained professional level himself, address the question whether a high ranked amateur would be able to beat John McEnroe, given his old age and dito style. And the coach said: you have no idea how strong pros are. Mac's style may look odd today but he has served a million times in his career and struck the ball a few million times more.

So even if by today's style, which has been changed considerably by AI, especially in the opening, old pro play looks odd, it certainly doesn't mean that any AI bred amateur of today would stand even the smallest chance against them.

Okay, and this also is a response to the OP: from what I gather, the correct way to determine the level of pros using AI isn't to watch winrate and attempt to interpret them as a scholar interpreting religious scripture, but for the top pros to play lots of quality two and three stone games against the top AI.

What we need are the top twenty pros like Iyama Yuta and Shin Quadruple-Crown to play ten two-stone and ten three-stone games, half with kata-go and half with other AI' and we can see exactly which moves make such a thing possible as a top pro losing a three stones.

Also, one reason why the ancient masters are not behind modern pre-ai pros at all in the opinion of ai is that much of the so-called progress in the opening theory since that time was not really progress at all but the invention of opening traditions. Shin fuseki proved that. When go was transferred from China to Japan, a bit of strength may have been gained but also a lot of pretentiousness. Is it historical destiny for a Chinese player to be one half of the revolution to get things straight, at least such that we must forgive it was basically the go hippie movement before the hippie movement.

Okay, either this is the first time in my life I doubted a pro sportsperson on current theory (discounting pro opinion that has changed like in the case of the post-alphago era), or more likely he was being polite, not wanting to imply that he could beat McEnroe today. I'll believe he could beat a high-ranking amateur when he actually does it.

Elom0 wrote: Okay, either this is the first time in my life I doubted a pro sportsperson on current theory (discounting pro opinion that has changed like in the case of the post-alphago era), or more likely he was being polite, not wanting to imply that he could beat McEnroe today. I'll believe he could beat a high-ranking amateur when he actually does it.

Here's the video: https://www.youtube.com/watch?v=i_1Uc0y3ySI

Knotwilg wrote:Haven't pros played AI with handicaps yet? Didn't those matches indicate a difference between 2 and 3 stones?

I am not sure how many handicap games there have been. Back in maybe 2018(?) Ke Jie played a games with 2 stones and regular komi (for white) but this is similar to a reverse komi game for one rank difference (-5.5 komi typically with Japanese rules). He lost and I think the reason for the strange handicap was that the program only supported a fixed komi.

I assume most pros would be losing almost all the time at two stones. For one thing there really aren't any wins by pros except that one magic move but on the other hand it is quite a big handicap when genuine mistakes are rare. Only if there were a tournament with pros and computers in which they changed the handicap every game until the pros started winning. I remember there is a book in English that has handicap games between top pros of the time, some sort of an experiment in how much handicap is worth, I wonder what it would take to get them to do that experiment with a program.

I think talking about strength is missing the point.

Sugar Ray Leonard was one of the greatest boxers of all time. But he would never have been expected to beat Mike Tyson, because they were in significantly different weight categories. Similarly, AI bits have significantly different computing powers.

But I could easily believe Leonard knew more about boxing than Tyson.

I am speculating that go pros can be considered close enough to bots as Leonard was to Tyson so that we may reasonably take their opinions on go as superior to those simply implied by inarticulate bots from their stronger play.

Many people seemed to have turned away from human pro commentaries in favour of playing pinball with bot programs. That I would regard underestimating what human pros have to offer us. I admire Tyson but would rather listen to a boxing commentary by SRL.

John Fairbairn wrote:Many people seemed to have turned away from human pro commentaries in favour of playing pinball with bot programs. That I would regard underestimating what human pros have to offer us. I admire Tyson but would rather listen to a boxing commentary by SRL.

I think it's the high availability which makes the difference rather than the strength. If I could review my games instantly and for free with (fill out with any pro) then I would definitely favor that over AI.

But I agree to your point that pro commentaries of pro games are probably more valuable than AI reviews, although I'd still like to combine both to see where pros deviate from AI favored moves, or their positional judgment does.

Knotwilg wrote:The last post's analogy backfires in my mind. It says that small margins are not indicative of small differences in strength. Anyway I don't think cycling can serve as an analogy for the question at hand.

I was not trying to compare cycling with go. At least not too much. I was just trying to point out that in a competition, even if the victory is decided by a small margin, (we) people will praise the winner, and say that the difference with the second is big, even if it's not. But that difference, if can't be closed, is there.

I don't think that professionals are being underestimated. I certainly haven't changed my opinion about their advice. When I have, or take, the opportunity to listen to them, then I generally enjoy listening to their advice, and generally take away two or three points from that. If I just want to review my game with a robot, well I won't particularly enjoy what they show me, but I will try to answer a couple of questions about my play with their responses.

John Fairbairn wrote:I think talking about strength is missing the point.

Sugar Ray Leonard was one of the greatest boxers of all time. But he would never have been expected to beat Mike Tyson, because they were in significantly different weight categories. Similarly, AI bits have significantly different computing powers.

But I could easily believe Leonard knew more about boxing than Tyson.

I am speculating that go pros can be considered close enough to bots as Leonard was to Tyson so that we may reasonably take their opinions on go as superior to those simply implied by inarticulate bots from their stronger play.

Many people seemed to have turned away from human pro commentaries in favour of playing pinball with bot programs. That I would regard underestimating what human pros have to offer us. I admire Tyson but would rather listen to a boxing commentary by SRL.

Well it's true too many people think they can just use bots. You have be either a strong Dan player, or if you're weak at least be on of the rare people with a natural intuition for AI. I personally like to glean very general stylistic notes from AI, but I leave strong opinions on specific positions for pros to interpret from the AI.

I said before that I did naturally like the early 3-3 invasion before bots said it was cool, but at the time I went against it because pro opinion at the time said it was wrong. Now I think as an Amateur I should just play what I like in the opening and focus on the middle-game and endgame, at least I think that's what amateurs should focus on. Before, pros would leave the corner sequences unsettled in the opening and then try to settle the overall board. AI, and my natural style, is to settle the corners early then in the middle game leave the entire board unsettled.

The early 3-3 invasion is still wrong if you hane and connect at the end. Before alphago, pros didn't know about the new 3-3 josekis.

It is not about pro versus ai.

They complement each other beautifully.

Maybe sometimes pros like to experiment, even if they have a system/style. How else does one learn?

In my opinion, pros giving Go lectures have long had the weakness that they can't really justify what they are saying, or they use overly fuzzy words, which means that it takes a lot of other resources to get strong at Go. Frankly, a significant percentage of the time they try to use logic, they aren't quite right (but they were the best we had) and too often completely wrong. I think that they are trained to make reasonable decisions under time pressure, and logic hasn't necessarily been useful enough compared to intuition for that. Even after the appearance of AI, it seems pros don't have a great handle on explaining AI (or rather deeper Go) logic, but it hardly costs them many points. Though top pros do seem to have progress in predicting AI ideas with their moves. Perhaps explanation is partially a job for mathematicians.

However, I am impressed that so much of opening theory, including each revolutions over the last few hundred years all the way up to active play of the Lee Sedol era (pincering rather than defending), Chinese opening etc. seem to represent real progress. I had long thought this was the most difficult part of Go to justify, and yet humans seem to have secured much progress here pre-AI. At the same time, I think that from the right perspective, Go isn't as hard as it is hyped up to be (at least not in 2d), so perhaps developers of this theory were able to get this sort of insight and combine it with their experience and strength.

The AI is at the pinnacle, so it may be blind to understanding of bad play. Humans are more useful there, for now. Perhaps that is why so many Go terms sound like insults ...