Having super strong go-engines has definitely changed how the game is played at top level. But how?
I watched a youtube video on subject and from there I got understanding that some basic josekins like san-san invasion are perhaps not the best options. Keeping sente is more valuable than previously taught. Are the other finding and do they have any impact on lower echelons of go players?

Hwang In-seong 8 dan's 2022 Congress lecture series vol.1 [Four Trendy Ideas]
https://www.youtube.com/watch?v=QgUnD434NH4&t=855s

I have not studied the impact on top players but have experienced a wall of those 5d+ opponents with AI study I could not beat before my AI study started. AI study has kept me busy for 9 months now and I notice an impact on my level.

In the 90s, studying the opening from books and pro games was one factor for the 3k to 3d part of my fast improvement. However, later as a 5d I realised that I did not really understand the opening as taught by old books and pros. The more traditional opening theory I learned the more I saw nothing but contradictions. Eventually, I felt that I had no useful opening knowledge at all. I saw stronger players playing stronger opening but did not really understand how to play alike. With no traditional source that could improve my opening, I was stuck.

Therefore, my AI study has been mostly the early stage of the game. What I hoped to learn from AI I can. AI confirms my prior impression that all (ok, maybe 99%) traditional opening theory is useless. In stark contrast, AI opening makes sense and is consistent. It is not always easy though because the exact position can matter, deep tactical reading can be necessary, fights can be correct or very long term developments can be correct.

I presume some top players would also do similar AI study and it is an easy guess that they might draw similar conclusions. Such as: it is insufficient to play just efficiently - it is mandatory to play the most efficiently!

Many observations about AI Go are only in the eye of the beholder. The computer, unlike humans, will do the same thing on every move. If it protects a cutting point that is not because it thinks the alternative is too much, it is something it basically does after considering the alternatives seriously. It is the same with sente and gote, just as the compute will play marvelous sente sequences it also plays astonishing gote moves. A human player can't always do the same thing, they will make choices based on what is good enough and if they can better spend their energy on some other decision.

I think lot of people should pay more attention when it plays simply. If it simply connects, simply lives, simply defends and simply ends in gote, then that is something everyone could have done

One thing I have noticed is that if you look at human games from the past that the computer often thinks there is something fishy. There are weaknesses that it finds and wants to exploit. Sometimes the players of the game have the same idea but don't jump on it as quickly and often the defense is successful. Were as if you study the game with the computer it is clear that the attacker could have succeeded; the computer most of the time is silent about moves that don't work. However, if you look at computer vs. computer games there can be a ton of simple moves and both sides seem to have been ready for everything. The simplest explanation is probably just that the computer is much better at many aspects of the game than top pros were and if a much stronger program came along it would give old computer-computer games a similar treatment

I recon than that the top players have learned a lot from studying with computers but exactly what might be something they have to tell themselves

His section about tenuki is wrong. He says that tenuki is always an option and the whole section conveys this idea but it is wrong. He should have said that AI tenukis more often than previously was common.

For many moves, AI has a clear choice on which is the only correct move, and it is not always tenuki.

Like Takemiya, he fell into the teacher's trap of telling the pupils that they can do what they want. Instead, he should tell what is correct: when to tenuki and when not to tenuki. Of course, this is difficult for everybody including the teacher.

I wonder if what they have learned might be mostly improved assessment and understanding of certain specific positions rather than new concepts and theories. In which case, the SL page on 'Missing Concepts' is sadly still going to remain fairly empty.

Probably many small things and that is just my observation. Possibly they have learned so many things about shape that quantity led to a change in quality. Even that could be something that started before AI, but top pros do appear to play more solid shapes then they did couple of decades ago. That is solid as in it is harder to get a defensive move from the shape. They also appear to have learned many things about soba go

It's things that you'd never have thought top pros could learn anything new about.

Or it could be the other way around. Maybe they did learn how to evaluate some positions better and this led to a change in how they play other aspects of the game. On the other hand, it could be the seed that fell on a good soil, those that found it remarkable how AI evaluates some positions then learned something entirely different. Which is how it plays up to those positions.

Something that is usually overlooked could be more important. It is that when you play something novel and it doesn't work out your opponent and your so called friends (because we don't have anything better) will never let it go and trash you and your move for years to come. And if it works out it is all the same

I don't remember his name but there was a pro that always played a normal approach move to Kobayashi fuseki. He didn't see what it was he was supposed to be avoiding. Which makes sense when you can find games were Lee Changho(?) only managed to punishing it because he is Lee Changho. Nowadays, he could just have showed them! That is showed them how AI evaluates the move

Remember we've been through all this before with Shin Fuseki. We are still unsure what pros learned from that. We can be sure amateurs learned nothing from it despite books selling over 100,000 copies. But those who did learn a lot from it were the publishers, sponsors, media folk and the like. Bling, bling, kerching, kerching.

As Robert says, openings are the obvious change: new variations for the 3-3 invasion, re-evaluation of "old" joseki, rehabilitation of shapes we were taught are bad, reduced popularity of high concept openings, ... What can we say about the middlegame? I think there has been a real influence there too, but it's harder to describe in a few words.

To my mind, a big factor is that the new AIs are very good at responding to an attack. By this, I don't mean "defending". For an amateur like me, if my stones are threatened with attack, my first instinct is to try and keep them alive. But there's a lot of other options. Sacrifice and build influence. Make a light shape and allow the possibility that part of the "group" is cut off. Tenuki and do something more important elsewhere. Of course we already know this. Most of us learn "light and heavy" some time in the single-digit kyu ranks. But knowing the concept and finding the moves to put it into practice are two different things. And I think AI is taking it to the next level. This is why moyo strategies are becoming less popular, and it's so much more common for games to break out into early and chaotic fighting rather than building a framework first.

One factor is that the AI sees the board anew on each move. It doesn't have an emotional attachment to previously played stones. So it stands to reason that it should be more prepared to make trades in a way that feels risky to humans. (Why shouldn't I take risks in a close position? If it's risky for me, then it's equally risky for the other person. But somehow it seems to go against instincts.)

To some extent this is an acceleration of trends that were already happening in the 1990s. So it's hard to pick out how much AI has given things a push and how much it would be happening anyway.

There's two other things I'm noticing when I review games with KataGo. It seems to play forcing moves earlier than was recommended a few years ago. I've always been taught to avoid pushing the opponent around unless I can see a clear benefit. Leave things open, because you don't know what options you might need later. But it's a fine line between leaving it open versus waiting too late and losing the chance entirely. Except that KataGo seems to be suggesting it's not such a fine line: just do it as soon as the chance comes up (usually, not always).

And I think the definition of "probe" has become a lot broader. I see KataGo saying "Why don't you just drop a stone into that area that looks like your opponent's settled territory? It costs nothing (they have to reply), and might provide some useful aji in the endgame." Again not a new idea, but happening earlier in the game and more often than I'm used to.

Sorry I'm too lazy to pull out specific examples from pro games today. I might come back here next time I notice something. Meanwhile, people can tell me where my opinions are wrong, and we'll have another interesting conversation :-)

I'll also go out on a limb and say that AI has not changed the late endgame (the part where you're down to 6-point moves or smaller). Humans already have a good conceptual framework for that part of the game, and in slow time controls I think the top players have been able to get pretty close to perfect endgame play for some decades. Computers might be stronger in terms of being more consistent, less likely to make mistakes under pressure, but I don't think they've introduced any new concepts for this part of the game.

Actually, no. There's enough randomness in the algorithms (partly from a random number seed, partly from races between multiple threads), and enough positions with two or more moves that have similar evaluations, so that monte carlo based algorithms will not actually do the same thing every time. That's one of the reasons why we don't need opening books for go playing engines (in contrast to computer chess).

You make some good observations but I disagree about the following:



I have also paid close attention to the timing of forcing moves: it differs greatly. In some environments, forcing occurs early. In other environments, it occurs late. In yet other environments, there is a long period during which it can occur, but there are exceptional moments when temporarily something elsewhere is more urgent (such as preventing a big cut in a joseki) when the forcing option is interrupted.



While this is so to some extent, closer study of AI evaluations has often revealed that standard developments emerge from initial randomness after sufficiently many playouts. Therefore, there can be AI style opening books if the author relies on enough playouts and study of positional variation.

I think you misunderstand what I meant. If there is enough of randomness is something different. Clearly there is enough randomness to make a formidable Go playing program, one that in my experience also plays the same moves again and again, but certainly not always, and this wasn't what I meant. What I meant is that it will run the same neural network and MCTS based algorithms on every turn, were as a human will often just glance at the position and not go into deep thought at all. This makes for a huge different between Human and Computer Go.

A relevant paper from 2021 that I was unaware of until it was referenced on Facebook today:

Human Learning from Artificial Intelligence: Evidence from Human Go Players’ Decisions after AlphaGo https://escholarship.org/content/qt6q05n7pz/qt6q05n7pz.pdf

in this blog, "After AI beat them, professional Go players got better and more creative" https://www.henrikkarlsson.xyz/p/go

Here are my comments on the peer-reviewed paper Human Learning from Artificial Intelligence: Evidence from Human Go Players’ Decisions after AlphaGo by Minkyu Shin, Jin Kim and Minkyung Kim in Proceedings of the Annual Meeting of the Cognitive Science Society.

The paper contains some useful study and mathematical terms but also paints a bad light on quality control in cognitive science because the formal study contains a few mistakes and the informal text has more mistakes. Some mistakes I mention as follows:

- The aim of the game is not to make more territory. This mistake I highlighted countless times for territory rulesets. The aim of the game, ignoring komi, is to make the larger sum of territory and prisoners.

- The terms strategy and reasoning process are abused. Some more appropriate terms should have been chosen.

- The photos are caricature more than evidence.

- It should have been distinguished between the possibility to AI access and actual AI use.

- Two implicitly used asumptions have been made but not stated: while playing, AI does not learn from human play; while playing, humans do not cheat using AI.

- A set of matches and a set of moves have been used but their independence has not been declared and any correlation of duplicate data has not been studied.

- The stated equivalence ∆k = 0 ⇐⇒ akHuman = akAI is wrong because there can sometimes be alternative actions with the delta 0.

- The speculation that humans could learn less during the stages after move 50 of the game is premature because an alternative explanation could be that humans are at least as strong as AI then. This would have to be ruled out or partially confirmed for some such stages.

- It should have been noted that the evidence for the impact due to using AI "reasoning" is weak because the alternative explanation of using a larger number of AI actions has not been excluded statistically.

In the paper After AI beat them, professional Go players got better and more creative, what is the move decision quality in the two charts?

In the context of this thread and its amateur students of the game, it is irrelevant whether this paper is right, wrong or just flawed.

Centuries of experience have taught us that the best way for humans to be good at something is to spend a long time at it so as to build up a huge database in the brain - one's personal reference library, or one's intuition - whatever you want to call it.

Experience has also taught us that, while time spent may be the most important tactor, there are other factors at play. We know, for example, that it is good to start young. But, as any parents will know, starting does not imply finishing. Think of all those trombones and violins languishing in the loft. Or adults' rowing machines. Psychological and social factors come into play, too. But if a child can stick at a subject, it is likely that starting early has some influence on the way they build up their intuition. With still malleable brains, it seems that whatever they learn can be stored in the most efficient way for that subject. But the way something is learned can have a major effect on how something later is learned - even something similar. The effect of that can be seen best with language learning. Everyone learns their own native language to a very high level. No-one, essentially, can learn a foreign language to that level as an adult.

No doubt natural talents are also prominent factors. For example, in go, it may be that people like Sumire can memorise evaluated chunks of (say) 10x10 points, whereas most of struggle with 4x4 chunks. In the same way that most people can hold 6 or7 items in their frontal short-term memory but a lucky few can do 8 or 9. Possible the size of chunks memorised in go depends on the age you started.

The precise mechanisms or numbers are not relevant here. The important point is that amateurs have clearly not spent enough hours yet to become strong at the game. What they need to become stronger is simply more time at the board. For some, the bling of AI has provided that incentive. They may end up stronger, but it is the effect of the extra hours spent that is giving the improvement, not any fanciful notion of understanding how AI "thinks". If free lollipops for every extra hour spent at the go board were enough incentive for someone, their strength improvement would happen just as inexorably.

Pros are in a different situation. Their internal go databases (intuition) are already huge and probably close to saturation point. But even then we know from people like Shibano Toramaru that they still spend a lot of time looking very quickly at games to note what is new and so potentially worth storing in the brain. That method has worked for them in the past, so why not try some more?

There are, of course, pros who also try to think about theory. But, there, there may not be a single right answer. We saw this with Shin Fuseki. Go Seigen thought the way to achieve improvement was to achieve more speed in the opening. Kitani thought the key was the integration of joseki and fuseki. Both ideas together sound remarkably like whatever it is that characterises AI go - though probably there are other concepts.

One concept that appears to be relatively new in modern human go is the willingness of Korean (especially) and Chinese players to gamble with the odds. Wang Xi wrote a seminal paper on this, pointing out that the Japanese insistence on playing souba go may be what is holding them back internationally. Players of the Japanese school still prefer to choose a safe line that seems to offers a small lead that translates into a 51% chance of victory, whereas Koreans and Chinese have found more success with choosing risky lines that may offer a 75% chance of victory. The much faster time limits of Korea and and Chinese go favoured that style of play, which is now ensconced, and it seems to have given them an advantage in the AI era.

Top pros seem to be able to play games with around 90% of their moves approved by AI. The very best, Sin Chin-seo, are scoring about 96%. The difference between the second-tier pros, Sin and AI seems (according to what is being said in the Far East) to be down to deeper reading. But this does not mean tsumego-type reading. Rather, it seems to be the ability to make EVALUATIONS of positions at deeper and wider levels - encompassing bigger areas of the board. 19x19 chunks instead of 10x10 or whatever. Covering a wider area of the board like this still mainly demands building up one's intuition as the prime way forward (not theoretical understandings) but because of the size of the chunks they are now tackling (i.e. whole boards) concepts such as Go's 'speed' and Kitani's 'integration' are being seen in a new light. Because of the need for speed, which in a sense comes about because you feel the need to be in two or more places on the board at once, more stress is now being put on the initiative (not sente in the usual western sense), and the value of miai (as per Shuei, of course) is likewise now enhanced. My sense of the new approach to miai is that, while you can't be in two places at once, you can at least guarantee to be in one, as per the traditional view. But nowadays you can try to leave aji in the other, and with a bit of luck you can go back there later and so achieve an effect close to being in two places at once. We can be sure the top players are not giving all their secrets way, but these seem to be the things they are talking about from what I read.

But, interesting as all that is for go fans, amateurs who want to improve will probably get much more mileage out of studying go the traditional way and worrying about the bling only once they have mastered that.

[Incidentally, it seems that we can see the emergence of Japanese economic strength a few decades ago as an application of the souba theory (and this is where the term comes from). Traditional evaluation methods based on that, such as candlestick theory, worked well in that environment. But with globalisation (bigger boards!), the risk-loving methods of the West that offer much bigger rewards have come to the fore, and now Japan is going bnackwards, in economics as well as in go.]

It sounds like you've looked into it more deeply than I have. I'd be interested to hear more about your conclusions some time.

If simply this were true, I'd be 100 dan.



Evidence? What means "approved"?



You try to construct a competition between non-AI and AI learning methods but all methods are useful.

The more relevant question is about go theory (other than theorems): only traditional versus also AI. Traditional theory has a limited scope. Also using AI-born theory expands the scope and provides solutions where the former was wrong.

Yes, I was underwhelmed by that paper. There are some more fundamental problems. Correlation does not imply causation! They haven't commented on the most important confounding variable: the passage of time. Yes, their data shows that people play "better" go in 2020 than in 2014, for some measure of "better". But they can't separate "had access to Leela Zero on their own computer" versus "had an extra two years to study AlphaGo games" versus "spent more time learning go by other means".

The pseudomathematical notation on page 1797 is irritating. It's just an opaque way of saying "measure the average winrate drop". Dressing it up in symbols this way makes it harder to read without adding value. (I've seen equally bad things done in music theory papers! It's a way of trying to improve your chances of publication by impressing reviewers with your "cleverness",)

What I'd like to see is:
- Use of historical data as a control. By these metrics, did humans get any better at go between, say, 1950 and 2000?
- Critical examination of the metrics used and consideration of alternatives. When Leela Zero first came on the scene, there was plenty of discussion in this very forum about how to interpret "winrate". And whatever it's measuring, I don't think it's measuring on a linear scale, so taking a simple average isn't good mathematics. Other metrics that occur to me are frequency of a human choosing the AI-recommended move, frequency of "large mistakes" (winrate drop over a certain threshold), and comparisons of outputs from a neural network trained on a database of human games versus the networks trained on AI self-play games.

I do think AI is less of a black box than it's painted out to be. It's possible to have a "conversation" with the AI if you allow the possibility of non-verbal communication, and I believe that humans can and do learn from AI. And I think there's potential to slice and dice datasets in different ways to show evidence of this. But the paper at hand is only a small first step.

Yes, but beware of confirmation bias! It's not enough to ask "What do the successful people do?" The real question is "What do they do differently from others?" For example, the many people you can see on IGS or Fox who've played thousands of games but are still well inside the kyu ranks. They've also built up a huge database, but it's sometimes a case of "garbage in, garbage out". Access to pro game records -- or access to AI -- lets you put something else in.


Now this is interesting. Hard to pin down in a few words, but it starts to make sense of the chaos that is modern high-level go.