Impact of Go AI on the Professional Go World – Response

Published 25 Aug 2020 by antti (last edited 25 Aug 2020)
tags: tldr

Lee Hajin recently published an interesting article on the ‘Impact of Go ai on the Professional Go World’ on medium.com. I found myself disagreeing with many of her points, and so decided to write my own remarks on the topic. Note that my intention is not to attack or ‘bash’ Lee’s opinions, but rather to bring forth another point of view.

Lee writes on three areas of the professional go world that go ai has affected:

  1. go as a career or life path,
  2. professional players’ race to learn, and
  3. the demand for professional teaching.

Lee’s first point is that with the advent of go ai, the way of reaching for the ‘ultimate level of play’ has shifted from being a lifelong inner search, not unlike that of a philosopher’s or monk’s, to study involving computers. My interpretation of Lee’s message is that the ai’s surpassing of humans has demystified the game, making it something that can be measured and optimised like computers do, and that this is a bad thing.

Lee recalls that when she was a student of the game, she lived at her teacher’s place with other students and they studied and learned the game together. This is called the ‘uchi-deshi’ system (literally ‘inside student’, meaning students who live at their teacher’s place), and as far as I know, it was also common among Japanese professionals until the late 1990s – the Kitani dōjō is probably the most famous example. In the 2000s, it seems that the system fell out of fashion earlier and not for reasons related to the ai. Possible reasons for this include people starting to have fewer children (and therefore not wanting to ‘give them away’), daycare services becoming more common, and child welfare law becoming stricter.

Two sentences in particular piqued my attention: ‘. . . no one questioned that go was a path you walk for a lifetime’ and ‘The belief was, as a professional player, you explore and endeavour to reach an ever higher level of understanding.’ Lee seems to combine these two with the silent assumption that the professionals’ goal is to strive for the ‘highest possible level of go’, which is no longer possible because the ai cannot be beat; incidentally, Lee Sedol remarked something similar when he announced his retirement. To me, on the other hand, it seems like nothing has changed because I have always reached for the ‘highest possible personal level of go’, and this should be the same for most players who are not near the top of the world.

Lee’s second point is that the advent of ai has reduced the variety in personal playing styles; that, ‘today, everyone is trying to imitate the ai style, and the pros judge each other only by who is better at playing like the ai.’

While I agree that ai-inspired moves are now increasingly common among top players, I predict that the situation is a bit more like if Ancient Greek mathematicians were given a graphing calculator: of course they will first lose themselves in playing around with the new toy, but eventually they will get used to the new possibilities and life returns to a slightly more normal state. For top professionals, it is valuable to be aware of the newest developments (or ‘meta’) so that one can on the one hand avoid larger mistakes, and on the other save valuable time during tournament games. Besides top games, I suspect the ai has actually made more styles of playing possible. My kifu newsletter subscribers have read some of my research regarding the (surprising) viability of the sanrensei and Southern cross opening, and the third-strongest ai on the Computer Go Server plays the Black hole opening, of all things. And, of course, the ai has the great added benefit that it makes it possible for players not from Asian countries to possibly someday become the world champion.

Lee’s third point relates to teaching pros; she remarks that ‘the demand for pro-level teaching games and private lessons has plummeted.’ I can see this being the case especially in Asia; and, in fact, some teaching professionals may now find it difficult to even make a living from the game. The sad thing is that this should not be so, because good teachers still have a lot to give what the ai cannot. While the ai may point out your mistakes and show you better moves, it cannot tell you why the other move is better, or teach you higher-level concepts that make it possible for you to avoid similar mistakes in the future. My teacher remarked that people who learn moves directly from the ai are ‘studying’ the game similarly to if they were memorising jōseki, and I generally agree with the notion.

Personally I have found one new challenge to my job as a go teacher, which many readers are familiar with from my other posts: that of ai-related cheating. There have been many reported cases of cheating in online study leagues, and I even know of cases where a student used an ai in their games against their teacher. As this kind of conduct is directly self-harming – you are paying for doing a disservice to yourself – it is clear that its motivation lies somewhere else than learning. As for whether it is in perceived social status (e.g., getting to a higher rank or impressing your teacher) or the joy of winning a game, even if undeservedly, I cannot possibly say. The main threat I see in this conduct is that it seems to devalue go as a pastime altogether: what should have been an intellectual pursuit in learning is becoming something more like ‘press a button to win the game’.

In fact, recently it seems like more and more people have a natural dislike of anything arduous or difficult. I was reminded of this the other day when I was listening to The Brain by David Eagleman, where he explains how the human brain has evolved to conserve as much energy as possible (as, even in its current state, it still takes up one-fifth of the body’s total energy consumption). Quite possibly as an extension of this, during my training games I sometimes get an urge to check what the ai would do in that situation, so that I could avoid having to do the hard work myself; and I would be surprised if I was the only one who experiences this. How I cure this is by reminding myself why I am playing in the first place.

To me, superhuman go ai has changed nothing about the reason why I play go. The game is still an intellectual pursuit and a journey of self-betterment, and the ai is merely one more tool that one can use. However, I am also a go teacher, and as a teacher I am getting increasingly worried about the laziness-inducing side effects of the ai tool.


Comments (9)


yakago wrote 1 month ago:

This gives a perfect opportunity to invent some 'ancient wisdom story'

A student and their master are seeking the divine path to enlightenment.
They are dilligent in their search, and are always disciplined, for they know that this is a goal worthy of devoting their lives.
One day, much to their surprise, a light appears and the divine path is revealed before their eyes.
The student proclaims in despair "It is over, we have nothing left to do, what are we to do with our lives now that we have found the path?"
The master answers calmly "Now we must walk it"


Farodin wrote 1 month ago:

What I've wondered about the Black Hole AI though, is whether it's winning games by using the strengths of the Black Hole fuseki, or whether it's winning games 'despite' using this fuseki. I'm not strong enough to analyse this (EGF 3kyu), but I can't rule out the possibility (maybe someone stronger can) that the first 4 opening moves are hard-coded in the system, and that the rest of the game is played with a LZ or KataGo instance that is backed by insanely beefy hardware, which allows it to outfight the opponent regardless of what fuseki it uses.

When I look at the Black Hole AI games, I don't really see a tendency to use the 4 opening stones to carve out a big territory in the centre. Instead the AI still really likes to take the opponent's corners away, which to me seems like it's wasting the value of its center stones.

On another note, I have noticed the laziness to put the work in when it comes to my own game reviews, so I'm tempted to not use the AI at all and study in the more traditional way for a while.


MarcelGruenauer wrote 1 month ago:

Two more aspects.

First, in the e-Go Congress Opening discussion - https://www.youtube.com/watch?v=-cEL7I6BWTc - at 56:07, there is a quote from Lee Sedol 9p: "Professional Go players, in old times, they study Go and invent something that only belongs to themselves. Now Go players go online to learn new things."

It sounds to me like some professional players considered themselves to be the Keepers of Secret Knowledge and now Tom, Dick and Harry can use AI to gatecrash their monastery.

Second, at 1:33:00, Lee Ha-jin says that she believes that Go is moving more towards being a mind sport and less of an art form or philosophy.

But why is the art or philosophy diminished if the answer comes from a machine instead of from a strong human master? If, instead of AlphaGo, a human started playing the same way, would we not admire this human master? Also, you can still appreciate the art and philosophy of Go if you study and play the new way.

Maybe what some people find disappointing is not so much the superior way of playing itself. Maybe it's that the top pros, the ones closest to the truth, were elusive; you couldn't just go to and ask them all the time. Now AI is available to everyone, so part of the magic is gone.


lightvector wrote 1 month ago:

Regarding diversity of styles of play - I would also hope that before too long things will settle down, with many players discovering their own varied styles of play, and daring to return to play openings that the bots would never play and even other new ones. It is very common in the opening in Go for there to be multiple moves that the bot dismisses out of hand, yet when actually set on the board, are barely a percent or two worse even at the superhuman AI's level. It's hard to see why such moves would not be playable. And perhaps even lead to an advantage rather than a disadvantage, if they are also unusual enough such that player playing them has studied them more thoroughly than their opponent.

The Chess community is one that has had to grow in the presence of super-human AI for a a fair amount of time now, and it is well understood in that community that there is a difference between the moves that a computer would play, and the moves that you yourself should play. A move that you understand and that fits your style can be superior, for you, to the move the computer wants to play. And its top player, Magnus Carlsen, is even known for deliberately playing minor "inaccuracies" in openings specifically to get out of the space of memorizable openings and back into wilderness of Chess where both players must improvise and judge for themselves.

Anyways, thanks for the post and the links, and initiating such a discussion. :)


antti wrote 1 month ago:

@lightvector: My pleasure!

@Farodin: The Black hole AI is interesting in that, while all the AIs seem to agree that the opening is suboptimal, in the later parts of the game the Black hole AI can often turn the game around by its superior life-and-death understanding, killing the opponent’s invading group. In other words, the other AIs don’t give the influential positions of the Black hole AI the value that they (possibly?) should be due.

While I would be really surprised if the Black hole turned out to be the best opening out there, or even comparably good to current fashionable openings, the existence of the Black hole AI does raise the interesting question of how much the evaluation of the other AIs should be trusted.


kyyni wrote 1 month ago:

As Antti pointed out, the Black Hole AI (BHAI) raises questions about the validity of mainstream bots' positional judgement. For me, it seems not to be too far-fetched idea, if there really was this kind of issue with mainstream AI:s (either with unsupervised learning, like Leela or AlphaZero, or supervised like AlphaGo Lee/Master).

I'll state a hypothesis: both the human professional tradition and the mainstream AI:s play corners-4/3:th-line-first in principle because it has been easier to understand (not preferring to anthropomorphize, in case of AI:s, of course in a metaphorical sense). With supervised AI:s, they are obviously bound by human tradition into this schema. With unsupervised AI:s, starting ab initio, they too find themselves into the same schema. Starting with random play, they can indeed quickly converge into playing corners-first: in the beginning phases of the self-play iteration, with almost no (metaphorical) ability to plan ahead a whole board strategy, the kind of network at least makes some easy cash at the early game is relatively strong. And so the network weights get heavily skewed towards corners-first, in such a way that no iterative process thereafter can escape beyond that (possibly local) optimum.

I should add that even in the case my hypothesis holds, and in the light of *current* understanding of Go and the track record of bots to date, it cannot be claimed that corners-first is necessarily very far from theoretically optimal strategy, but just either not strongly favourable (e.g. BHAI corrently being at least second to KataGo), or slightly suboptimal. If it were seriously suboptimal, even self-play iteration within the AlphaGo-Zero-like framework would eventually divert from corners-first.


lightvector wrote 3 weeks, 6 days ago:

Two followup thoughts:

1. I don't think anyone should conclude much about anything in openings from the Black Hole bot on CGOS. I think it's extremely weak evidence of anything, so weak that it almost shouldn't deserve a mention - you should mostly discuss and think about the situation the same way as if this bot didn't exist at all.

Why? Because play on CGOS is not well-controlled. We know that KataGo can give *itself* at least 2 stones merely if you have a large enough difference in playouts. So if someone wanted to throw a bit of hardware at the problem, it should be possible to reach the top of CGOS with a "2-2" or even a "1-1" bot that opens on the 2-2 point or the 1-1 point, using simply KataGo itself. The task becomes yet easier if, e.g. you're using one of the few closed-source bots stronger than KataGo, such as if you've managed to get a hold one of FineArt's networks, then the hardware advantage you'd need for a 1-1 bot would likely be smaller.

If someone were to do so, then taking it as evidence that the 2-2 or the 1-1 point is a good first move would of course be ridiculous. Unless you in fact know the Black Hole bot and its hardware, the same applies to that opening as well.

2. My experience and observations with AlphaZero-style training make me feel that kyyni's hypothesis is much more likely to be false rather than true. Keep in mind that in the AlphaZero process, the networks is trained using the final game outcome, rather than, e.g. the evaluation of the limited lookahead of the MCTS. You can never get into a local optimum as simple as "I overvalue the corner, therefore every time I evaluate the board, my MCTS overvalues the corner, therefore I train my net to further predict this, and therefore the bad evaluation gets reinforced". You *can* still get stuck in local optima, but they have to be for reasons that are more complex and subtle.

So don't need to ever think that a shape is better to learn it is better, you just need for it to lead to a better result any time before the end of the game. Even if it only pays off in 100 moves, 100 moves is still before the end of the game. This, combined with massive training data, is why bots are so fantastically strong at the judgment of extremely long-term influence/territory tradeoffs and effective center play compared to human pros.

Another way to think about it: the same way that KataGo plays some small proportion of its games on 9x9, it plays some small proportion of its games with the first move on the 5-5 point, or on the sides, or in the center (there are opening randomization mechanisms). So just like "9x9" is a separate game that strong play is learned for, "19x19 with black forced to open on 5-5" is a game and KataGo learns how to use the 5-5 point well... and it still doesn't get good results.

Perhaps maybe the optimal move is the 5-5 point, but such a move is quite bad unless you very exactly play the 6-7 point in the next corner and then play the extra-large knight's approach from the opponent's 4-4? So even though each such move on its own followed by "normal play" has statistical evidence showing worse results, a succession of them together becomes good? Sure, this could be possible. Still I don't see any particular reason to specially elevate this kind of hypothesis at the moment if there's no more evidence for it than for it than for less-demanding hypotheses.

Lastly, one thing that makes me think that it is more likely that 4-4 and 4-3 are actually right is that KataGo's preference for 4-4 and/or 4-3 emerges already on 10 x 10, 11 x 11, 12 x 12,... (and KataGo has explicit training on these board sizes). This rules out some flavors of the above kind of kind of hypotheses. Bots are more than capable of learning that tengen = 5-5 is one of several good openings on 9 x 9. Size 10 and 11 should not be not so much bigger as to hamper the learning of effective combinations of higher-than-4th-line moves, if they exist. Instead, the fact that 4-4 and 4-3 are preferred is suggestive of the simpler explanation - 4-4 and 4-3 corner moves are simply optimal-efficiency moves, and 9 x 9 is perhaps the last board size in which a more central first move is better, 5-5 being okay on 9x9 because it influences all the corners at once.


antti wrote 3 weeks, 6 days ago:

@lightvector

The point about the Black hole AI on CGOS is fair. I had just silently assumed that the bots on CGOS were paired against each on similar hardware and time settings. If this is not the case, then, as you also said, throwing more hardware at the problem can easily be enough to twist the results.

It could be mildly interesting to have the Black hole AI network have a long series of ‘fair’ matches e.g. against KataGo, mainly to see how big of a winning ratio it could then manage, but it does not seem like the network is publicly available.


kyyni wrote 3 weeks, 5 days ago:

@lightvector Very good points. To be fair, I don't regard my hypothesis very probable, it just seemed to me something interesting and not entirely unthinkable. Indeed much more probable and boring explanation would be if BHAI just had OP hardware.