I’ve been playing Go for half my life. When I started, in 2005, there was no such thing as Youtube, everybody knew that AI had absolutely no chance of ever approaching human level, and the Kiseido Go Server looked and operated exactly like it does today.
Over the years I have heard a lot about how to study Go, how to improve at it, and how to approach it. I have my own thoughts on these topics, and will describe them here. My credentials: I’m 5 dan in the American Go Association ratings system, and in 2019 I played in the US Masters tournament (though don’t be too impressed - I lost 4 of my 6 games ;-; ).
This essay is one man’s opinions on these topics, and I haven’t surveyed others or looked at related scientific literature. Apply appropriate grains of salt. Advice here does not apply to child prodigies or Koreans.
“Lose your first X games as quickly as possible” is advice for 30 kyus, not you
When you started Go, somebody probably told you to lose your first 100 games as quickly as possible. (This number has undergone serious inflation in certain minds, and I’ve heard the advice is to lose your first thousand games as quickly as possible! This is ridiculous and way too much). The right number is probably more like 30 or 50 games, not 100: just play, don’t review, grind grind grind. At this stage - rank 30 kyu to 25 kyu - it is like you are navigating a cluttered room while being unable to see and unable to hear. The correct strategy is to just stumble around and bump into things to build your mental map. It’s all you can do - reasoning about a space you can’t see or listen to is a waste of time, so just be free and happy and bumble around.
So that’s the common advice, and it’s good advice. But people take it too far. If you are 10 kyu, and have been 10 kyu for months, the advice “just play” doesn’t make sense anymore. You can see now! You can hear! Your situation is very different from that of the blind 30kyu. If you like playing because Go is fun and you get to see your friends at Go club, yes, absolutely, this makes sense, not everything in life is about upskilling, and you should enjoy yourself and have a good time and just play. But if your progress is stalled, and you want to unstall it, “just play” isn’t going to work.
Like, I just opened up WBaduk and scrolled down to the 13 kyus. Look at these Win/Loss counts:
park1229 is at nearly 60,000 games and he’s still 14 kyu. bhch2010, 30,000 games, 13 kyu. And many sitting at above 5,000, or 10,000, games - 13 kyu. (For perspective: if you played a game a day for 5 years, you’d have played less than 2,000 games!)
People who give the advice “just play” are communicating a model that looks something like this:
Here’s a graph of real data from the site WBaduk, with the same axes:
There is no hint of that increasing line from the “just play” model.
This data is fairly confounded (why would the trend lines go down? very strange - and there’s clearly something weird about the range 26k - 18k - where are they? - and I expect most of the strong players have many games offline that don’t show here). But it sure is suggestive: the average rank of people who have played between 100 and 1,000 games is about the same as the average rank of people who have played over 1,000, and even about the same as those who have played more than 10,000 games.
So: playing lots of games, by itself, is not sufficient. Instead, you need to do purposeful practice:
Purposeful practice is practice which (1) has well-defined goals (such as doing something 3 times in a row with no mistakes), (2) is focused (the person is intently interested in improving, rather than having their attention elsewhere), (3) involves feedback, (4) involves getting out of one's comfort zone, practicing things on the edge of one's ability.
This is in contrast to naive practice:
You practice music by playing it. You practice medicine by doing it. You practice driving by doing it. Ericsson and Pool cite studies showing that this is not very effective at increasing, or even maintaining, skill. Doctors are (on average) at their best shortly after getting out of medical school, and gradually decline in skill thereafter despite their continuing "practice" of medicine. Similar observations apply to other skills.
“Just play” is naive practice. Ericsson and Pool surely have many good examples in their papers and book, but think of examples in your own life:
When is the last time you felt you really improved at driving? If you’re like me, in your first year or two behind the wheel, you got much better at driving just by driving, and since then, you’ve gotten better either very slowly or not at all. Most drivers only naively practice driving.
(Many people probably have the same experience as me, of marveling at how middling an older relative is at driving compared to yourself. My <unnamed relation> has been driving for 40+ years! I’ve been driving for 5! How is it that we’re equally skilled? A lack of purposeful practice.)At work. If you play Go, you’re probably a programmer. Think back to the times in your career when you felt like you got a lot better at programming in a short time. For me, these times of growth are all related to new ideas and new patterns of programming:
when I moved from R to Python and started doing real, no-shit Object Oriented Programming.
when I learned about functional programming, and how many problems overly complex object state can cause, and started deliberately trying to make the stateful logic of my classes as simple as I could.
when I realized how good Jupyter notebooks were, then
when I realized how bad Juptyer notebooks are, then
when I realized how good Jupyter notebooks are
Perhaps you are a real programmer, unlike me, who is a fake programmer; or maybe you don’t program. This is fine. The point is not that there are commonalities between Go and programming (besides demographically), the point is that you should have similar moments in your Go life. You should have specific memories of when you started trying out specific ideas in the game, and they opened your mind to possibilities that you had not before considered. You must be trying to create these mental “opening-up” experiences in Go. Here are some ideas on how to create these moments:
Every time you have black for the next month, play the Sanrensei, and try to make a moyo. Here is a good primer.
Become a 3-3 goblin. A 3-3 goblin is someone who invades at the 3-3 point in every opening. I am a 3-3aholic, and one of my favorite things to do as white is:
The game starts.
My opponent plays a star point as their first move.
I invade at the 3-3.
But it isn’t just the opening where you can change things up like this. Studying the first few moves of the game, and agonizing over which is better, or which joseki is appropriate when white has a 3-4 in the upper right instead of a star point, is a waste of time. It’s strangely fun, and personally I’ve spent way too much time in my life thinking about the first few moves, because they’re so fun to ponder, but they really, seriously don’t matter.
To prove this to yourself, open up a program like KataGo and play a few moves and take a look at the values for the next possible move. Here’s an example - take a moment, before scrolling down, to think about what black’s next move should be in this opening:
You might think things like: “is it better to approach the lower-left corner, or the lower-right?” - “I heard AI doesn’t like the Sanrensei, so I probably shouldn’t play that” - “when I approach one of white’s corners, is the high or low approach better?” - things like this. But I have great news. It doesn’t matter! Testing the various options in KataGo:
Playing the Sanrensei loses half a point.
The low approach, high one-space approach, and high two-space approach to the lower-right, and all three of these options to approach the lower-left instead, are all within 0.6 points of each other.
Black making a star-point-based corner enclosure with either knight’s move, in either of black’s corners, loses 0.3 to 0.4 points, i.e. about 0 points.
Even this stupid move only loses 0.6 points:
Or consider this abhorrent corner approach, despised by all right-thinking people:
THIS must lose multiple points, right? If you played this and had a stronger player review your game, they would spend 5 minutes berating you and making sure you never played it again.
It loses… 0.9 points.
So minutae about the interactions of the corner stones with the overall board position are a wash, unless you are professional strength. You can throw away all the knowledge about the differences between these options and be fine! The real answer in this board position, for anyone below amateur 7 dan, is play anything in any corner. The opening, for amateur players, is just that easy.
(Advice to strong players: focusing on these opening minutae - especially direction of play - when reviewing the games of kyu players is not a good use of time, for the same reasons).
This is why, when you go to expand your mind and frontiers, your experimentation must not be limited to fuseki patterns like “mini chinese” or “sanrensei”. These are good things to try, but also try things like:
Playing amashi in every game for a month. Grab territory, let your opponent sketch out a sphere of influence, then gleefully plop yourself into the middle of it and try to make your way out or live. This will develop your ability to live under heavy pressure; hone your instincts on how much space you need to live, and how to exploit weaknesses in an opponent’s walls to eke out a life.
Playing Cosmic Go for a month, or at least make humungo moyos. There are two high-level approaches here:
desperately defend your moyo from any encroachment, and bet the win on your ability to keep your moyo intact.
allow incursion into your moyo, and take it as an opportunity to attack the invading stones, gaining benefit elsewhere via the attack.
Try both.
Fight like a dog. Viciously do every unreasonable thing that occurs to you, and if you fall behind, become even more vicious, and attack whatever group looks weakest until it dies. You either win by 100 points or lose by 100 points. You don’t need to read any books or watch any videos for this. Just make an account at WBaduk and play literally any Korean on that server. Their Go is a bloodsport.
Intentionally grinding these styles is of great benefit. Once you return, after your months of experimentation, to playing your usual style, your time with Amashi will have prepared you for those occasions where your opponent has built a large moyo and you need to live in it or lose. Your time making moyos will let you take advantage of an overly-territorial opponent, and crush them with thickness and influence. Your time in the bloody trenches of WBaduk will have equipped you with knives strapped in concealed locations all over your body, to unsheath when you get in a fight.
Review your games
The above, about playing different styles, is fun, but it’s probably not the most effective thing. The most effective thing is reviewing the games you play to see what went wrong. You can use AI for this.
Some argue that AI has idiosyncratic preferences about moves that only work because of some complicated 30-step plan, and fine, yes, this is true. But when people say this, they are talking about things like KataGo rating the Sanrensei above as losing half a point. This is not the normal case when reviewing with AI. When you do AI review, you will find that you lost 30 points on one move, and 8 or 15 points on a couple others. Here is a graph of the AI-estimated point difference in a game I recently played as Black:
The three things I’d review in this game are:
Why did I pull ahead around move 40?
What went wrong at move 135, to bring the game back to even?
What godforsaken fuckup did I make at move 155?
There are little wobbles in the graph everywhere, but those three moments are the ones that the game hinged on. They’re not about AI preferences that depend on superior AI reading ability - they are major mistakes by the players that can be understood after examination.
You can download LeelaZero or KataGo and run them on your own computer, or if you’re lazy like me, you can pay OGS $11 a month (and they have a tier even cheaper than this) to use their AI analyzer. There are other sites too, but I strongly recommend you use KataGo, because KataGo gives you the estimated point difference, and other engines, like LeelaZero, only give you the estimated win probabilities, which are hard to understand. (I understand probability well - Bayesian statistics is one of my hobbies. But I do not understand win probabilities).
Getting a human to review your games is good too, and the best is a teacher who learns your strengths and weaknesses and can track them over time and emphasize specific areas you should focus on improving. The Yunguseng online Go school is excellent at this.
The correct schedule is play → review. Play → review. You might play a game online and lose and not want to review because it feels bad. Review it anyway. You might win a game and think reviewing a won game is not useful. It’s very useful! Review that game too. It is better to play one game a night and review one game a night than it is to play 5 games a night and review none.
Play → review is purposeful practice. Play → play → play → go to bed is naive practice. Naive practice is the least effective form of practice. Play → review.
(A warning: studying without playing will likely get you as far as playing without studying. I’ve been watching Redmond videos multiple times a week for years, because I like them, but I barely play, and almost never do directed study of joseki or mistakes I’ve made in my games, and as a consequence, I haven’t gotten stronger in a long time. Beware the trap of Go infotainment).
Study stronger players’ games
There are four pillars of Go improvement:
Playing.
Reviewing the games you play.
Doing Go problems
Studying the games of people stronger than you
Everyone is already convinced of 1-3. But many are not doing #4.
All kinds of wonderful things are revealed by reviewing the games of players better than you. The group you think is weak, they play away from. Why? It turns out that group is strong. The group you think is strong, they reinforce. Why? The group is weak. You think they should invade at the bottom; instead, they play an extension at the top? Why? Why? Why? He played THERE?? Much is revealed.
If you could perfectly emulate a 7 dan, you would be 7 dan. So even trying to memorize the games of professional players, until you can play each move from memory, is worthwhile. (Maybe skip the last half of the endgame, to save time). Memorizing a list of 200 coordinates (Q-4… P-4… C-16… 197 more of these) is impossible for most people, so if you can remember 200 moves of one game, it means you’ve succeeded in internalizing some of the reasoning of the players. (Analogy: memorizing 50 random words in sequence is extremely difficult, but there are lots of songs you & I can sing along to. This is because we’ve internalized things about musical structure and the structure of English). I used to sit in front of a board with a book and play every move by hand, then do it again, then try to play it from memory, with the book’s help. I had a great time, learned a lot, and I wasn’t stronger than 5 kyu at the time.
Commented games are excellent. The Redmond Reviews series on YouTube is great, though I find it much harder to focus on the details of a video, so I still recommend books too. The Invincible series is heavily commented, and focuses on the 1800’s, when Go was simpler and the moves were easier to understand. And you can download a few commented games from the pre-AI era here.
You can get a million uncommented games at Go4Go. Even reviewing uncommented games is useful, to see what areas the players prioritize, when they consider something alive, how they invade (especially how they invade! The way strong players use weaknesses in their opponent’s shape to make their own groups live is extremely instructive). And you can memorize these too.
If you don’t want to spend any money, you will run out of commented games fairly quickly, so Go4Go or other sources of uncommented games will be your main resource for seeing how strong players play. The iOS app Go Eye gives you a stream of up-to-date pro games, and costs $20 for a lifetime subscription (it sources from Go4Go, so if you don’t have an iPhone, you can still use the Go4Go website).
Count
You need to be dan level to play dan-level moves, but you don’t need to be dan-level to count. Very few kyu players count the game. You can study counting for a week, start counting your games while they’re happening, and have critical information about the state of the game that almost none of your opponents have.
This is such an advantage. Your opponent invades your moyo. What should you do? The answer depends on whether you are behind, leading, or about even. If you’ve counted that you’re 30 points ahead, you can chill and let them live while surrounding territory, and coast to the win. If you’re behind, you’ll realize you need to make a desperate effort to kill the group. Weaker players make the mistake of betting the game on a dangerous fight, when they were 15 points ahead before the fight. Or the opposite mistake, of sedately defending and being safe when in fact they are 10 points behind and nearly doomed. To make the right strategic decisions, you need to know the state of the game.
Pros can count with an error of less than 1 point in either direction. I can count with an error of 3-5 points. A kyu who learns to count will have less precision than this, because they know less about what is likely to happen in each region of the board. But also, a 10 kyu game that finishes with one player winning by 6 points is considered a close game. Often, games at this level end with margins of 15 points or more. With these margins, a count with an error of ±10 points in the middle game is still an informational advantage. I can’t count nearly as well as professionals - ±4 points for me, ±0.5 points for them - but I consider a 3-point final difference to be close, and they don’t. So counting still helps me a lot.
I actually count in a pretty bad and inefficient way: I count pairs, and remember how many pairs each player has. One one hand, counting pairs is way faster than counting single points. On the other, there are much faster ways to count: identifying shapes of 10 points, 5 points, 20 points, and so on. So I’m a bit worried I’ll ruin you forever with this advice, but if counting is so difficult that you can’t even bear to begin, I’ll just say: count in pairs.
Study joseki
Now and then, stronger players will say that weaker players shouldn’t study joseki. Don’t listen. All of them studied joseki.
Pros study joseki. When AlphaGo Master was unleashed online in 2016, many of the times it pulled ahead were cases where it played a new joseki and the pro they were playing didn’t know it and didn’t play the right continuation. In the Edo period in Japan, schools would secretely invent and study new joseki to spring on their opponents in important games. In a recent US Go Congress, Stephanie Yin’s students had all prepared a joseki that they sprung on their unsuspecting opponents, and they won many games. I went to a tournament in November and misplayed a joseki in one of my games, fell behind because of it, and never recovered. In my next game, my opponent badly misplayed a joseki and I pulled way ahead early.
Joseki gives you a serious leg up. Knowing more joseki than your opponent, and how to take advantage when they misstep in the joseki, will win you lots of games. Another wonderful thing about joseki is, once you know one, it’s in the bank. This is unlike studying the middle game, where every situation is new, and it can sometimes feel like you’re making no progress in your studies at all. If you learn how to approach a high 3-4 enclosure, you now know how to deal with high 3-4 enclosures forever! This is concrete progress.
It’s easy to know when to study which joseki, too: if you play a game where a joseki you’re not familiar with comes up at move 25, and you do your best, and after the game during AI analysis, your score tanks at move 30... you should brush up on that one.
You can use AI to study joseki. Just start playing in a corner and see what it wants to do next. If you think a different move looks better, try that move and see how the AI punishes it. You can also study them on OGS or from Kogo’s Joseki Dictionary. Kogo’s is pre-AI - it says the 3-3 invasion of the star point isn’t joseki - but most joseki is from pre-AI times anyway.
I learned joseki from Ishida’s series, written in the late 70’s. The commentary, and exploration of alternatives - why the joseki is the joseki - is the best around. Volumes 1 and 2 are still useful, but you can skip volume 3, since it focuses on star points, and little was known about star points even then.
In summary
Purposeful practice is the most important thing. What did you get wrong in your last game, and why?
KataGo is the best AI for studying because it gives you point differences instead of win probabilities.
When studying with AI, ignore the hundreds of places where your score went up or down by a little. Focus on the places where the score changed a lot. You can find these with the graph that every Go AI program shows you.
Study the games of stronger players. Now and then, invest the effort to memorize the opening and middle game of a pro game.
You can increase your win rate over players near your strength by knowing joseki they don’t and by knowing the current score.