Proof of concept - small 2 eyes in the corners

Here another one!

This was actually a victory!!

OMG, that is really fun!

Look how for one single stone I didn’t managed to get the bottom right corner.

But since the strategy can really achieve a true victory, maybe I’ll improve it and play rated and go for the win.

I’m so happy! This is my SECOND VICTORY EVER… :joy:


Dr. Bèco


I loved the passage in the book The Master of Go by Yasunari Kawabata: Ch 10 (page 46, 2006 paperback)

Onoda of the sixth rank … His manner of play was equally remarkable. While awaiting a play he would sit quietly with his eyes closed. He explained that he was ridding himself of the desire to win.

I hadn’t really thought about playing Go in those terms before but in trying to win you often sow the seeds of your own defeat. So I would urge you to think less about winning and losing and consider more the process of the game and the meaning of each move.


p.42 (second half of chapter 10) in my paperback copy.

My “opponent” is not my foe. They are my partner in this leg of my journey to discover Go.


Does this also count when we win against bots? :innocent:

1 Like

Very much so. Bots have shown us HUGE swathes of the Go landscape that we humans might not have otherwise seen.

…and we’ve drifted off topic.


You can pull this off against amybot. And it may be a good exercise. But you should know that a skilled opponent won’t let you do this. In fact, there’s a maxim for it:

In 9x9, the general rule is to aim for one or maximum two groups.


When I was first learning the basics of go, motivated by the frustration of losing many stones, I asked of my friend who was teaching me, why should I not just make a small living group with two eyes and then cautiously extend outward one stone at a time, keeping everything solidly connected in order to avoid losing any stones to capture.

He explained to me that while that strategy would prevent losing any stones, it would also be very slow at making territory and hence I would quickly fall behind compared to an opponent that played more efficiently and ambitiously, even if that comes at the cost of keeping their stones less safe.

Consider this game, where white and black have each played six moves:

Black has an unconditionally living group and secured a small amount of territory in one corner. Even though white’s stones are spread out, they should still have no trouble later connecting up and making enough eyes to live. More importantly, those stones are very strong in exerting influence and claiming territory over a much larger area of the board. In fact, it would be very difficult (requiring a major blunder by white) for black establish any living stones or claim any territory in the entire upper-left half of the board. Hence, white is in a winning position with a very large advantage over black.

The central challenge of go strategy is figuring out the proper balance between ambition and safety as one tries to efficiently build up territory. If we build too slow and safe, we might not even seek to claim enough territory. If we build too fast and reckless, we might not be able to hold onto all that we claim as the stones we lay down get threatened and killed. It’s not easy to get it right, but this balance between speed and safety should always be on our mind.

On another point, I would recommend avoiding playing against bots too much. The weak bots generally have predictable, glaring flaws that one would soon recognize and easily exploit. Hence, playing too much with them might have the negative effect of reinforcing bad habits. Human opponents will each have their own strengths and weaknesses, but you will encounter a broader range of skill and styles that will teach you more. Even if you play the same human many times, they may also learn from their own mistakes, so while two people can benefit from mutual learning, the strategy and style of play will also not become stale.


:hugs: Well done!! While there is a lot of advice being shared here about how to play Go properly and how to view the pursuit of Go in a healthy way, I would like to congratulate you on your second victory! At this point in your development, sticking with it can be difficult at times.

One of the fun things about Go is that every time you think you’re finally grasping the game and beginning to feel confident with your skills, your knowledge of the game tends to advance. You see new layers of complexity, new possibilities, things you thought you knew evolve, and you realize you are strong at what you’ve learned, but weak as can be with what you’ve just uncovered. This journey never ends.

Pursuing Go can often feel like a rollercoaster of highs and lows. So right now, when you are feeling a high, I encourage you to embrace it. Celebrate it! Feel it in your bones and let it cover the walls of your heart with vibrant color. Hold onto this feeling and when times get tough in the future, remember this and strive to feel it again :wink:


Dear Yebellz,

I’ve saved this start position as a .SGF file:

Then I tried to play against LeelaZero.

Look how nice it turned out:

It seems to be a good exercise nonetheless.


Dr. Bèco

PS. Oh boy, I think I missed a win!! Wow!

Bots don’t care about their winning margin. If you play against them, they’ll give you some small corners for free, if you’re giving them the win in exchange.


It looks like the .SGF file had some problems. The first 6 white stones are not in the board. I’ll need to figure out what happened and upload again.

Here the correct game.

But after asking for a AI review, it shows black with 100% chances of winning. Very strange.

The AI review has problems with the initial position. It tries to play moves on top of stones.


I would actually say that it was a terribly misleading exercise, since the way Leela played out this position is not an accurate reflection of the absolutely dominating nature of this position for white.

Leela plays in a very inhuman-like manner throughout that game and even makes many irrational moves that needlessly lose points. For example:

  1. At move 26, Leela plays J4 which makes it possible for black to capture 3 stones with a play at H6.
  2. Leela shows no interest in trying to save those stones, but does eventually protect after black fails to capture them for several moves.
  3. Black’s entire invasion into the top-left could have easily been stopped. There are many opportunities where white could have killed black’s invasion and protect their own stones in the process, but Leela simply let’s black live and some white stones die, with a sense of profound indifference.
  4. Overall, white seems incredibly indifferent to maximizing their own score, even before move 26.

This might seem to contradict the fact that Leela is not a terrible AI, but I think what’s going on is that Leela only cares about maximizing it chances of winning, and does not care about allowing black some additional points in the process (as long as victory is safe). Essentially, starting from the initial position, Leela is most likely calculating that white has a winning percentage of nearly 100% and that practically every move during every point of the game has no noticeable decrease in that winning probability. So Leela essentially just goes through the motions of finishing the game, placing stones haphazardly, with complete indifference to black chipping away a bit at the score.

EDIT: as @flovo pointed out below, @drbeco seems to have been playing with Leela rather than Leela Zero.

EDIT 2: not really sure how strong the original Leela is, but I assume that it is not terrible.

Another strong bot, KataGo, can be configured to care about maximizing its score, which would lead to more human-like play.

Basically, as @flovo mentioned, there is a major bug with how the OGS’s AI analysis is handling this particular SGF file.


Thanks Yebellz,

It was very informative! I guess there are some adjustments I could do to minimize the effect, but all in all, it should be taken with a grain of salt.

Good to see the overall winning strategy, but to understand local fights humans are better for beginners.

My best,

By the way this is Leela, not LZ,


Good catch. I assumed that it was Leela Zero since @drbeco does say “LeelaZero” in their post above:

However, the specific bot name is listed as only “Leela 0.11.0” in the SGF, which does seem to suggest that @drbeco was mistakenly playing against the older Leela bot instead of Leela Zero, which is considerably massively stronger.


That is an understatement. IMHO original Leela is not even dan level on 9x9




I installed LeelaGUI, sjeng version 0.11.0, and until now I was thinking I was playing with the engine “Leela Zero” that is installed together with the GUI.

So I decided to test it, and also I went to github original current LeelaZero and installed it as well so I could compare the results.

I created another post here, so we keep the forum organized for off-topics.

1 Like

Now you know it’s called Leela, not LeelaZero. :wink:

How to know:

AlphaGo is 2016, everything before is standard Monte Carlo tree search (or worse).

LeelaZero plays only 19x19. It’s network is trained for a fixed grid size of 19x19 and 7.5 komi. If definitely cannot play

and last but not least

If you follow the link to senseis, you can read

Leela should not be confused with the more recent, stronger bot Leela Zero; the names are similar only due to having the same lead author, but they play quite differently.