Ai reviews for kyus (things the AI taught me)

I was thinking having a thread which tried to help people, in particular kyu players (TPK, DDK, SDK) to learn something from the AI, and ways to use AI, could be useful.

It might be often said that the AI reviews aren’t suitable for players unless they’re above a certain strength, or that the AI reads too deeply and so we can’t hope to understand its suggestions.

However I think we can successfully challenge this, and learn one thing at time, even just one or two things per game.

Initially I’d like to share some of the recent things I’ve learned from reviewing my own games, but also I’d like to be able to show examples of how one can try to understand why the AI calls something mistake for all levels, so that we can all find something small to learn from reviewing our games.

Of course feel free to share whatever you've found interesting yourself at any level

It might be a long first post so I’ll hide some things in “details” and peruse at your own pace. And because it’s long I’d like to focus on few things that are hopefully SDK type level, but would be happy to try to review some DDK and TPK games to see if there’s some small things one could learn to use the AI for.

Enough of the intros :slight_smile:

Ai to validate ideas and decisions

Often I think of moves in a game that I feel are probably the “right” move. Sometimes I have my own reason for it like “it might be slow but it settles the group” or “this feels big for both” etc, and other times it’s some received wisdom which one kind of has to believe and adjust to over time.

Here’s many example of these. It’s not always the case that whatever my own reasoning or own understanding of some received wisdom is entirely accurate or perfect, but still arriving at the (one of) “right” moves and then reinforcing it and hopefully improving the understanding is a good goal to aim for.

Some SDK ideas

It's just endgame pt1

Often go proverbs need to be taken with a pinch of salt, and can’t be understood and apply in every situation. One proverb might say corner, side, centre, and then white can certainly try to follow up the joseki moves by emphasising the corner with R18 below.

Other times people will say “keeping/taking the corner is big” but of course the sizes of these things change throughout the game, and the received wisdom in this position is that a move like R18 is actually small, it’s endgame in a sense.

So one is supposed to answer it not by blocking in the corner but by a bigger move on the side, typically something like R11. If one believes this received wisdom and plays it in a game like I did recently, then one can be quite happy to see some validation from AI :slight_smile:

Black does win this game, but of course it wasn’t down to just this move, there’s many many things that happen and go wrong after this :stuck_out_tongue:

It's just endgame pt2

Here’s another position below where it can be quite hard to accept that it’s okay to tenuki such a move by the opponent because it’s just endgame. When black has a strong stone on the second line like at B18, there’s always the worry that Black at some point in the game will monkey jump and greatly reduce the White area. However again while the monkey jump can be estimated to be worth 10 or so points, it’s something again that’s said to just be endgame, and there’s probably bigger moves elsewhere.

So in the game Black played the monkey jump quite early, far from the endgame. I naturally had to will myself to accept the wisdom that this isn’t big enough to answer, especially if Black can get this reduction in sente. White’s move previous to this was the turn at L4, and so my belief was that there was a big follow up here, not to kill Black’s bottom group, but certainly to try make it live small, or confine it to the side if possible (At least if black escapes White might be able to build the middle). So for that reason I played K3 and some further followups (which included mistakes by both players – we’re still kyus :stuck_out_tongue:) and let Black destroy the White area on the top.

Again White does win this game, but it’s not because of this tenuki necessarily, but certainly the tenuki did not stop White winning :stuck_out_tongue:

Whole board thinking

Now this next one it could’ve been the case that the AI absolutely hates it, but I did come up with a plan which the AI agrees with, and this was my though process. Naturally as is the style of the time, White has to decide how to deal with the 3-3 invasion.

Initially I think the idea is to block on the side you have more potential to develop, so naturally one might decide to block with D3. However these days the 3-3 invasions can be quite flexible in their directions. Since this came up for me in a correspondence game (feel free to have look through) I had plenty of time to come up with a plan.

So first block on the wider side or side you can develop. Now whether to extend straight up with D4 or try the knights move is quite a tough decision. My thinking was that if I play D5 I give black the option to knights move to B6, and to be honest I would be much happier with one of the simple versions that happens in the game after the knights move (there’s many complicated versions but at least White can get a wall in a good direction, even keeping it simple).

Now here’s some whole board ideas, and where the flexibility of direction comes in. Again being a kyung my ideas and reasoning can also be slightly be wrong, but one can still hope for certain variations.

Thinking like an AI, it might be nice to swap directions, get a thick shape on the left side and pressure the two Black stones in the upper left corner (and hopefully make some points) with the following joseki for example, but ideally if one can capture the stone in a ladder White is much stronger. So instead lets try another AI favourite joseki first to try get a ladder breaker in a good spot :stuck_out_tongue:

And this is how the game went, and nicely White keeps Sente :slight_smile: So now we followed the plan for the bottom left. Of course at any point our opponent could muck up the plan if they figure it out, but there’s always a chance we can get what we want, none of these variations look bad for Black. So we play out according to the plan, and once again manage to keep sente. There’s still some slight worry that Black will find a cool ladder breaker in the upper right, but that’s a future problem :stuck_out_tongue:

Well the idea and the plan was almost good. It turns out that now pressuring Black group at the bottom should take priority, a move like K3 for example. It also seems like a high move like D13 would be better if playing in the upper left.

If you’re a bot of course, you make some strange super human exchange and then tenuki but hey it least considers my plan after such an exchange (although again maybe the high move is still better for reasons maybe I’ll learn someday).

If you made it this far, you might think this is a waste of time to plan such things ahead, but I found it quite fun, and sometimes there’s things to learn when and why the plans go wrong.

Your opponents good move is your good move and vice versa

There was a game I played with a fairly standard 4-4 approach and pincer joseki. Then I played what I thought was a good move, but I overlooked the idea that the received wisdom to remove aji in the position on the right, following the joseki is the nice two space extension (below it’s J3). It’s a move that for example was said to be preferred by Lee Changho, not in this position, but I did even manage to guess a pro game was his even though there were no Black stones on the board in the following

forum game

So I got a bit greedy and tried to get more and my opponent immediately played on the good point of J3 and the resulting position and fighting got a bit tough for me.

Sometimes playing simpler and solid, in a handicap game, is enough

This was a four stone handicap game, and while I opted for a ko variation, there was (a) a better placement move inside, and (b) just a better simpler solid move outside.

Just connect avoid the trouble of the cut, and let White worry about living :slight_smile:

Why are some mistakes the values they are?

There’s often quite a lot of the time where one can find it difficult to understand why a certain mistake is a loss of points, with the value that’s given. It’s often quite difficult to know in most situations, like in the opening, why a certain move loses 2 or three points for example. It can be hard to see those 2-3 points on the board. In others when a group dies it can be much easier to try to attribute the score loss to dead stones. Still there’s others where one can, with a bit of exploration try to get an idea of why some moves

Some SDK ideas

capturing with a snapback instead

In this game, some trade happens where White gives up 3 stones for a strong ponnuki in the centre. It’s actually in quite a good position especially given Black’s stone in the bottom right (the 4-4 stone influence is maybe a bit neutralised). In any case, I went for the obvious choice of just taking the three Black stones off the board. However this is like a 3-4 point loss. The better option was to capture the stones in a snapback instead.

Now this is quite tricky to understand I think without some further exploration. Firstly it turns out that White has a nice continuation locally to more or less block of the centre when Black just takes the three stones off the board, which my opponent played much later.

Now if you could forcibly and maybe peacefully resolve the ko aji in the top right, this might locally be an AI recommended move. So it’s quite annoying that White is left with such a good move to block the centre with.

Instead suppose Black captures with the snapback instead, white can get an Atari to force Black to capture the stones, but it’s much harder to block the centre – there’s too many Ataris left behind if trying to block directly. So it’s certainly believable that not playing the snapback is losing at least a few points.

Losing the capturing race in the 'right' way

This one is a bit complicated and it takes some explanation to come up with even a basic understanding of why the point losses are what they are. However even if it can be argued that it’s not worth understanding this niche situation, I think the ideas of how to try to understand it, and the principle that sometimes there’s a good and a bad way to lose a capturing race is important. (Also maybe I did waste my time with it already so :P)

So here’s a complicated position involving a capturing race between a Black and White group on the left side of the board in a handicap game.

White answers the hane and connect and loses a combined total of about 17 or 18 points in doing so. Now the initial thought would be that, maybe White would’ve won the capturing race if playing differently right? I mean there’s about 9 or so White stones to save and maybe kill the same amount of Black. But that’s more like a 30 point swing, not 17 or 18 points.

So in the game white loses the capturing race (on the left) but lets follow the AI suggestions locally and see what the difference is. On the right, White still loses the capturing race but what’s the difference?

Essentially, White capturing two stones pretty much guarantees White is super strong in the bottom left corner. So even though White loses the capturing race, White’s group is strong and so there’s much less aji in the bottom left corner for Black. If we were much stronger, or happened to be katago one might worry about variations such as

I might not be able to play these in an actual game, but I can still appreciate the idea that there is a right way to lose a capturing race with style, and a way to interpret the score loss. If Black can live in that corner White’s group is not alive and needs to run. On the other hand, even if Black can live in after white captures the two stones, at least White isn’t under threat with the whole group. (If you like proverbs, sometimes it’s said that a weak group can be thought of as being worth -20 points :slight_smile: That seems to match up here.)


I think it’s also worth mentioning couple of things that people that use the AI review might already know, but people who want to get started with it might find useful.

Firstly there’s already some nice documentation on using the AI review on OGS

It tells you things like how to read the win-rate/score bar per move, and about the review graph overall. It explains that the blue move is the “best” move in a sense that one can explore more if one wishes. The green moves are other considered moves and the opacity is related to how well explored the move was by the AI.

At the end of the day the AI isn’t perfect, maybe even more so on 9x9 where there can be quite complicated fights for the life and death of most of the board!

Still it can be quite useful. One useful tip that gets iterated over and over though is

In particular for handicap games the win rate graph might look like the following which might hide a lot of Black’s mistakes until it was enough to change the result of the game:

The score graph however can show a lot more details

In particular the win rate loss per move might only be 0 or 0.1 percent points until it affects the result of the game where it then could be 90+. On the other hand for score, it could be a series of 1,2,3…,10 points mistakes that all added up over time until White eventually wins by a couple of points.

There’s also a top 3 moves (maybe more in the future) which will try pinpoint some big mistakes by either player which could be useful to review, even without being a site supporter.

You can click on them to jump to that point in the game. There’s going to be an update soon also which will now try to find these moves by change in score rather than change in win rate. Hopefully this will be an improvement, and can continue to be improved upon in the future


I’d summarize the idea with a kind of proverb: if you die, then use the aji of the dead stones as much as possible.

Other example: move 83 of this game European Championship Top 16 doesn’t save F3 but allows Black to make a ponnuki in sente.