Are AI reviews correct?

Do these AI reviews make any sense, or should they be taken with a grain of salt? I look at them and instead of learning something from it, most of the time I don’t understand it at all. I get -40 points for what seem to be good moves. On the other hand, I did lose, so I guess there must be some truth in them…

They usually do make sense. Of the “???”'s you put in the comments:

  • first one is a move on the first line while there is plenty of larger stuff in the middle to play, so yes, I agree with the AI.
  • second one is an invasion point, so it’s not a bad move. You can invade the black area still, later it will be a lot more difficult if black closes the area more solidly. The AI doesn’t think your move wasn’t so bad there, so it doesn’t matter as much.
  • third / fourth move is pushing from behind: you practically force black to grow his area, so yes, that’s a bad move.
  • fifth / sixth move is gote, your area doesn’t need it to be secure, and you just save a single stone. In sente black can now protect the middle from your bottom left (which he didn’t, but oh well)

Probably try to ask the following questions when the AI marks something as a huge mistake:

  1. Is there another group of mine that is about to die? Did I have to protect it?
  2. Is it gote? Did my opponent have to respond to my move?
  3. Is it small? Could I have secured more points elsewhere?
  4. Does it help my opponent more than it does me? Did it force him to secure something weak while I got nothing back in return?

Currently, Leela (which is not yet reflected by our implementation) is unable to cope with handicap or custom komi games, but will still review them. In these cases she may be quite wrong. In even games, Leela is way above us, mere humans, so she is probably not “wrong”.

That however, does not mean you should not take her review with a grain of salt.

First off, the numbers represent a change to Leela’s percieved winning chance, not a “move quality”. That means that in an even game what we would percieve as a very solid move, can be -40% just because another move is worth 2 extra points and now your opponent is winning by 0,5.

Sometimes what we might think was a “good move” can have a complicated tesuji refutation nor you, nor your opponent will see. What I mean to say is that even if she is again “correct” we may not understand why, and hence the information is not usefull.

There is certainly a lot of truth to them reviews, and it is a great learning tool. I would still ask for human reviews too though. We may not be as strong, but we are better at explaining :slight_smile:


Okay, I didn’t mean to sound disrespectful, and yes, I can see often that my moves are not very strong, but what I mainly have trouble understanding are the suggestions by the AI. For example:

  • First one: yes M1 is weak and unnecessary. But what would be the point of P10? It can’t go anywhere and seems completely meaningless. In fact, in this position I already felt I was losing because I didn’t see other big moves to make and I had only small potential territories.
  • Second one: fair enough, but why does it suggest Q6? Again: what could I possible achieve there?
  • Third/Fourth (move 102): I don’t know what you mean. O12, I don’t see as “pushing from behind”. I see it as defending against being completely locked in. I often get completely eaten if I don’t defend regions like this. So why is this -45? Moreover, the AI suggests L10 and again I wonder: why? Because it maybe can connect to bottom left or top? Seems farfetched to me.
  • fifth/sixt: okay, maybe I didn’t need to defend here. It was one of those “better safe than sorry” moves which cost me a tempo, fair enough. But again: L9? Why not K13 or G8 for example?
  • another example: move 104, K15. Maybe it’s not as strong as I thought it was, I am sure there are better moves. But -50 points? When I made that move I was sure it was the “typical sort of move a good GO player would make”: relatively big, expanding, going towards the center, etc…
1 Like

I second this excellent statement!

@Igiveup, good on you for asking these fair questions. I see that you have thought about Leela’s suggestions. She does not offer any explanation though. This shows you the limits of what the AI can do for you.

Why does Leela view those white moves so negatively? Because in three of the four cases, you defend a group that is already alive. It’s nice that you want to be “better safe than sorry”, but objectively, such a move is barely better than a pass. You could be getting big moves and points instead.

Why is Leela obsessed with playing in the middle? Apparently that is the best point to reduce the black potential.

In general, when you reduce a moyo that is open towards the center, you look for moves at or just outside the “sector line” - the imaginary border between the furthest extents of the surrounding moyo stones that surround the potential. This allows the reduction stone a way out. If you go in too deep, the reduction will be enclosed behind enemy lines.

The worst you can do in Go is induce your opponent to play the moves that you should have gotten first, like finishing a moyo (where you could have reduced). That is basically what you did.


I disagree.

P10 cuts P11 and can lead to a capture of this black stone and building a white territory around it.
Also it treathens to cap and even cut the black group on the right, if black pushes it in that direction.

If you look at the variation you can see a fight where both black and white have weak groups surrounded by the opponent: the one who escapes or captures gets a big reward!

So P10 isn’t meaningless: it’s fighting. :slight_smile:

Look at the variation: it’s a peep. Black needs to connect and white gains a free stone that can be useful to fight in the middle.

In general: it happens to me too to be surprised and baffled by AI choices but often this is because I’m just too weak to understand them. :smiley:
In some situations, though, suggested moves are just too aggressive for me and I wouldn’t do them even if I knew them before.


I think my point is also, that it’s a more useful tool to learn what not to play than to learn what the best move is. LZ is pretty ruthless, a little chaotic and a near-perfect reader in life / death situations, and since humans usually have an easier time playing solid, basic strong moves than creating a complicated mess of fights everywhere, I think taking precise suggestions by LZ with a grain of salt is not a bad choice.

It does show the general idea of what LZ thinks are good moves though. She plays in entirely different areas than you did, so she does point out where the big points are. You don’t have to exactly follow that, but it should be a good push in the right direction to try to figure out why LZ thinks a certain area is more important than another.


Also a good thing to remember: not every one of your stones has to survive the game. If a stone can be useful later to force an opponent to capture it (to avoid big losses), then it was a good move. Especially in invasions, you shouldn’t expect everything to live: the goal is to create something that’s alive in an area that was your opponent’s potential, or if that’s too difficult, to get a group running away and connecting to another of your (living) groups.


OGS isn’t using Leela? Please, can you elaborate?

You may also want to take a look at ZBaduk, which does use LeeLa Zero.
But runs on powerful GPU servers, just accessible from your webbrowser.
And best part: it’s free (for now).

Next release of ZBaduk will be able to play handicap games, different komi, etc …
But even without all those features, ZBaduk may be worth a try.

One thing to check - whether this question is about “supporter” or “casual” reviews.

You get what you pay for.

Personally, I think that the “free” reviews should be taken with a grain of salt. The settings are for a very quick and approximate analysis.

On the other hand, the “full” reviews run a decent analysis, and ought to be “correct” in as much as any Leela review can be correct.

1 Like

I guess in this case my opponent was a supporter because the review looked completely “full” to me.
The “casual” reviews (only 3 moves which are supposed to be the worst or something) are completely incomprehensible to me, but that I can understand.

Second one suggests q6 because it is a peep that threatens to cut off and kill part of the wall if black doesn’t respond. If black does respond, now you have a stone that you can potentially run away with, which would severely damage blacks influence in an area where black would prefer to make points. Living inside an area where black wanted to make points is obviously good for white.

1 Like

OGS is using Leela Zero. What I meant was that our implementation does not reflect the fact that she can’t play handicap or custom komi reliably, therefore reviews of these games may not be trustworthy.

Okay, I guess I would have to learn how to “potentially run away” with one white stone in an environment that is completely dominated by “B” (or at least that how it seems to me).

1 Like

Well, the important point to note is that you can get that one stone in there for free. Maybe there are other cuts/peeps that can help you to get one or two more stones in there. All these do not even need to survive to be useful. It may be enough if they help another invading/reducing stone later. For example the opponent may have to worry about you connecting your stones, which restricts his/her options.

Don’t get me wrong, I’m not nearly strong enough to really see how to efficiently use it, but I myself am in the slow process of learning that dead looking stones within enemy territory that I get in there for free can be useful.


I think the most important point is that at that point white has some solid territory but black had huge potential (as you pointed out yourself). Therefore you have to try and reduce that potential SOMEHOW. If it fails and you lose the stones, you will lose the game, sure, that happens. But if you do not try at all you are pretty much sure to lose by default and will learn nothing. I would therefore go for the first option.

Whether the peep is better choice than a cut, who knows… Of course you cannot see that (I probably would not be sure either), but by trying something you get experience on these invasion points. By getting experience you will see more options and will succeed more often and will climb in rank :slight_smile: By saying “man, black has more potential than me” and than doing nothing to fix that (or rather the opposite), you will learn nothing.

Potential is not territory. That’s why we call it potential.

I somehow overexcited myself with this rousing pep-talk, hope it did not came out as rude :smiley: not my intention at all.


You do it by abusing other forcing moves or cutting points in the same area and playing light shape, which means that you must be willing to sacrifice a small part of the group in exchange for saving most of the group. Just click on the move in the review and you will see an example of the potential followup.