Has AI taught us anything new recently?

A couple years ago, AI revolutionized the way we approach 19x19 games. It taught us about early corner invasions and attach moves, fixed some old joseki and taught us new ones.

I see that Leela Zero keeps improving steadily, as do Asian AI like FineArt. Has recent progress brought up new discoveries?


I’m probably the last person to tell you anything about bots, but from what I understand, our old ways of thinking were not appropriate for using influence. I’m not sure what we’re supposed to use influence for now, but it seems the “build big” mentality isn’t really that great any more.

AI is very comfortable with building territory fast and then massively reducing any large potential.

1 Like

One thing I noticed that seems to have changed recently (?) is the 3-3 invasion joseki.

First when AI came out we learned that we are supposed to do that long extension thing on the 3rd line.

Recently I heard Dwyrin saying “we don’t do that anymore”. I’m not sure what “we do” now, I think it is variations of this.

1 Like

@Kaworu_Nagisa and @Eugene I hear what you’re saying, but I would call this old news… in AI timeline that is :slight_smile:

@Kaworu_Nagisa a fresh example of how building influence is hard to handle.

@Eugene the joseki you are referring to comes from AlphaGo Master, which learnt from human games. It was later superseeded by pure AI learning (codename “Zero”). And that is correct, black P17 followed by O17 is not considered joseki anymore. Black would rather play O17 directly, or play O18 after P17.

I was more looking for something that might have evolved in the past 12 months. According to the chart, LZ has improved from Elo below 13,000 to above 15,000 during that period, and I was wondering what we might have learnt from her improved play.


We learned that you have to be able to read 40 moves deep for 30 different variations and judge them all accurately in order to understand the new joseki.


lol and that review just scratches the surface… Indeed the 3-3 invasion has become one of the most complicated joseki.

1 Like

I’m not sure if this is talking about joseki anymore.

The whole idea of joseki is that they are a sequence you can rely on to be locally even without full board analysis etc.

40 moves deep is not joseki, it is just full game analysis.


No no no no no, please never play joseki without looking at the whole-board situation. Never do this, and never recommend anyone you teach do this.

Joseki absolutely require you to do full board analysis. If you do not, then you have the potential to just play mindlessly and make the game unbearable for you at move 20.


There’s a difference between “full board analysis” and “taking into account the full board position” though, right?

I mean, I can choose this joseki because I have a stone over on the other star point etc, or because I want influence go with what’s happening … etc, sure.

But doing 40 moves of variation to see exactly what happens … that’s not joseki based decision making, that’s full analysis based decision making.

1 Like

@Kaworu_Nagisa agree, and I don’t think @Eugene meant it that way.

While most joseki are much shorter, some local sequences can be pretty long (think avalanche, magic sword, taisha). There is no absolute limit to the number of moves. When I said the 3-3 invasion was one of the most complicated joseki, I meant it! Here is an example of a 40+ move joseki starting with a 3-3 invasion (rewind for the details).

Note that my question is not specifically about joseki.


I am by no means an expert here…

But stylistiacally the biggest changes to have happned in total are that the whole order of play is very different to how it was before AI even came about. I dont really think that AI has mastered the game. I think that there is still much to learn about the game. The biggest changed have been in joseki, and indeed the way in which the game is being played, priorities have changed.

The biggest thigns we are gonna see from here on out are likely going to be more of the same of what we have had already. More joseki, and even longer ones. the machines do not care that we cannot read forty plus moves for forty plus variations… thats how they play, its not how we play. Especially the average amature player.

We very much get fed the drips and drabs of “New” moves, and hope for the best when we play them, because we dont have the same abilities that an AI has.

Though i do think it would be intersting to see AI play on larger boards, see what they make of them… and even train an AI to play on larger boards, and then make them play on the standard 19X board. See what happens.

But yeah, anything new is going to be more of the same, a new sequence here and there, and perhaps more change of style, but its really hard to predict.


Is that recent though?

I’d say that this describes the change that happened pretty much as soon as the AI started winning and the pros looked at how.


In terms of Go though, anything in the last fifty to a hundred years is comparitivly recent given the age of the game.

1 Like

Heh heh - I think in the context of AI, though, recent is “last couple of months” and “old” is a year or more :wink:


Right, I clarified that my question was about the last 12 months. Anything before 2015 is pre-AI-story.


Perhaps a newer thing is that strong humans are slowly getting convinced that 5-space extensions are not automatically good.

It was somewhat apparent from AlphaGo’s teaching tool 2 years ago (I made a video about it in early 2018 https://www.youtube.com/watch?v=OapPVew_stM). But it wasn’t immediatley picked up by stronger humans. But nowadays strong AI are widely available and they confirm AlphaGo’s judgement. And I see the popularity of 5-space extensions declining even in amateur games nowadays.


I’ve heard that the first joseki played more or less doesn’t matter. It’s from that point on you need to take into account the rest of the board.

More or less the first joseki starts probably within 5-10 moves, so there isn’t much whole board analysis to do really

I think more or less is the key part of that sentence. For a 20k none of the joseki in any of the corners will matter, for a 5k the first one or two will barely have a huge impact, for a 4d choosing the wrong first joseki can be painful and for a prof the wrong joseki will probably end you up losing the game.

1 Like

If black plays a 3-4 on move 1, and then white approached the 3-4 on move 2, then it doesn’t really matter. As soon as there is another stone in another corner, it does matter.

I think it starts to really matter around 15k,because most tend to have a grasp of territory vs influence.

That’s a great point, thanks @gennan! I’ll keep this in mind when reviewing modern pro games. I have also noticed that difference among teachers, with the new wave promoting AI style while older pros still teach 5 space extensions.

Let me take this opportunity to say that I find your videos very instructive, and I watch them over and over. I have already mentioned them a few times in this forum :slight_smile: