In a separate thread, it was asserted that “What the AI says is the best winrate should define what is Joseki”.
This has been a topic of intense debate amongst the contributors to the Joseki Explorer.
I would welcome more well considered input.
One specific reason why the AI’s highest win rate is not automatically and obviously “joseki” is because the AI’s highest win rate is for that specific position, with the rest of the board specifically empty.
Joseki is defined as the optimal play when all other things are equal on the rest of the board.
“When all other things are equal” is a loose English phrase which means more along the lines of “if there aren’t significant considerations” rather than “precisely equal”.
This is a very loose definition, if you poke at it, which is one source of debate.
However, there is an important sense in which it transcends this looseness. Humans are looking for advice about what is the best thing to play given that we can’t take into account the rest of the board in a deep way. AI is telling us what is the best winrate move given that the rest of the board is empty, and it has analysed that in a deep way. These are two different things.
It may be that nonetheless “if the AI says its best then it is”, but it’s worth discussion to validate that this will result in useful content.
(Right now the gold standard for “what is Joseki” in Joseki Explorer is “does a pro tell us that it is?”, and “have they said so recently”. We aim to have Joseki sourced from a credible source. We do have AI sequences, but they don’t have the same credibility at present).
BLUF: I really think we should just have the sequences presented “as is”, and have it be as complete of a dictionary as possible, with commentary (from humans) as to why and when you would want to play this. There is no value to a beginner or higher level player in having a 30 sequence joseki, ending with nothing more than a “52% winrate for black”.
I’m against telling the player anything along the lines of “you must play/avoid this”.
Can also think of edge cases where someone (likely me) plays some really old josekis that someone who religiously, dogmatically, only uses the most recent AI highest winrate joseki, and then loses horribly because they only care about what the AI says is “the best”.
Also, joseki are highly dependent on the rest of the board, especially as a lot of new joseki involve ladders and weak running groups.
Edit: We shoudl definitely include AI joseki and whatever else pros agree is joseki, but telling players to prefer one over the other I think is no good.
This is useful input about how the Joseki Explorer should be presented.
If I may observe, it is largely not about the question at hand however
You have included a recommendation that we include “AI Joseki”. This thread questions whether this is a legitimate and useful concept.
If it is a legitimate and useful concept, this thread also seeks for definition of what it means. For example, traditionally joseki are sequences played often because they are reliable in a range of situations.
The AI is perfectly capable of telling us that a move has the highest win rate on an empty board, but never actually playing that move in practice.
Well we shouldn’t be defining what a joseki is. It’s a little inappropriate for amateurs to do so.
AI joseki is simply what the pros have agreed to call a joseki, that was initially played out by AI. We should not make a new sequence in one of our own games and then call it a joseki because the AI review said our move was ok.
Edit: I understand that it would be useful to have a well defined definition for the joseki dictionary, but I don’t think we should be including anything that the pros haven’t already called joseki. What would be nice maybe, is to find some reliable resources that can be used to update the OGS joseki dictionary.
“we amateurs should not be deciding if a given sequence is joseki or not”, or whether you mean
“we amateurs should not be deciding that the definition of ‘joseki’ means”.
Can I address #2 first, although I think you probably mean #1.
Since I’m leading then curation of a Joseki Explorer, I kind of need to find out and lock down what the definition of what “joseki” means .
We need to be clear on what “joseki” means so that we can be clear on what we should include, and how we should label it.
Your suggestion for what “AI Joseki” means is different to the proposition of this thread.
It was proposed that “if the AI says the winrate is highest, then that is Joseki”. This is a topic that needs careful consideration, and is what this thread is here for.
About #1: “we amateurs should not be deciding if a sequence is joseki or not”
Absolutely true. The Joseki Explorer contributors are being very diligent about this, and insisting that anything marked “Joseki” has a credible pro level source… except they too agonise over whether “the AI says so” is also a credible source.
I also meant #2, but we shouldn’t find a definition that is outside of what the pros have said or done or commented on.
Now… I’m writing a book with Tokumoto on honte and what it might mean in a post-AI world and how it can still be beneficial, but we’re not going outside of what has already been said. It started mostly because there are actually no English books on it, and the topic has been kinda dead since Go Seigen, but it’s still worth talking about. It’s been about a year so far and it’s been mostly research and “reverse engineering” the one (and only one) Japanese book on the subject.
Joseki on the other hand, is a well discussed and known topic, and the topic on Sensei’s Library is, as far as I know, uncontested by anyone:
A standard sequences of moves played out in a corner that result in a locally even exchange. Note that “locally even” neither implies optimal nor even at the whole-board level. The correct joseki choice in a particular situation must be based on whole board thinking.
To me, AI is a sieve, not a sculptor. If a pre-AI joseki ends in 90% win rate for one player, that tells you that the joseki status was in error. Indeed, this has already happened (see Russian Go Federation twitter), and we needn’t wait around for professional confirmation when the evidence is right before our eyes. On the other hand, only pros can confer joseki status. This is by definition—joseki as a standard and studied sequence that is considered balanced for both sides. Of course, pros freely draw from AI analyses. But I feel that amateurs are not positioned to do that work for them.
I’m not sure why you keep asking me to “answer the question”, when my answer was already given in the other thread as you noted. I’m just writing down my thoughts and feelings for the record (since this is the right topic for them), and for other to tear them apart so we can figure something out hopefully.
I feel like we could make tens of topics and have detailed and long discussions on each, but I also don’t want to spam the forums. Maybe we should have some sort of chat going, and make a forum post with sorta the general conclusions and recommendations we come up with so there isn’t a wall of text to read?
Anyway, I realize now what you actually wanted from us (and sorry for rambling), and yes I would just leave it at “pros must identify joseki”.
I think as someone who might use the joseki tool, I would want the ability to turn on or turn off computer contributions.
The reason for this is that while the computer is capable of playing joseki, particularly intricate or complex ones, I, as a DDK player, will not benefit from playing these joseki. In fact, attempting to learn them would be foolish because the deep reading that the AI uses is not explainable by the AI, unlike Pro contributions.
Ideally I would like the ability to turn on and off computer contributions and potentially search for comments on joseki by specific Pro players commentary on those joseki. (Normally this include variations and key shape points that should be looked after.)
Obviously there are those who will want access to the computer additions and those who do not. Rather than exclude computer joseki, I would prefer having the option to view or not view them and to have an obvious tagging system for AI inspired or derived joseki.
In case it is not immediately obvious in the documentation, the Filter feature allows you to filter by source. This allows you to look at Joseki only from a non-AI source.
Right now it’s not perfect, in that you can only select one source to look at, at a time.
Hopefully before too long I’ll make it more flexible.
My take is that for current joseki, professionals have the same(or way better) AI tools as amateurs. Why don’t we let them dedicate their study hours and matches to figuring out what is joseki? I don’t think AI has changed the process at all really.
For older stuff when pros did not have AI tools available, it could be a useful check if used in moderation.
edit: if there are AI contributions, I think documentation would be important. The positions researched, which weights, how many GPUs, # of playouts… etc.
This is essential IMO. I’ve tried to be very careful in labeling moves as ‘GOOD’ or being ungenerous with comments to any moves. It can become dogmatic so quickly. The name is of the tool is explorer and I think that spirit is lost if you are placing judgments on every move.
This is why I like the AI 3-3 revival… not long ago there was “the one correct response to a 3-3 invasion” with only a few edge cases where something else would be chosen, but now there are many options that are regarded as good
The issue is NOT about whether or not to include joseki produced by AI, but whether we should be allowed to decide on what is joseki just based on an AI’s winrate that we explore on our own computers.
Do us weak amateurs have the right to choose what is joseki and what isn’t, before professionals have reviewed and vetted the algorithms choice?
No reason to jump the gun, in my view. Professionals will study and play the moves if they are joseki.
If we did rigorous and well documented analysis with a variety of super-human weights (who will tell different stories which have to be translated into human thinking) and we did manage to come up with a joseki… all we gained is months or a year of having it before a pro did the same thing and it sees regular play. That’s the best case. The likely case is just a ton of bloated sequences that are played in one or five professional games that nobody bothers to remove… (Josekipedia)
There is no instance of where I would want any algorithm to run without human oversight and having the final say. This is no different. The only issue is whether we are qualified or not to choose what is and isn’t joseki.
