I completely agree. KataGo analysis shouldnt be used. OJE variations should be from professional play. Those were added in the early stages of OJE before we had a lot of discussion about what should be included.
Everything that has been added in 2020 has been from professional play only. With maybe <10 exceptions (egregiously bad human mistakes, or articles/videos by professionals about AI moves) which are noted in descriptions.
We should make a list of AI only variations(usually from August to September of last year) and purge them. It is much harder to remove things than add them which is why I was so strongly against adding KataGo variations at any point.
I removed r18 and q18 because they donāt fit the inclusion criteria. I also corrected q17 to ideal.
With no pro play, itās hard indeed to call r18 and q18 ācredibly established.ā Although it was around a year ago, I believe I added these because of their similarity to the more common pattern, which I added at the same time. In retrospect I was mistaken to include them.
OJE is more of a hybrid between Kogos and Waltheri though.
Everything that Iāve added is based on observation from a more complete database than Waltheri - while trimming it a bit so that moves played <5% of the time are not cluttering the position search. Any variations that donāt meet that criteria or do and are mistakes are sourced (mostly by video sources at the moment).
AI analysis does come into play to blunder check some of the āTraditionalā joseki that have died since AlphaGo. Anything <5% mistake (in several test positions of professional games) is still treated as ideal. I believe I have made a few notations where a common joseki is >15% mistake and labeled it as such but my own judgement of joseki is limited to observations of the situation it is played. (Example of a joseki mistake)(Example of a situational observation)
For trick plays or common amateur mistakes, I/we had some discussion of loosening the AI restrictions. For instance, if a trick play appears in a book and the result of being tricked is actually not a disadvantage, we would relabel it. Or for common amateur mistakes, take a look at any of the one space low pincers where I downloaded 500,000 SDK games from Fox and found the biggest blunders and did use AI and some old joseki books) to sketch out a skeleton of refutations.
ā¦ and the way this quote appears makes it seem that this was a description by bugcat of OJE.
I want to refute this description (whether or not that is what bugcat was originally saying).
I want to refute the idea that ā[OJE has ] seamlessly shifted from the idea of providing a medium by which the student can access professional knowledge, to the idea of the players of OGS (all amateurs) dispensing their own judgements as factā.
This is specifically not the caseā¦
Of course there was. And itās decisions are documented here, and specifically here, with discussion and clarification of the intent of these words undergoing intense discussion by those actually contributing, here.
One specific principle is that
This point specifically refutes the implication of the quote above.
A āCredible Sourceā is " if we have a source for it who is or was a professional player"
Further:
" Use in AI-only play alone is not enough - we want it to have passed through the filter of āhumans have worked out how to use itā."
There was a period, while this was being worked out, where AI sequences were deeemed as ācredibleā, and this may explain why some of them are in OJE, in violation of this principle. Such inclusions should be seen as an aberration, and Iād support weeding them out.
Having said all that, there is one situation where AI inclusions are specifically allowed:
"Mistakes and refutations do not require the same strict level of credible sourcing, because practically speaking this level of information is not readily available; however they must be validated by AIā¦ [and] ā¦ An unsourced sequence of moves may be included in the Joseki Dictionary if that sequence starts from a misplay that a human player is likely to make, contains only those moves necessary for punishing the misplay, and has been validated by AI.
"
IE weāre willing to use AI to help us access refutations of mistakes (because this is a helpful thing to do), but not to establish āwhat is Josekiā.
Youāve seamlessly shifted from the idea of providing a medium by which the student can access professional knowledge, to the idea of the players of OGS (all amateurs) dispensing their own judgements as fact.
I used you to refer only to Finrod, who I was speaking to and about the conversation we were having, and not about the OGS Joseki project. But aesalonās comment doesnāt seem to rely at all on misquoting me.
I donāt actually think that OGS Joseki has, in any significant way, fallen into that pitfall at the current time.
I should note that Iām not 100% happy with this poll, as my own cutoff for ācommonnessā would be more like 25 games. But 50 seemed a more reasonable poll distinction.
I really donāt know so I canāt vote.
This project is ambitious.
Before AI all was easy. Just listen the pros. Now pros donāt talk yet that much. Joseki dictionary integrating new AI input are not yet publish. So what?
Maybe itās just too early and impatient. Big pieces of the āoldā knowledge didnāt yet pass through the examination and the test of real play. Itās really like a revolution, how can you say if something is still joseki when the reference are pre-AI ? How can we reject the not tested ones if afterall they are still of great value, who knows?
Of course itās a great project and we all are impatient to read it. But we want to go quicker as the informersā¦
I only skimmed through the topic, so forgive me (and let me know) if I missed something
It seems several players are against the idea of listing AI joseki, but I donāt understand why? Currently AIs are one of the top sources for us, why disregard such a great source of knowledge/ideas? The joseki explorer comes with source metadata and filter optionsā¦ Simply ignore AI sources if you need, but why consider limiting it for all?
Sorry for full quoting bugcat. I just wanted to explain ways we are trying to be different from the other noseki tools basically.
Because it is a major rabithole. AI are extraordinarily sensitive to the global situation so how many positions does it need to be a blue move in for it to be included? What is the cutoff in terms of %/katapoints? There are many other issues but including AI positions is one of the worst things about Josekipedia.
Professional games are a good filter from AI to us.
Thatās amazing because in that case itās almost the death of any attempt to create a selection of josekis: arenāt even AI more pro as pro?
I mean each time I read professional, I want to replace the word by AIā¦
It comes down to the meaning of ājoseki,ā which we defined as ācrediblyā āestablishedā sequences. For OJE purposes, amateurs arenāt credible:
I feel this is even true of an amateur wielding AI, otherwise novice DDK players could contribute ājosekiā simply by reading AI output.
Weāve also agreed that we need to know how and why a sequence is used for it to be part of the dictionary:
I donāt like the idea taking professional games out of context though (filtering games by just the local position) for whether something could be joseki.
We could say oh hey this move appeared in 47 games on waltheri is that joseki?
Probably more useful the other way Bugcat mentioned, to potentially rule out some other moves
Regarding the poll I think the best judge is how the sequence fairs out on an empty board. Is it an even result for both sides. Proās can judge this quite well. Aiās (at the moment) probably a bit less so because a lot of them just want to play away in an empty corner and win the game.
The percentage matters more than the number does. A move played 0.06% of the time isnāt credibly established to be among the best that you can do locally. Put differently, professionals chose to play somewhere else 74,124 times. 47 means a lot less in that contextā¦
Letās remember that there isnāt just ideal move and āmove we donāt addā. We do have the good move category as well. It could be that good move could be allowed to require a lower level of sourcing than ideal move, but this would require its own discussion.
We could also make more use of the categorisation potential we have by creating more categories to express finer detail of sourcing and opinion ā and we could build in a filter tool to allow the user to ignore some of these categories and thus tailor their study to their own preferences.
eg. classical for moves with low play since 1950, pre-bot for ones with low play since 2016, untested for moves suggested by pros that havenāt been well-tried yet, and so on
I still think AI suggestions without professional commentary are redundant, though, since the user can just run the AI at home with whatever global context he or she wants to give it.
I think that variation contains an earlier mistake. I donāt think whiteās attachment at R16 can be considered a normal move with blackās one-space jump O16. Itās really not the same as when black responded low (O17) on whiteās approach.
Josekipedia says R16 is situational (in the comment) and in Waltheriās R16 is played less than 1% (the exact percentage depending on the size of the search window) and R16 is not in the top 10 of most frequent moves.
Yet, R16 is marked as IDEAL - traditional in the OJE.
Because: " Use in AI-only play alone is not enough - we want it to have passed through the filter of āhumans have worked out how to use itā. "
I donāt think that the source itself (AI) is being disregarded.
Rather OJE contritbutors are asked to find Pro level validation of those great new AI ideas before adding them.
Iām pretty sure there entries like this validated by Yeon Woo, where she has said ācheck out this cool Josekiā and provided analysis of itā¦ that makes it āCredibleā.
