… I was quite surprised to see 6(!) moves in the (supposedly) joseki continuation that lost more than ~2.5 points, according to the level IV AI game review.
At my level (3d) I would qualify score losses of that size as fairly significant mistakes (not just inaccuracies), that basically disqualify this continuation from earning the joseki label.
black 1 lost 2.7 points
white 2 lost 2.4 points
black 3 lost 3.6 points
white 4 lost 3.5 points
black 9 lost 3.9 points
white 10 lost 2.6 points
I know that the OJE policy is to basically ignore AI variations and evaluations, unless corroborated by pro games and/or pro analysis.
But I feel that this policy is a bit problematic for variations like these, marking them as joseki while AI shoot numerous holes in it.
In particular, black’s E18 tesuji isn’t discussed in the OJE, nor does it appear in Waltheri’s. Maybe pro analysis of this variation has been a bit too shallow and we should include new AI insights to the OJE? After all, this variation hasn’t appeared in pro games for 20 years, so this joseki is basically dead and pros won’t be investigating it anymore and mistakes will never be fixed under the current OJE policy.
About this particular case, now that I see the AI critique, I have some vague recollection of being shown black’s E18 tesuji in the past, so maybe strong amateurs and pros know about it, but sources for this are just too hard to find for the OJE composers?
I believe the “joseki” in question is a variation that requires certain circumstances to be feasible for the side with the 5-4. It is not a joseki that yields an even result in most positions. Most of the time the hane on the outside is superior …
… and even before that, black should probably attach in the corner at 3-3 in most cases. Of course it is fine to go for these variations if your style is to play for influence.
The inside hane can be a powerful weapon if the side with 5-4 has reinforcements nearby. However it is rather easy to dodge these cases, for example by invading at 3-3 instead. I guess the joseki in question is rather outdated due to these and similar reasons.
But from there on, the one-way street continuation shown by the OJE seems to become quite dubious according to AI. If AI stuff is not allowed in the OJE, perhaps the OJE should stop at the point above (white S19), or at least note that everything after is outdated and perhaps dubious, instead of continueing and calling the following result “joseki”:
I studied joseki most actively in the early 90s, using Ishida’s Joseki Dictionary, dating from the 70s. The inside hane did appear quite often in my games back then (played by me or my opponent), although move 10 in my 2nd diagram would usually be played at F17/O17 (I think that had to do with the E18/P18 tesuji), which can still be found in Josekipedia:
I agree. I just think that the process of labeling and editing should involve this kind of discourse that you have started rather than one editor with an AI open making a decision (no matter how easy it may seem, in this case)
Joseki are common sequences, not necessarily “correct” and certainly not under any and all circumstances. Unless you have an impeccable understanding of the entire board situation, you will never be able to choose the ‘proper’ joseki.
These sequences make sense and prevent you from losing immediately (by trading something of value for no compensation at all). Viable returns on investment in human games are things like “my group is settled”, “my opponent’s group is not settled”, “I gained influence here”, “I negated my opponent’s influence there”, etc.
A compromise would be to add Kata’s moves in a different color, which would… save people the hassle of having to recreate that position to check it with their own version. I, for one, prefer the KJD way of doing things: lots of actual commentary and notes for each move or sequence.
As for the suggestion of including “bots don’t like this”… sure, as a comment that’s fine, but unless you can explain WHY bots don’t like it, refrain from using cheap “reasons” like “it says -2.5”.
I said that I like KJD’s way of doing things. Nowhere do I claim that the actual information in it is more accurate or that I’ve seen all branches and found them to be suitably well documented. Overall, KJD has high comment density and in order to know what to even consider when picking alternatives, short notes like “this trades w life in the corner for B influence toward the middle” at the beginning of a sequence is probably more helpful than to have a move listed that’s doing the same thing but slightly more efficiently.
Okay, so in the interest of clarity, what do you think should be done now and in the future with sequences bots don’t like (or like, again), and what standard 19x position should be used as a reference for what bots like? I’ll help you out with a few suggestions.
Rename OJE to OJA (joseki archive)
Split OJE into human vs bot variants: OHJE (German and Dutch users will love this - “Ja, dit is een OHJE’tje”) and OBJE
Isn’t it complicated to say categorically that AI disagree with a specific joseki? What board state do we use to test the joseki? How vehemently does the AI need to object? What about situational joseki that are correct or playable only in specific circumstances?
I agree that we should be incorporating AI somehow in evaluating joseki, but the specific algorithm needs to be carefully thought out. If you could have Kata analyze all the pro games in which the joseki occurs and spit out statistics (average score, variance, etc), then that might be very useful.
I’d say that any move that loses more than about 2 points according to AI in common early game positions (say just 4-4 stones in the other corners in a parallel and cross fuseki to account for ladders), should be up for debate (to potentially mark with “Kata’s seal of disapproval”).
I can’t tell if you’re being serious here, but one way to find group consensus is to ask knowledgeable individuals to share their theories or opinions on the matter. This can then serve as a framework for a larger discussion.
I didn’t mean to demand that you, smurph, solve everything by yourself, and though I replied to your comment, I did not intend my comment to be addressed solely to you. I was just discussing how one could create a “Kata’s seal of disapproval” as mentioned in your post.