Black hole Joseki

I would like to know if it was possible to add to the Joseki dictionnary the new “Joseki” who seemed to have appear thanks to the “black hole” AI. Currently, the joseki dictionnary only gives the old Joseki (the ones from Chô u and So Yôkoku I bilieve). Though, as pointed by Alexander Dinerchtein in this video "Black Hole" Fuseki Lecture by Alexander Dinerchtein 3p eng subs. - YouTube, the AI seems to play other Joseki (see for example at 6 min).

I think it would be nice to at least add to the Joseki dictionnary the Joseki pointed out in this video. Would it be possible ? (I can do it myself if you want)

Thank you :slight_smile:

Edit : for clarification : When I say Joseki, I do not refer to the black hole Fuseki, but to the approach of the corner by black after the fuseki has been played

That’s a fuseki that requires both players to agree. It’s not a joseki and you shouldn’t ever expect to see it unless both players agree to it before the game starts…

It’s more an opening (fuseki) strategy than a joseki. Also, it mostly disregards white’s moves, so there will be a lot of variations with white mixing all kinds of moves in between black’s moves. So there is not 1 sequence.

No I think you are misunderstanding me. Of course, the black hole is a fuseki. What I mean by joseki is, given the black hole opening, the way to approach the corner for black (usually taken by white) and the following sequence, after the 8 four first moves of this Fuseki has been played.

And obviously there is many sequence possible, but the AI seems to often play the same “Joseki” (which is not the one currently in the dictionnary)


OK, but the policy of OGS Joseki Explorer is to ignore AI sequences that have not (or very rarely) appeared in pro games.

There are not many pro games where 5-7 was played in an empty quadrant (only a few games from 2012 to 2015, mostly by Ohashi Hirofumi or So Yokoku), so there is really not enough data to designate some sequence as standard (i.e. joseki).

When I look up pro games where black 5-7 was the first move in an empty quadrant, the white approach move in the OGS Joseki Explorer is what pros tended to play. For example move 10 in So Yokoku (8p) vs. Yamashiro Hiroshi (9p) | Waltheri's go pattern search and Cho U (9p) vs. Hane Naoki (9p) | Waltheri's go pattern search.


I think the balance we have right now is pretty good.

The OGS Joseki page Play Go at! | OGS links to a Waltheri local search, where the user can go to explore the actual professional variations instead of interpreting them through OGSJ.


Is this just from the AI vs AI games on computer Go server? I don’t really think we want to make an exception and use those patterns. (Just my opinion though) There have been a few exceptions where the variation was interesting and useful so I don’t know.

If it were a bot playing vs professionals then I would be open to the idea of making an exception because to me it would check some boxes as potentially useful and interesting.

1 Like

This is, indeed, AI VS AI games on computer Go server (it was this AI who reached the first rank in the AI competition at the beginning of last year). Nowadays, no pro plays the black hole Fuseki in official games (too risky, and too much time to spend to learn to play it properly, time which could be used to study “traditionnal” Joseki for example).

I wasn’t aware of the rule that every Joseki have to be played by pros to be present in the database, and it clearly do not check this condition. So I’m sorry about my message, I wasn’t aware of that point.

Though, I find it a little sad since the black hole fuseki is more a “fun” thing (so not played by pros), but with very specific Josekis. Thus, amateurs who want to play this fuseki for fun, but to play it properly (since there is no fun by being crushed by your opponent) will never have access to these Josekis since it is likely they will never be played by any pros. Though, even if these Josekis are not played by pros, that’s not because they do not work, but because pros do not play this Fuseki. Since nowadays AI are better than pros (and specially an AI able to be the best AI last year), I thought it could be legit enough to be present in the Joseki dictionnary.


I’m glad you linked the video in any case. It’s an interesting watch (I don’t think i’ve finished it yet). The subtitles are very helpful :slight_smile:

Do we know what happened to the black_hole bot? Did it eventually get beaten regularly by the strong networks or was it just an experiment? It’s not there any more I don’t think?

Well to be fair I never followed AI server very closely and became aware of this Fuseki very recently so I do not know :confused:

If you want my guess, this AI was performing well mainly because the traditionnal AI have never had a chance of training against black hole opening, so they probably were playing a bit unoptimal. After that the traditionnal AI trained against the black_hole AI, and they probably learned to manage how to deal with this particular Fuseki. It was clearly an experiment, since this AI was playing only both black hole and white hole Fuseki, the latter being clearly a bad opening (but it makes the result of the black_hole AI even more impressive). It would be very interesting to know the winrate of the black_hole AI playing black before it disapears though.

1 Like

It’s played some games this year.

v7 is rated 5th under 4 different versions of Katago.


Ah right I was looking at this page (CGOS) 19x19 Computer Go Server

That’s why I couldn’t see it!

It was not because AIs were unfamiliar with the fuseki. Likely they’ve always been able to deal fine with it, and the black hole opening has always been a not-so-good opening (at top levels) that is not particularly harder to deal with (for AIs at those top levels) than any other not-so-good opening.

The probable reason that black hole bot did fine for a while was because it was either using a bigger/stronger net (possibly), or using more hardware than the other bots (pretty likely).

Think about it this way: if a human player merely one stone stronger than you wanted to troll you, they could open every game by playing, say, the 2-10 point, and still have the advantage on average. As bad as 2-10 is, being on the second line, it’s still not even remotely close to bad as passing and giving up a whole move, so being one stone stronger than you means they’d still have a clear edge on average.

But this doesn’t mean that variations based on the 2-10 point are worthy of inclusion in joseki dictionaries. And if you lose the game, it has nothing to do with you being unfamiliar with the 2-10 point or needing to train more against it. And probably there’s nothing special about specifically the 2-10 point being hard for you versus any other bad second line move. It’s simply that the opponent was stronger than you by enough in the rest of the game to win despite their self-imposed handicap. :slight_smile:


You can generate your own corner sequence collection with any move that you like
just play 5-7 point in all 4 corners, then

you can do it with each corner separately and receive 4 bot moves after each game
your collection will grow fast and it would be even more useful than professional joseki because instead it would include opponents that you really play against - OGS players

2-10 maybe is not effective enough. But any strange move that is close enough to corner may really confuse not professional human.
And if you will train a lot with using a bot, you may start to use it in most games more efficiently than standard fuseki.

I’m going to explore 5-6 next.
Some may say that its sandbagging, but its not true. I didn’t play normally for very long time. So if I will use normal fuseki, then I will play much worse - I forgot how to use it.


Whatever you play, there is no sandbagging if you are trying to win instead of a fake loss.