I’m reviewing an AI variation, and don’t understand why move 6 is what looks like an unnecessary defensive move. I would love to be able to play the move I think is correct, in a variation, and have the AI evaluate that position, showing a new playout.
I think this is a similar request to this:
Is this feature already available? If not, is it on the roadmap?
I think this is a great improvement for beginners users to understand the reasons why AI suggested the variations.
It would be a great feature, like it’d make uploading finished games that little bit nicer to review for instance.
However I expect some resistance to it as an idea.
Eg Prevent SGF download until game conclusion - there was discussions about how hosting ai on the OGS website might reduce the barrier to cheating. My assumption is that at the moment it’s probably quite painful to use the ai to cheat - I think you’d need to (re)upload an sgf and run analysis on the whole game, for every move you wanted to check.
In principle if you could ask the ai about any move/variation it might make it easier to cheat on ongoing games.
Obviously you wouldn’t make it available for ongoing games, but you’d probably have to follow some ideas in that thread - make sgf downloads unavailable, maybe you could do a check in uploaded games, which sever the game came from and refuse to allow a move by move check if it’s OGS, or maybe it could do a board comparison to ongoing games (+flips+rotates) to see if it’s an ongoing game before running the analysis.
There’s still probably ways around those (and really someone could probably just download a strong ai anyway and use that) - but my guess is that “not making it any easier to cheat” will be high on the list of reasons this might not get a go-ahead.
I was looking for the mentioned feature too. I’m coming from chess where the engines can be really helpful for exploring ideas and seeing refutations you wouldn’t spot on your own. If ease of cheating is a worry, I would recommend reaching out to the dev team at Lichess. They run the preferred online chess site for most, and it has full engine integration and great cheat detection.
Lichess could be a good source for ideas on how cheat detection could be implemented.
If something like that does get implemented on OGS, I guess the next question/roadblock would be how much resources it takes to run an interactive ai analysis for X amount of people at a time, as opposed to just the one off full game analysis.
I don’t know much myself about it, just trying to help the discussion along 
Moderator use various approaches like the above.
I’m in two minds about whether it’s helpful to even discuss it, because it is an ongoing “war”. The more you tell people about how you counter it, the more they counter those techniques. Sure, whatever is discussed a botter can think of themselves - but not all of them do, so I find it unappealing laying it all out in the open.
Also may be a good feature to allow play against strong bot in variants, for donated accounts. so try many variants and their possible outcomes. Also add bot to joseki dict. Now i not see much use for ai review, because it not making a tree of variants, just some (random ?) branch
You can download leela or similar AI programs to review for you
If you can to play certain variations you can, each move will give you certain winrate percentage, it will analyze which move is best then shows you variations for specific moves you want
AI only give you variations, in order for you to understand the variations, you yourself also need Go theories and reading skills
Certain variations you don’t understand is mainly because you lack the logic of why this variation is superior to yours
Whether the logic is about whole board, sente/gote, local territorial/influence gained
This all has to be understood by yourself first to understand the reasoning behind why such variation is superior to yours
Like for example, invading 3-3 immediately was considered bad before the AI came out, because you get surrounded and gives too much influence immediately so it was considered bad
But once they saw the merits of 3-3 immediately (sente+territory then neutralize the influence with other plays), the Go theory didn’t change, what’s changed is the way people think
Sente became the new big thing, it was important before but now it’s even more important
Every game if you didn’t fight for sente advantage, then you might be wrong
Some AI even has cosmic style, i forgot the AI’s name but it was playing random center moves in the opening, from that we can conclude, opening theory(corner, side, center) maybe isn’t that important anymore.
Fighting(reading abilities) becomes the most important, if you fight properly, you can play random openings all you want
Anyways, treat AI as a tool, don’t become a slave to them, understand meaning behind them, maybe you can come up with your own and better variations one day
I’ve tried reviewing with 2 different AI and the variations they gave were totally different, but what they gave me was insights to how i should play instead
Knowing i should play certain variations doesn’t do me any good, i might as well do tsumegos and to improve my reading abilities, but the idea/concepts behind their variations is what i think is the most important
I think one problem with this is that the ai loves to tenuki. Sometimes it ‘finishes’ joseki sometimes it stops early because there is something more urgent on the board.
Unless there’s a setting to force local play on some of the ai (like with an m x n grid/window on the board) it probably wouldn’t be that useful.
If Lichess level anti-cheat is implemented, I don’t think interactive AI, which sounds super resource intensive, would even be necessary. Just implement tree AI review. The only argument against it as far as I can see is the potential ease of cheating, and personally I don’t even see it as that big an issue. People who want to cheat will find a way, and you could even hide the feature behind the $10 pay wall or even higher so that if people are misusing it at least they’re also funding moderators to find and remove them.
AI reviews on variations need not be to the point of interactive AI, but it would actually make the feature far more useful and user friendly. Heck even if you limited it to 3 trees per game or something; the ability to go back and question the AI and test its reasoning is, in my opinion, the main reason why AI is such a great learning tool. I would really like to see this implemented in some capacity.
Totally agree. I understand that implementing this sort of thing would require more server resources, and would be willing to pay for using it.
This topic was automatically closed 375 days after the last reply. New replies are no longer allowed.