I believe this is generally true, but unfortunately those who do like to cheat with bots tend to be strong dan players. Within the 4d-9d rank range on OGS, I feel that somewhere closer to about 50% of the players use AI… It’s frustrating for the actual high dan players, who want to play humans and have to face bots all the time.
So, the 5% don’t affect 95% of the people on OGS, but almost 100% of the stronger player base.
But, @hexahedron, I think this is a very interesting project and probably can be used very helpfully for teaching and reviewing purposes.
How does it decide between playing a move that is more human-like but bad, and a move that is good but more AI-like? Is there some gradient such that you can decide how much “humanness” you want to include in the decision, or is it really only trying to imitate humans as closely as it could?
Do you think it’s possible to guide it so that it plays in certain styles (like aggressive, or building big moyo’s, etc)? I guess the problem is procuring a suitable training set for such a thing
Also, on the flip side, this may be actually useful to catch some of the high dan rank bot users: if this version of KataGo can recognise human-ness, it can be used to flag both those users using this version of KataGo to play (which line up better with human-KataGo’s predictions than an average human, most likely), and those who use the AI-like KataGo to play (since it makes such unhuman moves, which human-KataGo can recognise)
I’m very happy that you released it.
Yes, some people may exploit it for botting, but I think that’s not a good reason to withhold this feature. For me the potential benefits clearly outweigh that potential downside.
I’m hoping that Benjamin Teuber will incorporate this feature into AI-Sensei soon.
Reading the Reddit post, it seems like this new feature acts more like a “value network” than a “policy network”, i.e. it evaluates moves by simulating human intuition of some level rather than human reading of some level.
Its reading is still superhuman, so as you allow it more visits for move selection, actual reading would gradually overrule weaker human moves because it discovers refutations for those. So when the number of visits increases, its play would gradually become more AI-like and less human-like, regardless of the human intuition level you set it to aim for.
To simulate human reading of some level, one may need to carefully calibrate the number of visits for each level to come close to 50% winrate against humans of that level. And perhaps this is not easy to do for DDK levels, because even the minimum of 1 visit may already result in better “reading” than actual human DDK level reading.
Still this may lead to the issue of some nerfed AI bots (although perhaps to a lesser degree), having a bizarre playing style of playing god-like most of the time, mixed with some percentage of silly game losing blunders, to reach 50% winrate against a range of weaker human players. Even though such bots’ ratings seem fine, using such bots as sparring partners seems not very useful to me.
At the same time, if at the end of the day you felt exactly as if you had played a human opponent who just slightly outplayed you (like one would normally expect), this doesn’t sound like that much of an issue for a Go server.
It’s still cheating and I’m not advocating for this being tolerated; but still, the actual impact for the opponent doesn’t seem that concerning.
Lol did you just say cheating isn’t an issue? And some people agree?
Firstly, if I want to play an AI who plays like a human, I will play the AI directly instead of a human who is cheating using AI who plays like a human.
Secondly, if many people do that, the whole ranking system will be meaningless. It will be ranking the AIs instead of ranking the humans.
Thirdly, even if it doesn’t matter for casual games, it will be a serious issue for tournaments, especially those that are still quite loose in cheating prevention.
In light of the arguments that (1) botting is rare, (2) it doesn’t affect me, so I don’t care, and (3) the teaching value of KataGo v1.15.x makes the risks unimportant, perhaps it is time to consider removing botting from the list of reportable violations. If the majority of OGS players don’t care about botting—we would of course need to try to find out whether that is true—then why should moderator time and effort be wasted trying to mitigate botting? On the other hand, if mitigating botting still has value, then it is unjust to mock concerns about how KataGo v1.15.x might increase botting at weaker playing levels.
My view is that botting is wrong if you don’t declare it (and OGS facilitates that declaration in the form of marked bot accounts), and it takes away from the qualitative value of the game if one is not playing against another human bearing the image of God, but that a tool which can simulate weaker human-like play is far too valuable to pass up on even if it (in the worst case scenario) forces us to accept that undetectable botting will become the norm
Its boring for most humans to manually copy moves of bot. Few know how to automate process of botting in not bot account.
So currently, there are still very few online players who use bot instead of playing themselves.
I don’t agree with mocking concerns about botting, I have those same concerns, but I’m open to the idea that the best solution may be to simply accept it as inevitable without actually agreeing with it
I’m not sure what’s the point of playing online anymore if botting is so prevalent. You might as well just play against KataGo at the human level that you desire.
I don’t think the new Katago net will make botting much more prevalent. People who want to play will continue to play, and people who just want to win by copying bot moves can already do that.
The only way to prevent cheating with AI is to use a webcam that shows screen + hands. And even then, you can’t prevent cheating completely.
Maybe next step is to combine kata with some language model so it could also write its reasoning for each move. That could be the perfect teaching tool ^___^
Does that mean that all previous warnings for botting on OGS were folly? Sounds like an argument for removing botting from the list of reportable violations.
Yes, you can. Don’t play.
More seriously, does anyone really imagine that cheating can be completely prevented? If not, the comment seems pointless. This calls to mind the old saw about the perfect being the enemy of the doable.
I don’t see anybody here that claimed they did not care about botting, let alone that this would be a majority opinion. I assume this is not referring to my comment as I explicitly reiterated it should not be tolerated.
I also don’t see anybody here who mocked anybody else, so don’t understand the defensive stance either.
I don’t think that botting will double (for instance).
No. As long as “police” is there to maintain “criminality” at low level, I’m OK to introduce a new “weapon” that may increase “criminality” a bit but whose positive effects will likely outweigh the negative ones.
Yeah this is like saying if you don’t want to get your things stolen or don’t want to get hit by a car then don’t leave the house.
These things can’t be completely prevented. That doesn’t mean they should be tolerated. At the very least there should be measures to prevent them from happening.
