This is exactly what I’ve been waiting for.
1 Like
(Bot on single GPU will give handicap stones to top professionals)
2 Likes
What’s special about this?
1 Like
It might be a new AlphaGo.
AI Go is long past AlphaGo. The current best AI are already 2+ stones stronger than AlphaGo, even when run on consumer hardware.
What are they called and where can I find them?
KataGo is the most popular as far as I know. You can download the latest model here or download a client that comes with it out of the box like Katrain.
1 Like
I want to point out two things.
- Back in May 2024, analyses were already showing that KataGo’s first 28-block weight had surpassed FineArt. Just as AlphaGo disappeared from public view about 10 years ago, FineArt too has gone more than 4 years without any further improvement. The difference is that KataGo kept improving after surpassing FineArt, and it has continued to grow for another two years since then.
- I heard that the sponsor of that event is Google or a big tech company on Google’s level. Professional player Cho Hyeyeon mentioned this during a Kiseon Championship broadcast, so it sounds like someone from that side may have contacted her. If that’s true, it really could be a new AlphaGo.
1 Like
Despite what the commentator said, not really impressed so far. Feels like it is adversarily trained with human games, and playing lots of trick moves instead of good moves (so an attempt at Go devil? play the players instead of optimal?)
Against pros who rarely play handicaps this might work, but for those who played or teach in handicaps or studied old games with high handicaps (there were literally books in the past, called 3 handicaps, 4 handicaps, etc.), or just trained with AI using handicaps (like those trained on foxwq with fineart consistantly), then it is pretty clear how to handle high handicap games without any fancy play or tesuji, but just focus on playing thick and solid, don’t be greedy, and make sure all the groups are safe.
1 Like
:
Those two 5 handicap losses are played “unprofessional-like”. In the first game, black just peeped into the opponent to help weak white group to connet, and damaging their own group on the other side, attach a stone to a strong group, get cut, instead of extend, immediately atari (really bad habit), and connect to save like four unimportant stones, instead of saving a ground worth at least 30 points on the side, etc. (like several kyu, low dan level obvious mistakes, and keep following unreasonable overplay).
The other one is more solid, but mostly due to poor whole board judgement, giving up a fat corner helping white get thick, in exchange for an one space jump from 4-4 and loose framework (trading certain points with unknown gain), then afterward, when white encroaches the corner and play inside the framework, go to capture 2 stones in gote, and give up the corner and the side.
And the last game with a 3-handicap time-out. Feels like a technical issue.
2 Likes
AlphaGo was special because it was the first AI to beat top human players. This milestone is done and won’t happen again.
Since then there’s been many new AIs stronger than AlphaGo. I’m still not sure what’s special about this one?
https://online-go.com/player/2002470/cGreenSpot
green spot on avatar is small
https://online-go.com/player/2002471/cBlueSpot
blue spot account already created
the next test will be with bigger model?
1 Like