Article - AI is rewiring how the world’s best Go players think

[towforce]

I do not know that go is less computer friendly game than chess.

Well computers were unable to beat top humans until AlphaGo took on Lee Sedol in March 2016 (I can't believe it was 10 years ago!): that's my basis for saying that Go is not computer friendly.

[Uri Blass]

I do not know that go is less computer friendly game than chess.

Well computers were unable to beat top humans until AlphaGo took on Lee Sedol in March 2016 (I can't believe it was 10 years ago!): that's my basis for saying that Go is not computer friendly.

It was not computer friendly before 2016 but it seems today it is not the case.

[tapio]

Lee Sedol, 2026-03-09:

https://www.youtube.com/live/5W28E5Fux5 ... H8JXp8forg

10 years from the confrontation with AlphaGo in 2016. This is the entire video of the event with Lee Sedol 9th Dan and Enhans.
An era in which AI and human relationships transition from 'confrontation'to 'collaboration',
Enhancement has unveiled an Agentic AI OS based on Ontology technology.
See the scene where five AI agents are running real-world operations on one OS.
March 9, 2026 / Four Seasons Hotel Seoul
Sponsored by: Anthropic, NVIDIA, Microsoft

[Eelco de Groot]

If you would like to finally learn the game (myself, sadly I keep forgetting the rules ), or get some game advice from superhuman Senseis *, or just enjoy watching a very strong Go engine, the very latest trained network from KataGo run 1 seems very strong. Provisional rating is 14156.4. Eat that Stockfish but needed at least 1000 more games to have some confidence in the rating. The error margins are much larger (than ratings in computerchess especially) because typically every network just has about 2000 games. Apart from their training games, those are not listed here I think. Have not really followed the KataGo discord channel yet I just occasionally download the latest new network if it seems strong.

kata1-zhizi-b28c512nbt-muonfd2 is the name of the network, can be found here: https://katagotraining.org/networks/kata1/

*) Sensei in Japanese translates to something like "teacher, instructor, master". Strongest players like in Judo have a Dan rank; Professional Players (Pro, "p"): The elite level (1 dan pro to 9 dan pro). In East Asia, they are known as Kishi (Japan), Gisa (Korea), or Qisheng/Guoshou (China, ancient/professional).

[Uri Blass]

If you would like to finally learn the game (myself, sadly I keep forgetting the rules ), or get some game advice from superhuman Senseis *, or just enjoy watching a very strong Go engine, the very latest trained network from KataGo run 1 seems very strong. Provisional rating is 14156.4. Eat that Stockfish but needed at least 1000 more games to have some confidence in the rating. The error margins are much larger (than ratings in computerchess especially) because typically every network just has about 2000 games. Apart from their training games, those are not listed here I think. Have not really followed the KataGo discord channel yet I just occasionally download the latest new network if it seems strong.

kata1-zhizi-b28c512nbt-muonfd2 is the name of the network, can be found here: https://katagotraining.org/networks/kata1/

*) Sensei in Japanese translates to something like "teacher, instructor, master". Strongest players like in Judo have a Dan rank; Professional Players (Pro, "p"): The elite level (1 dan pro to 9 dan pro). In East Asia, they are known as Kishi (Japan), Gisa (Korea), or Qisheng/Guoshou (China, ancient/professional).

rating of 14156.4 tells me nothing without knowing the rating of the best humans in go.

It seems that the rating of the top human is 3851
https://www.go4go.net/go/players/rank

[Eelco de Groot]

No you are right of course, there is no way to compare that to human ratings unless you let them play some competition. (In chess , 2000 was set for strong clubplayers but I have no idea what is used here, 0 is maybe something like random moves? I do not know.) The KataGo ratings are only in relation to other KataGo networks (in this run 1) and in the graph networks are rated on visits per node that will give larger ratings to slow networks and lower to small, fast networks. You can correct a little but that depends on hardware used again.

https://katagotraining.org/

See graph. Correction is possible by checking the box.

[tapio]

On Windows, use Katrain, on Android, Baduk Ai https://play.google.com/store/apps/deta ... r.baduk_ai - nets at https://katagotraining.org
There's a human net as well.