Honestly, this course is incredibly enjoyable and gives me chills from the beauty and creativity of AlphaGo. It keeps evolving and getting better and better.
However, I want to share some important information: searching online for “which AI is the best” is not accurate. I got the same results myself, but after more than 4 hours of deep research using AI, here’s the real situation after AlphaGo.
DeepMind published the research paper, then the first AI that appeared was FineArt. At that time there was no KataGo. It was extremely strong — on the same level as AlphaGo — and surpassed it over time.
Later another AI called Golaxy came out (also before KataGo). It was slightly weaker than FineArt. Both were closed-source, so you couldn’t download them to your own device, and they were very close to each other in strength. There was also Leela Zero, but it wasn’t as strong as them (though still powerful).
Then FineArt decided to go commercial. Its owner Tencent merged it into Fox and stopped participating in AI tournaments. Golaxy did the same.
After that, KataGo appeared as an open-source AI — anyone can see the code and download it. Leela Zero development stopped and the community focused on KataGo. It received a lot of support and contributions. In the beginning it was weaker, but now it has become extraordinary. I ran experiments and it can beat Golaxy by a huge margin, and also FineArt (on their website) by a huge margin.
But you need to know that KataGo’s strength changes a lot depending on the computing power. On a strong phone it is already stronger than a human, but with powerful graphics cards it completely dominates. That’s why it needs a decent computer — not necessarily extreme, something like an RTX 3090 or RTX 4090 is enough.
The reason is Playouts/Visits (the number of calculations performed). The higher the number, the more accurate the result. A phone is not enough to show its true strength, so it’s better to rent a cloud service. There are many options, but one of the best in terms of price and mobile compatibility is the Chinese platform Zhizi.
Another important point: network size plays a huge role. The larger the network, the stronger it is and the fewer Playouts it needs to reach the same quality. Weaker networks need more Playouts, and each Playout on a larger network is more powerful. The difference is massive.
If you download it yourself, always check the network size and your computer’s power. Some platforms don’t use the strongest networks, so the strength you get is much weaker than current KataGo.
Existing networks from weakest to strongest: 10B → 18B → 28B → 40B → 60B.
Then a new and much stronger technique appeared called NBT (bottleneck). It uses the hardware much more efficiently and is a real qualitative leap. Any version with NBT is significantly stronger. Current ones are 18NBT and 28NBT — they can beat a 60B network without NBT.
There is also a giant and very powerful network trained on the Zhizi platform: 40B NBT. It was released recently, is very heavy and extremely strong, and makes a big difference.
The number of Playouts needed for an accurate evaluation depends on the network. Larger networks need fewer. For 40B NBT you roughly need 12,000 to 50,000 Playouts — lower numbers are not accurate enough.
There is also an option called PDA (Playouts Double Advantage) with values from 0 to 3.
–3 makes it more conservative (safer play).
+3 makes it play more complexly, which is useful in handicap games — it complicates the position so the opponent is more likely to make mistakes. It may play moves that lose 0.3–0.8 points but create a very complicated situation.
I know this is long, but I explained an important topic because it really makes a big difference. Vadim mentioned that FineArt is the strongest and Golaxy is stronger, but in reality KataGo is much stronger.
Honestly, this course is incredibly enjoyable and gives me chills from the beauty and creativity of AlphaGo. It keeps evolving and getting better and better.
However, I want to share some important information: searching online for “which AI is the best” is not accurate. I got the same results myself, but after more than 4 hours of deep research using AI, here’s the real situation after AlphaGo.
DeepMind published the research paper, then the first AI that appeared was FineArt. At that time there was no KataGo. It was extremely strong — on the same level as AlphaGo — and surpassed it over time.
Later another AI called Golaxy came out (also before KataGo). It was slightly weaker than FineArt. Both were closed-source, so you couldn’t download them to your own device, and they were very close to each other in strength. There was also Leela Zero, but it wasn’t as strong as them (though still powerful).
Then FineArt decided to go commercial. Its owner Tencent merged it into Fox and stopped participating in AI tournaments. Golaxy did the same.
After that, KataGo appeared as an open-source AI — anyone can see the code and download it. Leela Zero development stopped and the community focused on KataGo. It received a lot of support and contributions. In the beginning it was weaker, but now it has become extraordinary. I ran experiments and it can beat Golaxy by a huge margin, and also FineArt (on their website) by a huge margin.
But you need to know that KataGo’s strength changes a lot depending on the computing power. On a strong phone it is already stronger than a human, but with powerful graphics cards it completely dominates. That’s why it needs a decent computer — not necessarily extreme, something like an RTX 3090 or RTX 4090 is enough.
The reason is Playouts/Visits (the number of calculations performed). The higher the number, the more accurate the result. A phone is not enough to show its true strength, so it’s better to rent a cloud service. There are many options, but one of the best in terms of price and mobile compatibility is the Chinese platform Zhizi.
Another important point: network size plays a huge role. The larger the network, the stronger it is and the fewer Playouts it needs to reach the same quality. Weaker networks need more Playouts, and each Playout on a larger network is more powerful. The difference is massive.
If you download it yourself, always check the network size and your computer’s power. Some platforms don’t use the strongest networks, so the strength you get is much weaker than current KataGo.
Existing networks from weakest to strongest: 10B → 18B → 28B → 40B → 60B.
Then a new and much stronger technique appeared called NBT (bottleneck). It uses the hardware much more efficiently and is a real qualitative leap. Any version with NBT is significantly stronger. Current ones are 18NBT and 28NBT — they can beat a 60B network without NBT.
There is also a giant and very powerful network trained on the Zhizi platform: 40B NBT. It was released recently, is very heavy and extremely strong, and makes a big difference.
The number of Playouts needed for an accurate evaluation depends on the network. Larger networks need fewer. For 40B NBT you roughly need 12,000 to 50,000 Playouts — lower numbers are not accurate enough.
There is also an option called PDA (Playouts Double Advantage) with values from 0 to 3.
–3 makes it more conservative (safer play).
+3 makes it play more complexly, which is useful in handicap games — it complicates the position so the opponent is more likely to make mistakes. It may play moves that lose 0.3–0.8 points but create a very complicated situation.
I know this is long, but I explained an important topic because it really makes a big difference. Vadim mentioned that FineArt is the strongest and Golaxy is stronger, but in reality KataGo is much stronger.