非常好的app
In fact, the model is very stupid
Response from developer
Thank you for this report — you were right, and it was a real bug on our side, not just weak play. Below Lv.10 the app deliberately picks among KataGo's candidate moves, but a clamping error piled most of the probability onto the *worst* candidate (about 80% at the default Lv.5), and the search was so shallow that those candidates had only 1-2 visits — moves the engine had never really evaluated. So the AI genuinely did play near-random moves. We reproduced it with the same b10 model and fixed it: the AI now only plays moves the engine actually searched, each level caps how much win rate it may give up (Lv.10 always plays the best move), search was deepened, and the default is now Lv.7. This ships in the next update. 感谢反馈！这确实是我们的程序缺陷：低于 Lv.10 时选点逻辑把大部分概率压在了最差的候选着法上，而且那些着法几乎没有被搜索过，所以 AI 看起来像在乱下。我们已复现并修复，下个版本会更新。
