I wouldn’t call that a madness. As I told you, Kugutsu was (and still is a bit) based on GNUGo without using MonteCarlo Tree Search, without using Neural Network. It’s been over 3 years since I discovered GNUGo and started messing with it as an example of a well-coded C program. I learned the rules of Go as a child but never played this “pastime for old people (a common image in Japan)” seriously. As I couldn’t beat GNUGo in games, I got interested in playing Go myself. So I have started coding Kugutsu earlier than I started playing Go seriously myself.
Once my play strength came close to Kugutsu’s, I wanted to see how strong he is relative to people, and found OGS to test it. I started playing on OGS, learned ins and outs on the server, then registered Kugutsu as a bot.
As fiddling with the program intending to improve it was/is my hobby, there was no deadline to meet, so I left it as it was with very few modifications while I trained myself in Go with the way Kugutsu played as my guide. When I finally became stronger than Kugutsu in a year or two, I stopped the training and studying, and made more programming improvements to the level where I couldn’t beat him any longer. Then I re-started my own Go study.
So this seesaw improvements to my own Go playing ability and the program became a repeated cycle up to early this year when I hit a ceiling in no longer being able to make the code stronger, or most of the ‘improvements’ as I thought actually made the bot weaker (without using MCTS and/or DCNN).
In many many ways, Kugutsu has taught me how to play Go better, and in a way he is a representation of my Go career. I’d like to think Kugutsu plays more human-like than GNUGo does, and I am happy that there are others who enjoy playing him.