I have recently been playing against the b18c384nbt-humanv0 models, with different strengths (say 20k-10k). I was wondering if some-one has tried to verify their nominal rank.
I am asking because I started playing GO quite recently (few months) and by my last games I evaluate my rank to be something aroung 20k-18k, but it seems that I can routinely win the human-like model even at quite stronger levels (as low as 10k).
I guarantee you I will not win against a 10k player
I hope I am getting better than my current rank shows (and I don’t play many ranked games on OGS), for sure my game-losing mistakes still look 20kyus
The point is that even then the AI does not look 10kyu to me
Supervised trained human-like models output policy are statistical distributions of the training data, and since the peak dateset centered around 5k, there is a lack of data as the ranks goes down, and they had to be “extrapolated” when there are a lack of datapoints. And since as the ranks in DDK, tend to have wildly differ opening and fuseki patterns, the statistical distribution of opening moves are almost random like or very slow, since the extrapolation for the follow-up to the opening would be “easier” to “emulate”. That is the supervised trained models tend to have very weak opening, and early games, but regression to norm in the mid-to-end games (very few can be extrapolated for mid-games or end-games). Hence a very uneven quality of moves below SDK levels. And they tend to be disjointed in their direction of play but somewhat decent in main line joseki (since the most common patterns they will share).
