#kata iirc is mostly self-play but gets told “hey you misread this situation study it a lot” manually sometimes

[kaiasky]

alphago master was trained on thousands of human games; we might very well say it just learned to imitate the masters and steal their moves. alphago zero never saw a human game, it never learned to imitate anybody. and yet for how they changed the game of go, it matters not at fucking all. The impact on the game is exactly the same.*

Obviously 2-player turn-based perfect strategy games are a very special case of AI applications. And like, in the art world there's probably a meaningful difference between an AI that can be carefully prompted to generate a derivative style and one that already has weights for "xyz style".

But ultimately, does the source matter or does the impact matter? Because I'm pretty skeptical that ethically sourced training data is that much of an obstacle longterm. and so it feels like the impact of the tool is the thing that matters in any analysis. (I'm not saying anything new here I realize. Smarter people have said this.) It's just that. Training data and how it's used and handled can be important, certainly, but it seems like it gets outsized importance in discussions compared to capabilities. It seems like the Go AIs were a useful demonstration of that.

(*the impact is 0, bc google didn't release them publicly and katago is stronger anyways gotemmmm)