Comment by p1esk

The way I think about QKV projections: Q defines sensitivity of token i features when computing similarity of this token to all other tokens. K defines visibility of token j features when it’s selected by all other tokens. V defines what features are important when doing weighted sum of all tokens.