For every token, model activations are computed once when the token is encountered and then never explicitly revised → “only [seems like it] goes in one direction”
For every token, model activations are computed once when the token is encountered and then never explicitly revised → “only [seems like it] goes in one direction”