Post #3553513

I've just invented the term "dominant priors" for when probabilities in an LLM "shout louder" than those calculated from the context, so the model "wins" over the information you give it e.g., when it uses out-of-date API documentation even after you gave it new docs Main cause: attention dilution