The same issues are still happening in frontier models. Especially in long contexts or in the edges of the models training data.
reply