Comment by nukanse

Language models are efficient because they are trained with millions of examples that don't exist for newer architectures, and they are still mediocre to bad on stuff where they hadn't been trained