Will transformer derived architectures still be state of the art for language modeling in December 2025?