“The structure of post-training recipes for large language models changed more in the last year than in the preceding three years.”