“An older version of the Llama model was reportedly used to generate complex puzzles, which were then used as training data for the next-generation model to improve its problem-solving efficiency.”