FORTH and associative/applicative languages may be more suitable for transformer architectures than traditional top-down problem-solving methods. Generating outputs before their constituent parts could enhance the effectiveness of large language models.