Examines whether Transformers are necessary when LLMs write model weights without gradient training. When allowed to design the architecture, the tested LLMs compile facts into sparse tables, reaching up to 54.4% accuracy on ARC.