“Large language models implicitly develop a world model that improves with scale, and it is not necessary to explicitly engineer a separate world model component.”