The paper introduces infinite-parameter large language models that can generate and adapt their weights directly from live data streams. This approach aims to enable models to scale beyond fixed parameter counts while continuously updating knowledge.