The obvious extension of generating models with large corpi, eventually, is generation of models of models. The LMM, the large model model, is trained on models, such as the content of hugging face. Right? All the fine tuners and abliterators, and every variation of a model can be used as training data for eventual large model models.