“The recently released MAI model was initially trained without any synthetic data to avoid potential distillation from other models.”