This maybe also dovetails with deep mind (?) or openai (?) article yesterday talking about adapting and updating general models with limited data for significantly different tasks.