I think trained knowledge is less and less important - as these multi-modal models have the ability to search the web and have much larger context windows.