Is this bringing us closer to combining image and language understanding within one model?