There is some flexible encoding of terms to images there, too, that would take up some space. Though I’m not really sure you could decompress the weights in so

There is some flexible encoding of terms to images there, too, that would take up some space. Though I’m not really sure you could decompress the weights in some way to regenerate the 400 million images and metadata.