Comment by sva_

Great article. Perhaps some part of this magic number simply factors in the amount of compute necessary to run the image through the CNN (proportional to compute use per token in the LM).