Maybe it knows earth is flat after all?
With multimodal abilities, I wonder if it could eventually check a raster of the SVG it created and then make iterative improvements that get a better result than vectorizing a raster from stable diffusion.
I once asked it to create a SVG of Earth and it create a green ellipse on top of a blue circle.