What? Minecraft "voxels" are huge. And then the cubes on the creatures are a different size. How is it not just using meshes to start with?
Because there are order of magnitude differences between rendering each voxel in a dense landscape compared to just the borders. Consider a cube of 10x10x10 voxels. They each have 6 sides, so that's 6000 quads to render. If we only render the sides that can possibly be visible, that's 6 sides on the cube times 10 * 10 voxels on each side, totalling 600 quads. I'm sure the creatures are not treated like this; they're…
AFAIK culling like that is standard in any mesh-based 3D engine, and most minecraft landscapes aren't uniform enough to benefit from the adjacent-face merging described. Standard 3D engines handle landscapes with way more polys than I can imagine fitting into a minecraft screen. I don't see what the big deal is here.