I can't even begin to fathom how someone would be able to do this. Wouldn't they have to perform many (100s?) pixel-perfect actions in order to get everything exactly as intended?
If you watch the video you'll see that one of the first things he does is to make the coin counter show Mario's x-position in the level. But up until that point yes, he has to do a lot of pixel-perfect actions. Extremely impressive.
He starts by writing dozens of bytes to check if the written bytes are correct or not (along with other side effects for bootstrapping), and some more to see the exact bytes to be written before applying them. And then finally 300+ actual bytes are written. Not as difficult overall, but initial steps would still require extreme patience.