it takes less space to write a 3 minutes HD animated video and music clip than than it takes to store a 32×32 pixel icon
The video very clearly demonstrated that using only a 9x9x24bit image you can encode a program that, when executed, will produce many MBs of data. It's the same concept as x (k)Bytes of data producing an infinitely scrolling procedurally generated terrain.
See, to me this means "reading the colors from the printout". It's ridiculous that anyone could get the least significant bits of the color right by just eyeballing it, and I don't think the author ever claimed he was doing that.
If you look at the video carefully, you'll see that the color picker displays numeric RGB values and that it has been sped up. Do you think the code would work if he didn't get the values 100% right?
Off to the right side of the screen (not visible) is probably another image with the colours in that he's picking from.
Maybe it's fake but it's theoretically possible. I don't have a windows pc to try on. He's provided the image though so you could try for yourself.
Am I wrong?