If I gave you a source code of a Friendly AI, all you'd have to do would be to run the code.
If I told you to do whatever human CEV is, you'd have to find and interrogate some humans.
The difference is that by analysing the code of the Friendly AI you could probably learn some facts about humans, while by learning about prime numbers you don't learn about the pebblesorters. But that's a consequence of humans caring about humans, and pebblesorters not caring about pebblesorters. Our values are more complex than prime numbers and include caring about ourselves... which is probably likely to happen to a species created by evolution.
There seems to be a widespread impression that the metaethics sequence was not very successful as an explanation of Eliezer Yudkowsky's views. It even says so on the wiki. And frankly, I'm puzzled by this... hence the "apparently" in this post's title. When I read the metaethics sequence, it seemed to make perfect sense to me. I can think of a couple things that may have made me different from the average OB/LW reader in this regard: