And for how hard the problem is, I'd emphasize how likely we are to be wrong about determining or extrapolating exactly what we want, in each sense listed above.
An engaging way to open a talk is to challenge the audience, then give them this info as first steps. If I can suggest an example, I'd say something like "Asimov's 3 Laws looked innocuous and even in our best interest, but they were terribly flawed. So can we find a real solution? One that will hold for powerful AGI? This is an open problem."
If you don't run out of time, connect it back to self-modifying code (if only to ground the problem for a group of programmers). We need software that will self-modify to satisfy dynamic criteria that may exist, but are obscured to us. And we may only get one shot.
That will make someone ask about take-off and cooperating agents, and you can use the Q&A time to mention FOOMing, negentropy etc.
Next week I'm going to be doing a 10-15 minute presentation on Friendly AI to a local group of programmers. They're already familiar with concepts such as the singularity. My basic plan is to cover what FAI is, why it's important, and why it's a hard problem, based on the material on this site.
Does anyone have any specific suggestions of things that should be included, questions that I might need to answer, etc?