http://intelligence.org/research/
The scariest part of the problem to me is this: http://yudkowsky.net/singularity/aibox
If an AI can't be contained, then there has to be some failsafe that prevents any self-modifications which could turn it evil OR remove the failsafe. Or you have to believe that you can neutralize an evil AI before it destroys humanity.
That being said, yes, what's the point of having an AGI if you're not letting it "out"? It will probably escape anyway, so containment is not a solution.
By the way, there is a lot of precedent in human history about the effectiveness of containing "dangerous" ideas by imprisoning the people who developed them. It generally doesn't work out.
If you're developing AGI, you've pretty much already resigned to the idea that the world will change. The destruction of the human race is only one extreme outcome. Especially if the takeoff isn't very hard, there are a lot of scenarios where a net-beneficial AI could be developed. And all the engineering experience we have does indeed point to a very soft takeoff.
Ultimately, I think the goal is NOT to generate a slave race with super powers. Yet that's what a lot of people are talking about. It's not only morally wrong but also self-evidently stupid. A lot of influential futurists are fans of that idea though, including Kurzweil.
Useful AGI scenarios range from mutually beneficial coexistence to outright merging with the AGIs, and I think that's a goal well worth pursuing. Another strategy would be to keep augmenting ourselves and to continue on a transhumanization path until we are the AGI. You could indeed make an argument that's what's already happening.
Probably it went something like this:
Eliezer: I give you 100$ if you write on the list that you let me out
Nathan: Done.
Which makes me think they use blackmail or something similar in the transcript.