Say we build a self-improving AI that's twice as smart as a human.

It's going to go off and try to improve itself, right? But why should we expect its improved version to share the goals and desires of the original? The AI still has to solve the control problem.

Then when the second-generation AI tries to build the third-generation AI, it has to solve the control problem again.

Since it's not clear that controlling something smarter than you gets easier as you yourself get smarter, and assuming each iteration has some nonzero chance to fail to pass on the values it intended, you're eventually going to have values drift. Probably several times, or at least until you get iterations that don't feel like self-improving any more.