Way in over my head here, but if the project is done in a series of steps with a series of boxed AIs helping to work on the next step, wouldn't it be better get something that is at the very least(*)…

• always trying to be friendly.
• always trying to be completely honest and non-evasive (to the programmers).
• pretty damn stable.
• pretty damn good at avoid killer-Gandhi pills, lost purposes, etc..
• will alert & stop itself if it thinks it's drifting on any of the above.

..at a much earlier step than either…

• being smart enough to outwit the programmers in dangerous ways.
• being capable & allowed to do potentially foomy self-improvements.
• doing any work on the next generation itself, as oppose to offering advice.
• being able to talk to anyone who hasn't been vetted.

(*) - if not perfectly/certifiably, if that's feasible before doing stuff on the second list.