An IQ70 AI is little more able to self-improve than a IQ70 human is able to improve an AI

This is not obviously true. We're a lot less well optimized for improving code than some conceivable AIs can be: a seed AI with relatively modest general intelligence but very good self-modification heuristics might still end up knocking our socks off.

That said, there's a much larger design space where this isn't the case.