This seems like a bad analogy. If you could simulate a group of smart human going at 10,000 times normal speed, say copies of Steven Chu or of Terry Tao, I'd expect that they'd be able to figure out how to self-improve pretty quickly. In about 2 months they have had about 5000 years worth of time to think about things. The human brain isn't a great structure for recursive self-improvement (while some aspects are highly modular, other aspects are very much not so) but given enough time one could work on improving that architecture.