I agree, the scale problems seem to outweigh the scale economies at N=2. At some point you get positive network effects where the older ones contribute to the care of the younger ones, but so far we mainly observe a lot of resource contention with frequent livelock exceptions requiring intervention by the parent processes.
At N=4, I have observed substantial reduction in required average per unit parental engagement. However, we found that parental unit genetics increased the probability of encountering a different instruction set (autism), which involves recompiling a large fraction of the standard library for one unit.
