I think you are making some unsupported assumptions:
-- you asusme that AIs are separate entities.
-- you assume that augmented humans are separate entities.
-- you assume a polarization between humans and AIs
I quit using the term AI some time ago, during an earlier round
of this discussion in 1996. I prefer the term "superintelligence",
or SI. IMO an SI will very likely start as an augmented human,
a collaboration between one human and a computer. This SI can
rapidly augment the computer part of itself, so there is no
reason to assume that a second SI will come into existence prior
to the first SI taking over the world, or that a de novo AI will
ever come into existence. The SI may choose to decouple from its
human component, but such a decision, and indeed any decision made
by the SI, is beyond analysis by humans.
There are variants on this scenario, such as a multi-human initial
collaboration, but IMO a single human suffices. Further, there
is no real necessity for new I/O interfaces: mouse and monitor suffice.
Since the human can provide creative direction and strategic pattern
matching, the computer need not start at "human level". This means
that the SI may be achievable at today's level of hardware, and that
the only reason the SI does not already exist is that the software
has not yet all been invented. If this is the case, the SI will
come into existence suddenly, as a phase change rather than as a
mere rapid increase in development. Please see:
http://www.shirenet.com/~dgc/singularity/singularity.htm