> There does seem to be a stumbling block if the AI needs the net's
> resources to become superhuman, but it can't exploit them until it
> achieves superhuman competence sufficient to redesign itself to run on
> such a constrained system as the net. But Eliezer points out that it
> does not need superhuman competence in every arena, only in software
> architecture design, and perhaps an idiot savant super-coder AI can make
> the leap.
Um, like I keep saying, my original statement was made in the context of
"Will there be a Slow Singularity?" And I said, "If it works for a Slow
Singularity, it works for superintelligence." I.e., threshold for a Slow
Singularity is over-and-above the threshold for a hard takeoff.
If the current AI needs the net's resources to become superhuman, but
Vastmind isn't a household name, latency is too high, and the AI is too
dumb to use the 'Net much less eat it... then you don't get a Singularity
until you can buy a sufficiently large supercomputer, and that's that.
I deem this scenario to be unlikely because I think that the major
constraint is likely to be software, not hardware. I.e., given a smooth
rate of progress, your AI is "good enough to do a takeoff given the
Internet" a month or two before the AI is "good enough to do a takeoff on
the fastest supercomputer around", and then it's just another month or two
before the AI can do a takeoff with whatever hardware you developed it
on. A few months could easily be ten million lives at the current death
rate, so yes, Vastmind is worth it, but it probably won't be the
Singularity itself that's at stake.
-- -- -- -- --
Eliezer S. Yudkowsky http://singinst.org/
Research Fellow, Singularity Institute for Artificial Intelligence
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 10:00:08 MDT