Re: How hard a Singularity?

From: Eliezer S. Yudkowsky (
Date: Sun Jun 23 2002 - 08:43:35 MDT

Ben Goertzel wrote:
> So far, nothing that you have said has specifically addressed this issue.
> I.e., nothing but your own unsupported intuition has been offered in
> favor of your posited "one month upper bound" for the transition from
> human-level AI to vastly superhuman AI.
> Nor have I offered any real evidence in favor of my own intuition that it
> will be a bit slower than this!
> What I am questioning is not your confidence that the feedback loop
> itself will exist, but your confidence in your quantitative estimate of
> the speed with which the feedback loop will lead to intelligence
> increase.

Look at this way: Given what I expect the overall shape of the curve to
look like, if you specify that it takes one year to go from human-level AI
to substantially transhuman AI, then it probably took you between a hundred
and a thousand years to get to human-level AI. If you're wondering where
the ability to name any specific time-period comes from, that's where - the
relative speed of the curve at the humanity-point should be going very fast,
so if you plug in a Fudge Factor large enough to slow down that point to a
year, you end up assuming that AI development took a century or more. Even
so I'm not sure you can just plug in a Fudge Factor this way - the
subjective rate of the developers is going to impose some limits on how slow
the AI can run and still be developed.

Seed AI will be a pattern of breakthroughs and bottlenecks. As the AI
passes the human-equivalence point on its way between infrahumanity and
transhumanity, I expect it to be squarely in the middle of one of the
largest breakthroughs anywhere on the curve. If this mother of all
breakthroughs is so slow as to take a year, then the curve up to that point,
in which you were crossing the long hard road all the way up to human
equivalence *without* the assistance of a mature seed AI, must have taken at
least a century or a millennium.

And if it takes that long, Moore's Law will make it possible to brute-force
it first, meaning that the AI is running on far more processing power than
it needs, meaning that when self-improvement takes off there will be plenty
of processing power around for immediate transhumanity. Still no Slow

Eliezer S. Yudkowsky                
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT