Dan Fabulich wrote:
>... suppose that in order to get a linear increase in
>intelligence, you had to solve a problem requiring an exponential
>number of computations. ... progress would get slower and slower.
>... For all we know, it'll turn out that exponential increases in
>intelligence requires solving a problem requiring a linearly
>increasing number of steps, and just as soon as we get the tools to
>start messing around, our intelligence will go through the roof.
We actually do have data on this! Human IQ *has* been increasing
dramatically over the last century. It's called the Flynn effect.
The graph there looks linear, but if it were exponential the
growth rate would be ~3% per year.
At the same time measures of humanity's effective compute cycles have
increased exponentially, such as population (~1.5%) or world product (~4%).
If we posit that IQ per person is a power law in terms of humanity's
effective compute cycles so far, and that IQ grows exponentially we'd get IQ
being population to the power of ~1.4, or world product to the power of ~.5.
If artificial computers come to dominate the compute cycles so far, and if
those compute cycles double every two years, then these power laws predict
IQ would double every 1.4 or 4 years. This would be a dramatic speedup, but
still far less than the super-minds overnight scenario. And if IQ growth
has been linear, future IQ growth would also be expected to be linear and
hence more limited, even if faster by a similar magnitude.
Notice that under this conception it makes little sense to imagine a
computer in a small lab suddenly breaking through to super-intelligence.
Intelligence increases would be the result of breakthroughs all across the
globe, and the compute cycles in any one lab would only be a very small
fraction of that total.
Robin Hanson firstname.lastname@example.org http://hanson.gmu.edu
Asst. Prof. Economics, George Mason University
MSN 1D3, Carow Hall, Fairfax VA 22030-4444
703-993-2326 FAX: 703-993-2323
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 09:50:14 MDT