Re: One or more FAIs??

From: Dani Eder (danielravennest@yahoo.com)
Date: Wed Jun 02 2004 - 14:31:25 MDT


> This logic doesn't work at the most basic level.
It
> seems to me that the
> Singularity Institute will *not* be the first to
> create an AGI

Looking at the top500.org supercomputer list, there
does not seem to be any obvious dedicated AI research
machines, but I would make a guess that 2% of the
available computation power on the list might be used
for AI work, mostly as timeshared capacity on
university machines.

The aggregate computation power for AI might be then
10
Tflops, and the largest single machine might be
1 Tflop.

The aggregate power of the top500 list has been
growing by 80% per year over the past 10 years.
Meanwhile the performance/cost ratio for cheap
commodity clusters has been
increasing at 75% per year over the past 2.5 years.
Adjusting for inflation, the total cost for the
top500 systems has remained near constant.

A dedicated AI research machine cranking 1 Tflop
would cost $150,000 today if built from cheap
commodity components. A comparable future machine
to equal the most powerful machine being used for
AI research elsewhere should also cost $150K in
future years.

Therefore, whatever processor power is needed to
implement AGI, if the SIAI can be expected to have
>$150K to apply to the hardware, then they might
be first. If not, then it is more likely that
someone else will be first.

If AI shows prospects of being useful, then the
present low level of research could shift to a
much higher level. If as much money is sunk into
AI as, for example physics or weather forecasting,
then the probable 'largest machine' that SIAI
would be competing with gets much larger. 7 of
the top 10 machines are being used in those two
fields, and they average 10 Tflops. In that case
you would need to spend a couple of million $ to
be competitive.

Daniel

        
                
__________________________________
Do you Yahoo!?
Friends. Fun. Try the all-new Yahoo! Messenger.
http://messenger.yahoo.com/



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:47 MDT