From: Bob Seidensticker (firstname.lastname@example.org)
Date: Tue May 02 2006 - 13:51:56 MDT
Dani: 3x10^6 GFLOPS is a human equivalent? That may be the equivalent in
computing power, but doesn't that ignore the software side? Suppose we had
such a supercomputer right now. Or we harnessed computers over the Internet
to collect that much power. Could this thing pass the Turing Test?
Or, said another way, if my laptop has the power of 10^(-6) human
equivalents (it can't quite do 3 GFLOPS, but let's imagine), then a Turing
Test could be done in one-millionth real time. That is, if you could
respond to a question in 5 seconds, we should have a laptop today give a
roughly equivalent answer in about 2 months. Does that software exist?
My concern is focusing on the hardware side (admittedly, Moore's Law is
making amazing advances) but ignoring the software side, which has been
notoriously hard to make.
From: email@example.com [mailto:firstname.lastname@example.org] On Behalf Of Dani Eder
Sent: Monday, May 01, 2006 5:42 PM
Subject: Re: Anti-singularity spam.
> When it becomes possible, for the first time, to build something that
> generates new knowledge in such a way that it is not limited by the
> speed of the human brain, then you're just not in Kansas any more.
If that's the point, I'm well aware of it, and if you read my archived posts
on "hardware progress" I've even tracked when I expect it to occur.
To sum up an argument I've made in the past, a machine with one
human-equivalent in potential should take on the order of 20 years to train,
because that's how long we do. Given a 2 year or so doubling time in
computer power, two years later a 2x human potential machine would take 10
years to train, etc. so the earliest expected date of more than trained
human ability is 8.5 years after the first human-equivalent is reached (6
years for technology improvements + 2.5 years to train).
After that a rapid run-up to the singularity can occur at any time. One
superintelligent machine may not do that much by iteslf, but if one can be
built, so can many.
Businesses, being economically rational, will employ machines when they are
cheaper than people. If a machine can do the job of a design engineer, it
is worth on the order of $3 million. That's because a machine can work 5x
the average hours of a human, and an engineer with fringe benefits and
overhead runs ~$120K/yr x 5 yr econonmic life of the machine.
I use 3x10^6 Gflops as human-equivalent, so it works out that when computers
cost <$1/Gflop, you would expect them to be deployed in large numbers
because they would be cheaper than people, and after that the pace of
technology development should accelerate.
The last time I projected a date for $1/Gflop, it was about 2016.
Dani (not Danni by the way) Eder
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT