On Tue, Nov 19, 2002 at 12:17:35AM -0600, The Fool wrote:
> I doubt their calcutions for human brain processing power are correct.
Their calculations are in the right ballpark; definitely they aren't
clearly incorrect. Here's a message Bob Chassell posted in 1999 that
gives a good background:
***
Date: Fri, 5 Feb 1999 11:19:21 -0500 (EST)
From: Robert J. Chassell <[EMAIL PROTECTED]>
Subject: Re: petaflops computing (L3)
"J. van Baardwijk" <[EMAIL PROTECTED]> asked:
Petaflops computing?
A computer that makes 10^15 floating point operations per second.
This is roughly the activity of a human brain.
Consequently, if you can program and build a computer that works at
approximately this speed, you may be able to manufacture an
intelligence that works as quickly as a human.
Unfortunately, the same line of reasoning that leads us to think this
may be possible in a generation or two is the line that led a young
1950s science fiction reader to think that interstellar flight might
be possible within a generation or two of then....
At
http://www.hedweb.com/nickb/superintelligence.htm
Nick Bostrom argues:
The human brain contains about 10^11 neurons. Each neuron has
about 5*10^3 synapses, and signals are transmitted along
these synapses at an average frequency of about 10^2 Hz. Each
signal contains, say, 5 bits. This equals 10^17 ops.
Bostrom remarks on the past failures of AI:
The computers in the seventies had a computing power comparable to
that of insects [10^6 ops]. They also achieved approximately
insect-level intelligence.
Calculations by Ralph C. Merkle in
http://www.merkle.com/brainLimits.html
range from from 10^13 operations per second to 10^16 ops for a human brain.
The fastest contemporary computers run at about one teraflop or 10^12
ops. If the lower capacity calculations are correct, we should soon
be entering the `human' range -- we should be be within a factor of 10
or 100 (3 or 7 doublings) of the necessary hardware capacity. If
Bostrom is correct, current computers are 100,000 times slower than
needed for human capacity (17 doublings).
But even supposing that Bostrom is underestimating human capacity by
another factor of ten, presuming that Moore's law holds and depending
on your choice of doubling time, we should be able to construct the
hardware for a human-level intelligence in nor more than 30 to 60
years time. If human capacity is lower, we should be able to
construct the hardware sooner.
Regardless whether you adopt the lower or higher estimates of human
capacity, we are in the same state with regard to artificial
intelligence at a human level as we were in the 1950s with regard to
interstellar flight.
In the 1950s -- I remember doing the calculation in either the late
'50s or early '60s -- the maximum speed of human travel had been
increasing for years and increasing at an *increasing* rate. (I wish
I could remember the figures; I only remember the results.) The
extrapolations were like our extrapolations of Moore's law, but even
more dramatic, on account of the increasing rate of increase.
The law for travel speed had worked great in the past, and it had
worked for a longer past than Moore's law. It even worked for another
8 or 10 years .... humans did go to and return from the Moon,
traveling at speeds up to 11 km/sec.
Then, CRASH!, the law of increasing speed hit a wall and maximum
speeds stopped increasing. The fastest humans are still the lunar
astronauts.
We did not build 20 km/sec nuclear-thermal-engined rockets in the
1970s nor, most sadly, did we achieve superluminal speeds in the late
1980s as my extrapolations predicted.
Similarly, Moore's law may fail. No AI at human levels.
But it may not fail. The human brain is a parallel arrangement of
individually slow components. Even if Moore's law fails for the
speeds of individual components, it may be possible to learn to
manufacture or grow many components and connect them together to make
a parallel arrangement, like a human brain.
Travel speed has just one way of increasing -- you must go faster so
as to reach your destination sooner. Intelligence capacity has at
least two ways of increasing: a single component can go faster or
multiple components can be added together.
This means that even if, as many people think likely, one route fails,
that of direct speed, the other route may succeed, that combining
components in parallel.
This means we can see fewer barriers to human level AI than 50s people
could see preventing high speed travel.
Speculation is not yet crushed by understandings of reality... :)
--
Robert J. Chassell [EMAIL PROTECTED]
Rattlesnake Enterprises http://www.rattlesnake.com
***
--
"Erik Reuter" <[EMAIL PROTECTED]> http://www.erikreuter.net/
_______________________________________________
http://www.mccmedia.com/mailman/listinfo/brin-l
