From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!utgpu!cs.utexas.edu!swrinde!mips!think.com!news!moravec Tue May 12 15:50:00 EDT 1992
Article 5514 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!utgpu!cs.utexas.edu!swrinde!mips!think.com!news!moravec
>From: moravec@turing.think.com (Hans Moravec)
Newsgroups: comp.ai.philosophy
Subject: Re: AI failures
Message-ID: <MORAVEC.92May9113029@turing.think.com>
Date: 9 May 92 16:30:29 GMT
References: <1992May7.152447.7930@waikato.ac.nz> <727@ckgp.UUCP>
	<uc2m8INNn5d@early-bird.think.com> <1992May9.013815.24244@nas.nasa.gov>
Organization: Thinking Machines Corporation, Cambridge MA, USA
Lines: 38
NNTP-Posting-Host: turing.think.com
In-reply-to: eugene@amelia.nas.nasa.gov's message of Sat, 9 May 92 01:38:15 GMT


eugene@amelia.nas.nasa.gov (Eugene N. Miya) writes:

> Knowing Amdahl's other law, how long do you expect to take to copy?  Minutes
> or a human life time.

Extrapolating from the ratio of neural tissue in the retina to the amount
of computing required to produce a similar end result with computer image
processing programs, I've concluded in past that a whole human mind should
be implementable in real time with a processing power of very roughly 10^13
(ten trillion) operations/second.  I've also guessed that with reasonably
efficient encoding (including some data compression) the state of a human
brain should be encodable in (even more roughly) 10^14 bits.  If that ratio
is maintained for later AIs, then it should be possible to transfer an AI
through the main data paths in a few seconds: Analogous to a ten megabyte
(10^8 bit) file on an existing 10 MIPS (10^7 operation/second)
workstation.   You just have way too much data for you processor speeds,
Gene!  The future AIs would be similarly overwhelmed if they tried to store
all their super-hi-res sensory data, instead of reducing it to its
important parts immediately.

By cheap reproduction, I had in mind keeping them as processes in main
memory.  With continuing persistent 1000 fold/ 20 year growth in computer power
and memory, within 100 years it should be possible to have a BILLION
human-scale AIs running simultaneously at human rates in a (vintage 2090)
workstation-class machine!  (Presumably it would be better to have fewer,
faster, superhuman-scale AIs)

			-- Hans Moravec








P.S. Thanks for the tour of your robot lab a few years ago.


