Message-ID: <32D96F7C.357E@tale.com>
Date: Sun, 12 Jan 1997 17:10:52 -0600
From: Ken Jenks <MindsEye@tale.com>
Reply-To: MindsEye@tale.com
Organization: Mind's Eye Fiction: http://tale.com/
X-Mailer: Mozilla 3.0 (Win95; I)
MIME-Version: 1.0
Newsgroups: misc.writing,rec.arts.sf.written,comp.ai,sci.space.shuttle,rec.arts.startrek.current,rec.arts.sf.movies
Subject: Happy Birthday, HAL 9000!
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
NNTP-Posting-Host: max7-11.ghg.net
Lines: 182
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!swrinde!howland.erols.net!newsfeed.internetmci.com!news.ghgcorp.com!max7-11.ghg.net
Xref: glinda.oz.cs.cmu.edu comp.ai:43357 sci.space.shuttle:53138

Today is the fictional birthdate for the HAL 9000 computer from the
landmark science fiction tale "2001: A Space Odyssey" by Arthur C.
Clarke. 

   "I am a HAL Nine Thousand computer Production Number 3. 
   I became operational at the Hal Plant in Urbana, Illinois, 
   on January 12, 1997."  

Urbana is the home of my  alma mater , the University of Illinois, and
the home of the famed ILLIAC I, one of the first digital computers built
and owned by an educational institution. (EDSAC I was the first, built
by the University of Cambridge in May 1949.) Urbana and its sister city
Champaign have a proud tradition of high-tech computer invention,
innovation and discovery. Urbana is a fitting place for Hal's
construction. 

Hal served as the "brain and nervous system" of the 400-foot-long
nuclear-powered spaceship  Discovery  that carried astronauts on a
thought-provoking voyage to the planet Saturn (changed to Jupiter in the
movie version). I highly recommend both the movie and the book, but as
usual, you should see the movie first.  

Hal could play chess, run a complicated spacecraft, perform predictive
diagnostics on the spacecraft systems, display emotions ("That's a
relief. You know that I have the greatest possible enthusiasm for this
mission."), plot, scheme, lie, murder, understand speech and even do
some lip reading. 

In the story, Hal's fatal flaw was an inability to reconcile the lies
with the truth. The conflicting goals of keeping the true purpose of the
mission secret from the crew and supporting the crew in that mission
drove Hal to abherant behavior, eventually leading to multiple murders.
This represents Frankenstein's monster reawakened, but without the
creator/creation aspect, and it sheds literary light on the devastating
effects of organized lying on a naive (but inorganic) mind. The literary
metaphor invites the comparison with humans and how they deal with
lying. 

But I'm not just interested in the literary aspects of Hal -- I'm
interested in his technology. Today, on the birthday of the imaginary
HAL 9000, how close are we to building computers that run spacecraft
autonomously, play chess, hold an intelligent conversation, propose
logical (albeit psychopathic) arguments, plan, scheme, lie and even
lip-read? Pretty damned far. 

(We're pretty damned far from having a nuclear-powered manned
spacecraft, too, but I'll leave that discussion for another day.) 

Let's take a look at Hal's capabilities and see where 1997 technology is
with respect to Arthur C. Clark's ideas in 1968. 

You should understand that Clarke is a technologist. He thinks broadly
and clearly about technology and its effects on people and the future.
Many of his science fiction stories deal with these effects. Clarke was
a pioneer of radar, radio and satellite communications. Today, he lives
in beautiful, war-torn Sri Lanka and communicates with the rest of the
world through the global communication system he helped invent. 

I'll explore the current state-of-the art in artificial intelligence
(AI) and compare it with Hal's impressive capabilities. 

Today, the term "artificial intelligence" is often avoided by people who
practice the craft. They prefer instead to talk about their particular
branches of the subject, such as vision systems, natural language
processing, inference engines, expert systems and machine learning. This
aversion to the term AI has come about because AI was over-hyped, in
part by authors like Clarke. People came to expect too much, too soon
from AI, and the technologies have not been able to live up to user
expectations. This hasn't been from lack of progress in computers and AI
-- it is simply inflated user expectations brought on by creative
wishful thinking and anthropomorphizing computers. 

Computers don't see, hear or think like people do. Putting a camera on a
computer doesn't make it see any more than a slide projector can see the
slides it shows. Although images come into the computer through a
camera, scanner or other device, the ability for computers to interpret
images and convert those images to concepts is very limited. Today, for
example, a powerful computer with custom-built stereo cameras and
specialty video processing hardware can look at a human walking slowly
across an uncluttered laboratory and usually figure out where the
person's head and arms are located. The special-purpose software
required to do this can't recognize chairs or tables. It can't
differentiate between a cardboard model of a human and an actual human.
It can't even tell the front of human from the back. This is the state
of the art. 

Vision systems are nowhere near ready to "understand" the subtle motions
of lips, mouth, eyes and body involved in "lip reading"  but computers
can use visual lip reading cues to improve (but not replace) speech
recognition. Today, computers with special sensors can barely tell where
your eyes are looking, much less the subtle difference between, say,
"Hal" and "hail" or even worse, "hail" and "hale." Research into lip
reading is being conducted by the Visual Perception Laboratory (VPL)
within Nippon Telegraph & Telephone (NTT) and the ICSI Speech Research
Lab in the EECS Department at Berkeley.  

Similarly, computers can't hear very well. (Actually, a computer with a
microphone can hear just fine; they just can't  listen  the way you and
I do.) Modern speech recognition software can "understand" some spoken
commands and most words and translate those commands and words into
actions and text, but the error rate is high, noise rejection is poor
and the speaker must modify his speech patterns to compensate for the
software's basic inability to understand conversational speech. For
example, you must pause between each word and enunciate each syllable
clearly, and the computer must be able to understand whatever accent you
have. This is a far cry from a computer understanding astronaut Dave
Bowman's impassioned plea, "Open the pod bay doors, Hal." (Movie
version, not in the book.) 

In 1997, computers don't plan very well. For example, it takes
specialized path planning software for a robot to navigate across a
cluttered floor. That software must have information about obstacles,
clearances, turn radii and other geometric constraints, and it must
always have human supervision. Today's planning software can only
operate in a specialized domain. That same software would not be able to
plan the path of a scuba diver through a wreck or the path of a
spacecraft through a cosmic debris field to Saturn.  

The computers flown on spacecraft have even more limited autonomy and
ability to conduct real-time planning. This is partly because computers
flown in space must be highly reliable and radiation hardened, so they
are often well behind the state of the art on earth. It is also because
NASA mission planners do not trust on-board hardware and software, even
after extensive engineering and testing. (And with good reason - many
manned and unmanned spacecraft have failed or had significant impacts to
the mission due to malfunctioning computers. Gemini-VIII, Apollo 11,
Mars 96 and Ariane 5 spring to mind.) At NASA, trust in AI software is
particularly low because AI has been often unreliable even in laboratory
conditions and never really proven in flight.  

Spacecraft mission planning and replanning are very labor-intensive
activities requiring human direction, intervention and supervision.
Although NASA has some very interesting, useful tools to assist human
mission planners, there is no software that can plan and execute a
complex mission without human intervention. 

Contrary to Clarke's prediction, computers have no trouble lying, but as
he foresaw, they have no conscience about committing murder. Because
epistemology is not well advanced in computers, computers don't really
"know" much about the nature of reality. AI systems use rules and
frameworks of knowledge in their internal representation of reality. If
one tells a computer that the sky is blue and that the sky is black, it
may have trouble understanding the concepts, but it could record the
information. If you tell a computer to segment its knowledge base so
that you tell Bowman and Poole that the sky is blue, but you tell
everyone else that the sky is black, it should have no trouble adhering
to the restriction. It's just another rule for the computer to keep
track of. 

But an additional complication of the bizarre structures used by AI
experts to represent reality inside computers is a lack of ability to
distinguish "right " from "wrong." Computers are psychopathic. Since a
computer can't tell the difference between a human being and the
cardboard picture of a human being, and since computers don't really
know what "harm" is, a real 1997 "Hal" can't know that it should never
harm a humans. Isaac Asimov's famous three laws of robotics are
impossible to follow, as Asimov himself demonstrated in his later
stories. What is a human? How does a computer know? When is it right or
wrong to harm a human? These literary themes explored in depth by Asimov
(but ignored by Clarke) have never been resolved by information
technology, and their resolution appears to be much further in the
future than 2001 - or even 2010. 

About the only place where Hal's capabilities can be met by 1997
technology is in playing chess. Within the limited domain of the chess
board, discussing only terms appropriate to the game, natural language
processing and speech recognition work fine. Programs exist that can
play chess at the world master level. Computer speech output is
sufficient for verbally communicating information about a chess game,
especially when augmented by an animated virtual reality chessboard. 

AI has advanced quite a bit from its pre-natal stages in 1968 to its
infancy today. The laboratory curiosities of the 1980's are finally
finding their way into the marketplace. You can buy hardware and
software today that implements, in much less polished form, a few of the
capabilities Clarke predicted for Hal. But you really, really don't want
to listen to a computer sing.  
  http://www.design.no/2001/sfx/wav/daisy.wav 

-- Ken Jenks, Editor-in-chief, Mind's Eye Fiction
   http://tale.com/ -- The First Web Publisher
   MindsEye@tale.com
