Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!think.com!zaphod.mps.ohio-state.edu!wuarchive!udel!rochester!pt.cs.cmu.edu!o.gp.cs.cmu.edu!andrew.cmu.edu!hm02+
From: hm02+@andrew.cmu.edu (Hans P. Moravec)
Newsgroups: comp.ai.philosophy
Subject: Re: Turing Test: opinions on an idea
Message-ID: <YcC8CRG00WBM83X4Fq@andrew.cmu.edu>
Date: 21 May 91 02:19:09 GMT
Organization: Carnegie Mellon, Pittsburgh, PA
Lines: 56

CC:

mcdermott-drew@cs.yale.edu (Drew McDermott) writes:

> In article <1991May13.133711.102@athena.mit.edu> mlevin@jade.tufts.edu writes
>>  ... hyperdimensional memory, or something). They then make an
>>    enormous 'game-tree' of all possible conversations in English ... 
> ...
> 2. Many conversational remarks have no context-independent responses, e.g.:
>  "What time is it?"  (due to Pat Hayes)
>  "Did you hear that? It sounded like a sonic boom."
>  "I don't know about you, but those heart tremors made me quayle in
>   fear."
> 
> There's no way to encode a single appropriate response to
> conversations including sentences like these.  There's also no way to
> rule them out that I can think of.  (Any attempt to do so would make
> it too easy for the human to always win, by subtly breaking the
> rules.)
> 
>                                            -- Drew McDermott

   This is not a good objection to the conversation tree idea.
The machine simply preambles its conversation: "It sure is quiet down
here in this inpenetrable bunker.  I'd go bonkers if I didn't have this
teletype, which is my only link to the outside world.  Thanks for taking
the time to schmooze with me.  We intelligent thinkers should support
one another." 

   Memory of the conversation that has gone before (context) is, of
course, encoded as the identity of the node reached so far in the tree
of possible conversational moves and responses (like a finite state
machine). Since
the tree is so large, this node address will be a pretty huge number--if
a typical question contains 1000 bits of essential information, and a
conversation is 1000 questions long, there will be (2^1000)^1000 nodes
in the conversation tree, so encoding the node identity will take one
million bits--not an unreasonable memory to capture this tiny fragment
of intelligence.

   A machine built on the same principle to respond intelligently to visual
and sound inputs (answering Drew's point 2 in another way) would have a
much larger state memory. If the sensory input data rate is 1 megabit
per second, then the state memory would have to have one megabit for
each second the machine is designed to exhibit its intelligence.  That's
still less than 100 gigabits per day, or 2,000 terabits per human
lifetime.  And if the table-encoded intelligence can forget some of that
deluge, then some nodes of the response tree (and all their successors)
can be merged, reducing their number.

   A neat thought experiment, I think.  If the tree is large enough to
cover a human lifetime of responses, then I would consider it as intelligent
as a human. Basically, the tree encodes all possible thoughts and reactions
of a particular person, in a very uncompact, but theoretically accessible, way.

                 - Hans Moravec