From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!think.com!spool.mu.edu!uunet!psinntp!scylla!daryl Tue Jan 21 09:26:55 EST 1992
Article 2862 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!think.com!spool.mu.edu!uunet!psinntp!scylla!daryl
>From: daryl@oracorp.com
Newsgroups: comp.ai.philosophy
Subject: Re: Table-lookup Chinese speaker
Message-ID: <1992Jan18.134742.4155@oracorp.com>
Date: 18 Jan 92 13:47:42 GMT
Article-I.D.: oracorp.1992Jan18.134742.4155
Organization: ORA Corporation
Lines: 101

Mark Rosenfelder writes:

> But let's pretend that the database of "conversations which pass the Turing
> test" can be specified and does exist.  It does NOT follow that a table
> lookup machine randomly choosing conversations from this database
> will itself pass the Turing test.  This is easily demonstrated.

If there is some sequence of exchanges that you consider to not to
pass the Turing Test, then that exchange shouldn't be in the database,
by definition.

> Suppose you start the conversation with "Hello".  The computer responds
> with "Hello, Professor Chung!"  It can say this because there is in fact
> a successful Turing-conversation in the database which begins this way--
> it's a conversation with Prof. Chung, naturally.

As I said, I consider a coversation to be the complete exchange
between the two people. When Professor Chung first talks to the
Chinese Room, it is not sensible that the CR would know that it was
Professor Chung until *after* Professor Chung introduces himself.
Therefore, I don't believe that a conversation starting out

   Q. Hello.
   A. Hello, Professor Chung!

*is* a sensible conversation. I didn't want to try to specify what I
mean by "sensible conversation" in too much detail, because I'm not
actually suggesting that anyone try to build such a database, but it
seems that we keep getting bogged down in the particulars.

     1. The Chinese Room must be input-total. No matter what the human
        types in, the CR must have a sensible response. I don't consider
        the computer crashing because I type in "XBJASDF@!" to be a sensible
        response; it is certainly not a response that would pass the Turing
        Test. A passing response might be "There seems to be some noise
        on the line, I'm just getting gibberish." This prevents a "Star Trek"
        style attack where the human tries to destroy the computer by giving
        it nonsense.

Since it is the Chinese Room that is trying to pass the Turing Test, I
am only going to require that the conversations be sensible on the
part of the CR; the human, who is not worried about passing any test,
can say absolutely anything, and the CR must come up with a sensible
response. However, given a complete set of two-way sensible
conversations (where both participants respond sensibly), it is not
too hard to make a larger database that includes the possibility of
nonsense from the human. If the human says something that does not
occur in the database, the computer can always say: "I have lost the
thread of this conversation. Could you say that again in different
words?" or something.

     2. The conversations must start with the CR having no knowledge
        of the human's identity, and may only gain knowledge through
        the conversation. That is, the computer simulates the situation
        of a man locked inside a room who has no access to the outside
        world except through his teletype conversation.

> Your next reply is "Huh?"  And the computer explodes, because it finds
> it HAS NO CONVERSATIONS which begin with these three sentences.  The
> successful conversation it used last time proceeded in a different way.

If the database has no conversations with the word "Huh?" occurring,
then it isn't a complete database. For absolutely any human exchange,
"Huh?" is a permissable response, so the CR has to be able to respond
to it.

> You can restrict the database as narrowly as you want (say "conversations
> with Daryl McCullough which pass the Turing test") without avoiding the
> problem.

The conversations are *not* conversations with Daryl McCullough or
conversations with Professor Chung; they are conversations that start
off with the CR having no idea who he is talking to.

> There must exist some conversation C in the database, which
> is the ONLY conversation which begins with the sequence s1, s2, ... sn,
> where these are particular statements.  (This claim must be true if the
> database is finite.)

My database includes *all* conversations of less than 100 years in length.
So, yes, if the human lives long enough he can force the computer to run
out of things to say. The same thing happens with people, though, if you
try to have a conversation lasting 100 years.

For conversations of more moderate length, there will be *no* conversation
that has a unique beginning sequence.

> The computer has just uttered statement sn.  If you utter statement s(n+1),
> all is well-- the computer can still proceed with conversation C.
> But if you utter any other statement, (e.g. t) it has failed the Turing test,
> because it has no entry in the database of passing conversations which
> begins with s1, s2, ..., sn, t.

If the length of the conversation is less than 100 years, then the
computer will have a response. If the length is greater than 100
years, then the computer can pretend to be having a heart attack and simply
respond: "My heart! I'm dying!" and then nothing more.

Daryl McCullough
ORA Corp.
Ithaca, NY


