Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!mcnc!rti!ntpdvp1!kenp
From: kenp@ntpdvp1.UUCP (Ken Presting)
Newsgroups: comp.ai
Subject: Re: Searle and Radical Translation (was: Re: Searle and Biology)
Summary: How to get observations which imply intentionality
Message-ID: <612@ntpdvp1.UUCP>
Date: 26 Jul 90 00:28:52 GMT
Organization: SNA Solutions Inc., Contract Programming Group
Lines: 118

> (Daryl McCullough) writes:
>
> > (Ken Presting) writes:
> > . . .  The interpretation "succeeds" if some symbol in the
> > system's vocabulary gets assigned the semantics of the concept
> > "truth".
>
> This is a very interesting idea until the last sentence. Why not say
> that the interpretation succeeds if there is any interpretation which
> makes all the sentences true? Why must there be a word for the notion
> of truth?

The main reason is that "truth" is something we are all capable of talking
about, and we define most logical concepts in terms of it.  Also, I doubt 
that you would get anywhere near a unique interpretation for non-semantic
terms, unless you reach a point where you can discuss semantics with your
interlocutor.

My guess is that (for Davidson) the ultimate motive is Kantian - if you
can establish that some organism can use logic and be swayed by arguments,
then you have a moral duty to reason with it, instead of merely exploiting
it.

Anything that understands the concept of truth can, on that basis, understand
validity, soundness, truth-conditions and truth-functions.

> > IMO, real computers are more than mere TM implementations, so
> > Tarksi's theorem does not prohibit real machines from representing
> > truth.
>
> Ken, IMHO, it is completely bogus to use Tarski's theorem to "prove"
> that there is some notion of truth that physical objects can know but
> Turing machines can't. Tarski's undefinability of truth has just as
> much force when applied to humans:
>
>      Theorem: No human being can have a definition of truth and still
>      be consistent.
>
>      Proof: Consider the sentence "This statement is not true". There
>      is no consistent truth assignment to the sentence if we are to
>      assign the word "true" its usual meaning.
>
> My guess is the human's are just inconsistent.
>

Tarski's solution is to define separate "truth" predicates in a metalanguage.
(Cf. "The Semantic Conception of Truth").   There are many formal approaches
to resolving the Liar paradox.  Kripke would say that the Liar sentence is
"ungrounded" because it refers only to sentences, and that it has no truth-
value.

I tend to agree with your position, however.  But this position entails that
arguing in natural language is all but pointless.  (No news to Netnews).
To ameliorate the negative consequences of "semantic closure", I also think
that most of natural language is used metaphorically, whenever it appears
in arguments.


But I think it is perfectly obvious that real computers are more than TM's.
Not because they are physical objects, but because they have a permanent
memory.  Each time a TM is given a particular input, the tape is erased
and the internal state is reset.  So the TM computes a *function* - same
input => same output.  That is obviously false for real computers, which
have updateable databases.

Theorems about TM's, which compute functions, do not apply to real machines,
for purely formal reasons.  Real machines do not, in the technical sense,
compute functions.

> Notice, however, that if we weaken our claim to having a definition of
> absolute truth to having a subjective definition of truth, then the
> paradox is resolved:
>
>      Consider the sentence: "This sentence will never be considered
>      true by Ken Presting"
>
> The above sentence could very well be true, but if so, then there is
> at least one sentence that is true, but Ken Presting doesn't consider
> to be true. Therefore Ken Presting doesn't have a definition of real
> truth. On the other hand, the above sentence could be false. Then Ken
> Presting will consider it to be true, and so again Ken Presting's
> notion of truth is not real truth.

Your conclusion follows only if we assume an *extensional* concept of
truth.  For example, if the only way it can be established that "my
conception of truth is real truth" is by establishing that there is
some act (possibly internal), such that I perform that I perform that
act when and only when presented with a true sentence.  Then my
performance of that act (say, "considering P true") would be extensionally
equivalent to "real truth".

But that is not the only way to represent truth, or any other concept, for
that matter.  Even TM's are used more flexibly than this, to enumerate
recursively enumerable sets.

Davidson's idea is based on Tarski's:  represent the truth predicate of
one language with an axiomatic theory stated in a metalanguage.  For
a system to use this method of representation, however, it must have
intentional states with intensional content.

So the idea is that you collect enough evidence to convinve yourself
that your system is correctly using the word "truth", and (here is the
kicker) the word "truth", as used by the system, cannot be interpreted
as having any other meaning.  If you can collect enough evidence to
reach this conclusion, then you *have* to grant that the system has
intentional states.

Whether it is possible to collect enough such evidence is another
question, but it's for sure that a TM, without permanent memory, could
NOT provide the right kind of evidence.  According to Davidson, you
should look for the machine making mistakes, and acknowledging that
they are mistakes (which of course involves trying to correct them, or
avoid them in the future).  A TM with a (partial) extensional
representation for "truth" would of course make many mistakes, and might
well acknowledge them, but could not avoid them in the future.


Ken Presting   ("A heuristic is an algorithm that doesn't work")