Re: consciousness based on information or computation?

From: <hal.domain.name.hidden>
Date: Sat, 30 Jan 1999 14:46:42 -0800

Several people have noticed that a very simple program can create a
pattern or even, by some definitions, a process which imitates our minds.
A trivial counting program eventually creates all possible patterns,
and you can map any string into a four-dimensional space-time diagram
which is isomorphic to an entire universe. It seems, then, that even
a trivial program like this ought to be able to create structures which
are identical to our minds down to the smallest detail, and therefore,
perhaps, we should conclude that we are likely to be the product of such
a trivial program.

If I understand him, Wei suggests that a way around this is to notice that
the valid mind or valid universe produced in this way is buried amidst
an unimaginably large amount of garbage. To estimate the contribution a
particular program makes to instantiating a mind we should look not only
at the size of the program, but also at the difficulty of identifying
or locating the mind amongst its output. Here we have a very simple
program, but if we then wanted to point to where in its output we had
a mind-creating universe being produced, we would have to specify a
huge amount of information. Combining these two we find that the total
information needed to both *create* and *locate* the mind in the output
of the program is very large, making the contribution of that program
be very small.

This approach also offers a solution to the problem that even in our
own universe, you can select subsets of, say, water molecule collisions
which produce patterns identical to our own consciousness. A vat of water
as big as my brain has billions of water molecules in each region which
is of the size of my neurons, and in the time my neurons fire, billions
of water molecule collisions will occur in each such region. By paying
attention to only a tiny subset of these water molecule collisions, we
can find a pattern of collisions which is identical to the pattern of
firings of my neurons when I am thinking. Of course, to do this we have
to ignore the billions of other collisions we aren't paying attention to,
but it might seem that adding more collisions cannot change the existence
of the patterns that are sufficient for consciousness.

However, Wei's approach calls attention to the size of the program needed
to select those molecules to pay attention to. In practice we would seem
to need a full-sized neural simulation to do that. This would be a very
large program, so the actual contribution of such random collisions to
my consciousness would be vanishingly small.

This approach seems to move away from the question, is this system
conscious? That is no longer seen as having a yes/no answer. Instead, we
try to say, how much contribution does this system make to a given mind?
If we can have a short program to define the laws of physics which govern
the system, and another short program can identify the mind states with
states of the system, then it makes a relatively large contribution.
But if either of these two programs must be large, then it makes a
small contribution.

It is somewhat counter-intuitive to downgrade the mental contribution
of a pattern just because it is lost among a large amount of noise.
>From the external perspective, maybe it makes sense; information that
is buried is less useful than information which is easily exposed.
But mental activity seems like an "inner" quality; the string doesn't
"know" whether it is surrounded by noise or is out in the open; it should
be equally conscious either way.

However if we accept the initial idea that short programs make more of
a contribution, I don't see a problem in saying that short *output* also
makes more of a contribution. Why, after all, should minds produced by
short programs be considered to make a larger contribution than those
for long ones? We have to make some implicit assumptions in order to
make this work. Sometimes we imagine an actual computer cranking out
programs and running them, with short programs getting more run time
than long ones.

With this kind of mechanistic picture, we could also imagine that programs
which take more time to produce their output could also make less of a
contribution. Programs which produce larger output could be penalized
in this way or simply by virtue of their greater resource requirements.

There does not seem to be a unique measuring system for estimating
the contribution of any given program to the overall reality.
Penalizing programs for program size, run time, or output size all seem
somewhat plausible. With this kind of model it does make sense that the
contribution a given output pattern makes would depend on details of how
it was produced by the program, including how much other output there was.

Hal
Received on Sat Jan 30 1999 - 14:54:44 PST