🚀 go-pugleaf

In article <4689@itivax.iti.org> dhw@itivax.UUCP (David H. West) writes:
>Real intelligent systems (e.g. humans) function quite successfully
>at a finite temperature despite the influence of thermal
>fluctuations (Brownian motion), which cause finite random perturbations
>of everything.  A finite system embedded in a thermal environment
>cannot encode an infinite amount of information.

A finite *digital* system is limited in the density of its states by
thermal (and other) fluctuations.  You are right to point out that the
brain is limited in its computational power by this fact.  I think that
the most productive and interesting approach to cognitive science (if not
artificial intelligence) must take advantage of the brain's limitations;
CS wants to learn about how the brain computes and the algorithms it
follows.

But the brain viewed as a computational system is different from the brain
viewed as a causal system.  The same distinction applies to electronic
computers - there are plenty of hardware failure modes to complicate the
causal description of the system which are irrelevant to the computational
description.  If we want to claim that we've accurately modeled the
computational power of the brain by demonstrating that our model is
faithful to the causal interactions in the brain, we're stuck with
modeling the details (including thermal motion) down to whatever level a
critic might demand.

The word "encode" might be ill-chosen for causal systems.  If position &
velocity (et al) encode anything, the code would have infinitely many
"symbols," one for each state of the system.

David continues:
>This would seem to indicate that your argument has no bearing on
>what may be necessary for intelligence, (...)

This is true.  Certainly there is no reason to believe that infinite
processing power (or infinite anything else) is necessary for
intelligence.  What I think we can conclude is that a familiar "safety
net" argument for AI doesn't quite work (different from the original):

(1) The brain is a causal system that thinks
(2) We can model causal systems with numerical methods
(3) Therefore, we can make a model that thinks .

Searle (I believe) would object that we would get a simulation of
thinking, not the real thing.  I'm objecting that it is physically
impossible to make an adequate model.  If we could let the model crank
as long as it needed before responding (ie relax the real-time constraint)
then my argument would not hold.  The model would eventually converge
close enough to the real system to be indistinguishable.  But that won't
pass the Turing test.