Mark Waser wrote:
Richard,
I'm afraid that you have successfully talked me out of the complex
systems camp.
> Richard Loosemore wrote:
>> 6) A system is deemed “complex” if the smallest size of a theory
that will explain that system is so large that, for today’s human minds,
the discovery of that theory is simply not practical. Notice that this
definition does not imply that there any such systems in the real world,
it just says that *if* the theory size were ever to go off the scale
*then* the system would (by definition) be complex.
I just don't believe that the core of intelligence is complex
according to this definition. The combination of the core of
intelligence *plus* the world is clearly complex but I don't believe
that a boot-strap intelligence need be complex (by this definition)
This definition is *not* what I understand to be complex.
Okay, let's take an example to illustrate it.
You know about those (abysmally depressing) experiments in which cats
are raised so that they only ever see vertical stripes, and then as
adults their brains cannot detect horizontal stripes?
Assume, for the sake of argument, that a similar principle exists at a
higher level of the cognitive system. Specifically, assume that the
adult human mind is largely driven by "operators" which act on the
"symbols", in such a way that the operators are a bit like biological
catalyst molecules and the symbols are like the atoms and molecules that
are manipulated by the catalysts.
We will suppose that (quite plausibly) most of these operators do not
come hard-wired, but are grown as a result of developmental experience
(which means, experience of thinking and experience of the world, both).
Now one more assumption, which requires a little more imagination. This
has to do with what kinds of things these (finished, adult-version)
operators really are ... because we have two choices:
Choice 1: The operators end up being clean and modular in their design,
which means that if we were able to examine them from the outside, we
would be able to understand how they worked because their structure was
NOT deeply entangled with the design of the symbols, and the other stuff
in the system. Call this the "God Is A Smalltalk Programmer" choice,
because in this case the adult version of the cognitive system, after
all the operators have developed, looks like a nice piece of OO
programming, with no hideous dependencies between the entities.
Choice 2: [And I am sure you can see this coming already] Suppose that
the operators develop in such a way that there are hideous dependencies
between them ... like, really horrible design in which it is almost
impossible to see how the operators work because everything developed as
a big kludge. This is the "God Is A Spaghetti Programmer" choice, for
obvious reasons.
In choice 2, we are talking about cognitive systems that have the same
core mechanisms (the same basic drivers that are there before the
development process kicks off and the operators start getting built and
the symbols start being collected), but the actual format of the
resulting operators and symbols is deeply ugly. It may well be that
every individual mind actually looks quite different on the inside, even
though the net result is a system that thinks in a roughly similar way
to all others.
Now, am I saying that the cognitive system is definitely built this way?
No, but it is at least a serious possibility. I could make a stronger
case in favor, but rather than do that I will just say that this is at
least as plausible as the alternatives.
It should be clear that the second choice is a complex system. Can I
take that to be agreed?
Is the second choice more plausible than the first? Well, yes: when a
bunch of mechanisms interact in the way described, the usual result is
not Smalltalk-like OOPSLA beauty, but a horrible (though functioning)
kludge. That is just the way nature is. Agreed?
The question is, can we come along and try to build a clean, OO-like
version of this operator system and get a non-complex version of an
intelligent system working? My argument is this: if the only working
examples of an AGI are these nature-built kludges, what are the chances
that a clean version can be built? In reasonable time? Without being
able to dissect and play with a working model of a nature-built system?
I cannot see any reason that COMPELS me to believe that such a clean
version can be built. It may be that the only version of intelligence
that can ever be built, is one which grows up from some initial
(pre-operator) mechanisms, and that ANY system that attempts to function
with operators built after-the-fact will slowly diverge from stability
as it interacts with the world.
In fact, looking at complex systems themselves (the real examples that
people play with in the computer laboratory), I strongly suspect that we
cannot do this.
How does this relate to what you said above? Well, your statement that
you "don't believe that the core of intelligence is complex according to
this definition" seems, to me, to indicate that you have compelling
reasons to suppose that the above scenario really cannot be correct.
What makes you believe so (and believe it so strongly)?
Richard Loosemore
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com
