From: Billy Brown (bbrown@conemsco.com)
Date: Wed Feb 03 1999 - 06:29:35 MST
<continued from part 3>
OK, here's where I start venturing onto thin ice. Given what we now know
about intelligence, what can we say about intelligence enhancement?
Obviously we don't really know how to build an intelligent mind yet, so we
can't make any firm predictions. However, AI has advanced to the point
where it seems plausible that a mind could be build using complex
combinations of the problem-solving methods that are currently known. If
that is indeed the case, then we can anticipate some of the general
characteristics of IE:
Increased Speed
Assume we have an intelligent entity implemented in software (it could be an
AI, and upload, or some sort of hybrid entity - our only constraint is that
we assume someone knows enough about the entity to make changes to it).
What benefits does it gain if we give it faster hardware to run on?
If we make no changes at all to its software, this will simply make the
entity think the same thoughts faster - from its point of view, the rest of
the world slows down. However, the task of modifying the entity to take
advantage of the increased speed is simple compared to the task of creating
the entity in the first place.
Modifying decision-tree abilities to take advantage of the extra speed is
trivial - it could even be made automatic for an artificial mind. Modifying
data-transformation abilities to work better is an engineering problem of
reasonable complexity. Knowledge bases will not be much affected - they
simply do the same things faster.
What would this mean in terms of intelligence? Based on current experience,
we would expect geometric increases in processing power to yield at least
linear improvements in ability for both decision-tree and
data-transformation problems. Some abilities will improve even faster,
while a few (like playing tic-tak-toe) will become 'solved' problems and
stop improving. Learning speed will enjoy an increase somewhere between
linear and geometric - the knowledge bases and other infrastructure enjoy a
geometric increase in performance, but we would expect real-world
constraints (such as network bandwidth, or the time required to perform
physical actions) to place some limits on the rate of improvement.
As an interesting note, it would appear that the effort required to take
advantage of each speed increase grows relatively slowly - a mind that knows
enough to make the necessary changes should have little trouble taking
advantage of all available processing power.
Optimization
A potentially much more powerful approach to intelligence enhancement is
optimization - using better heuristics, more sophisticated decision-tree
searches, more efficient data-transformation algorithms, and so forth. This
approach can sometimes yield amazing results - performance improvements of
several orders of magnitude can often be achieved through relatively
straightforward optimizations.
Unfortunately, this sort of improvement also tends to be self-limiting. In
any given problem domain there comes a time when all of the known methods
for improving performance have been applied, and there are no more obvious
improvements that can be made. Then you are reduced to inventing new
solutions, which is a process of scientific discovery that requires large
amounts of effort and produces only erratic results.
Temporary Specialization
Another approach to IE would exploit the inherent flexibility of
general-purpose computers by dynamically allocating processing resources.
The idea here is to allocate most of the mind's processing power to whatever
problem it is concentrating on at the time, rather than allocating a fixed
amount of power to each cognitive ability. The result would be a fairly
substantial increase in effective intelligence, equivalent to the effect of
a substantial speed increase (something like x10 to x100, depending on how
much temporary specialization is actually possible for a problem).
Long Thought
An entity with access to large amounts of data storage space can make
another interesting tradeoff. As computer scientists have discovered,
allowing more time to solve a problem has exactly the same effects as using
faster hardware. As long as the entity does not run out of memory, it can
take 10, 100 or even 1,000 times longer than normal to think about a
problem, and reach solutions that would normally require a higher level of
intelligence running on faster hardware. Of course, many problems are too
time-sensitive for this approach to be feasible.
---- Well, I'm sure there are lots of other ways to approach IE, but I haven't thought of them yet. Suggestions are welcome - I would like to include more possibilities in the list. Billy Brown, MCSE+I bbrown@conemsco.com
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:02:59 MST