om

In order to help me get my mind off my despondency over flunking out of
CS350 again, I write this post. Before I get to my subject, I need to
make it clear that it is not because I am lazy but because I'm being
shafted. I've flunked so many classes that it has become a joke to me. I
like to make my failures spectacular too. I don't just stop working
after I learn there's no chance to pull up my grade and go quietly into
a menial dead-end job, instead I attend every single class session right
up until the very end, there being more glory in going down in flames
than bailing out... I turn in every assignment even knowing it's futile.
When I get my F I am not ashamed of it. I don't say "if only I could
have worked harder". I stand up proudly and say, I did my part, I'm
being shafted. This class, is a sterling example, even though I'm past
the days when I'm OK with this type of F. I logged two hours on the
first assignment. The program worked perfectly. The documentation was
flawless, the forms were filled out. The test cases were there. I got a
50. On the second project, I spent four and a half hours, (258 minutes)
It was designed according to the spec. (Apparently I missed one of the
computations because no input or output was specified explicitly.) The
numbers it did compute were absolutely correct. It was packaged with the
standard automatic build system that _ALL_ GNU projects use. Therefore I
got -30 just because the grader couldn't open the README file and, with
other mark-downs, ended up with a 32. It's probably the second lowest
grade I ever worked for. The first assignment I did for the previous
time I took the course was almost as good, It earned me no points
whatsoever. Do I deserve this?

Sorry for getting side-tracked but I can't help but comment on my own
life going down the drain. =(

om

The subject is, again AI. I have been reading Jeff Hawkins' book "On
Intelligence". For the parts of the book he's not talking about his cat,
he presents a very good theory that broadly agrees with what I've been
able to come up with myself. My subject, however, is not about on what
points I do and don't agree with Hawkins about but rather in _how_ my
thinking on the subject differs from his.

The meat of his book (which is not about his cat or "Somewhere Over the
Rainbow".) is about his "memory prediction model" of artificial
intelligence. He believes that the brain's fundamental operation is that
of making predictions about things. He seems to believe that this is a
new model of computation and that our present computers are somehow
inadequate to the task (more on that point later).

His choice of concepts has had the effect of directing his thinking
towards how the brain accomplishes this predictive task. This focus has
had the benefit of leading him to focus on how failures in recognition
at lower stages in the the cortical heirarchy can lead to a shift in the
attentional state of the mind towards that stimulus. This is an
interesting line of thought that I had not yet examined in great depth.

In my own study, I had decided to call what is essentially the same
idea, an "abstraction". The idea being that abstractions capture
information, in general, and that the process of detecting instances of
abstractions in the environment (and in lower areas of the brain) can,
in some instances, be called "prediction". The focus of my own work has
been on determining how to extract information from the environment. I
make no claims as to the relative merit of our two approaches.

The interesting thing here is the differences in how the varrious AI
researchers think about their subject. The terms we choose aren't as
superficial as they may seem. To be truly effective at what we do we
need to be aware of how we approach the concepts we use to understand
what we are studying. Ideally, each of us will develop several
ontologicly distinct theories in paralell and use them to cover the
holes in the others. I don't suggest that is an easy task.

[ I have a note card of about ten points that I want to cover in this
essay, finding the most "natural" ordering for them is prooving to be
non-trivial... om... ]

Keeping multiple cross-refferanced ontologies is a step in the right
direction. Just as important is taking a step away from the wrong
direction. Everybody who's worked in AI has run into crackpots. Most of
us have run into the same crackpot. These people are so enamoured of
having their name attached to a theory alegedly about AI, that they
don't care if their idea is rubish. Other people in the field tend to be
more diciplined but not nearly as much as they really should be. One of
the most neglected tasks in the field of AI is the critical evaluation
of concepts and theories.

Instead, what few discussions on AI I've seen, tend to focus on
shoehorning the facts into the trendiest theory or try to design a
system to implement utter nonsense. My favorite example of this type is
the notion of "Self-consciousness" or "self awareness". The idea being
that the brain magicly achieves intelligence the moment it becomes "self
aware". This notion can be found in many works of fiction as well as
reams of discussion on the internet.

What exactly is self awareness anyway? The brain is devoid of sensory
organs, so it's not that... People have been trying (and mostly failing)
to figure out how they think for aeons so self awareness is not the act
of viewing one's own thoughts. I'm running out of ideas here but the
point is that as soon as one tries to evaluate many of these concepts
critically (as opposed to struggeling to find supporting evidence for.),
they go "poof". These are concepts that are beyond useless, beyond even
false, they're outright lies people tell themselves in order to deal
with their own existential angst and, in doing so, obstruct any progress
towards AI.

We can follow the trail of bad ideas out onto the factory floor to see
the true destruction that has been wraught. At some point between the
categorical sylogism and the invention of boolean algebra, the notion of
intelligence has changed from the act of making order out of chaos to
the process of making logical deductions. While it is true, that in
humans there is a positive corolation between how smart we think someone
is and how well they perform logic. What has only become obvious in
recient years is the focus on logic, deduction, and ontologies (as the
substrates of the AIs themselves), are inadequate to produce real AI.

Some people are good at logic and others such as the people who are
generally considered to be incapable of logic (such as our president...)
are still considered human... (mostly..) indicates that the primary
function of the brain must be in performing operations on which logic
can be built rather than logic itself. When one thinks about it one can
almost be amused at all the attempts to create the cause [ what the
brain really does ] out of elaborate constructions of the effect [
symbolic reasoning].

A critical thing to keep in mind is that the brain was produced by an
unintelligent process. This is evinced by the fact that while the brain
contains plenty of concepts (with the exception of the example of George
Bush), the brain is composed of utterly no concepts. We need concepts to
understand the brain but in using such concepts we must not forget that
these concepts are artificial constructs intended to bring some level of
tractability to the vast complexity of the brain. Unfortunately, we
don't yet know which details we can safely ignore, or the relative
importance of details that are important.

Instead of looking for better concepts and deeper understandings, people
tend to blame the weather. This is the heading I've decided to give all
the Roger Penroses of the world who claim that intelligence will require
some bizarre new type of computation, quantum gravity effects, and the
like. It also includes people who claim that AI requires computers
sufficiently advanced to be so far in the future that they're in little
danger of being discredited before people have forgotten their names.
These people are just plain annoying. They are trying to hide their
cognitive deficitts behind lame excuses. Other than that, they can
safely be ignored.

The people who really get on my nerves, however, are the people who
aren't even trying to work on AI yet call it AI... (There are also
people who claim not to be working AI but are, infact, producing the
most valuble and promising research in the field!) They are the people
who caused people such as Ben Goertzel to coin the term "AGI"
(Artificial General Intelligence), and to use phrases such as "Real AI"
as I did above.

While these people may have, at some time, been working on AI, what they
eventually produced will, by their own admission, never come anywhere
near human intelligence yet they still call it AI resarch. A different
bunch of the same breed is actively trying to pawn off things such as
NLP and what have you for applications that really require true AI.


The last group of people who think about AI I want to cover are the
singularitarians, especially the SIAI cultists. There are people out
there who get, for lack of a better word, dazzled by the prospects of AI
and think it will solve all of their problems in an afternoon. This is a
somewhat bizarre line of thinking. The current cult leader is Eliezer
Yudkowsky. In his long rambling essays, (about a hundred times longer
than this piece but hardly saying as much...) he goes on about a
speculated capability of his AI to simulate the development of your mind
into the future in order to determine what you really want when you ask
it to do something for you. -- He expresses about a hundred permutations
of this basic idea but that's the gist... He supports this idea with
claims about the scale of raw computation available to the AI.

He presents a number of arguments about the subject which do have some
merit. However, he overlooks and vigorously denies the fundamental
undesirability of such an arrangement (its the central tenant of his
religion). A much more plausible (And desirable) role for an AI is that
of a highly adept R&D department. The client/user will need to set goals
and monitor the progress and logistics of achieving those goals. Far
from reducing us to mere dependents, the best AI will require the best
from us, our best rationality, our best morals, our best intentions.
There is no such thing as salvation, nor is there such a thing as a
savior (as Eliezer paints himself out to be).

Transhumanist singularitarians seem to think that the AI will respond to
requests such as "make me godlike" or "upload me into virtual paradise."
The AI may indeed do something but it won't be what you wanted. The
typical notion of transhumanism requires that you, at least in part,
become the AI. Uploaders tend to talk about the speedups they claim to
be possible with whole brain emulation. They talk about uploads being
able to do a thousand years of work in a minute. (or some speedup on
such an incredible scale). What they don't seem to consider is what it
would actually be like to sit at a simulated desk working on a
thousand-year project.

In the more desirable neural interface scenereo, you will have to train
your mind to be able to interface with the artificial substrate and get
useful work from it. Taken to its logical conclusion, your mind will end
up being something almost entirely alien.

I guess what I'm trying to reiterate here is that you don't get sumpfin
for nutin... When I'm not busy flunking classes, I make myself busy
working on AI, and its related subprojects. I don't pretend that when I
achieve it, I'll be any less busy on what comes next.

Thanks for your time.

om

--
Friends don't let friends use GCC 3.4.4
GCC 3.3.6 produces code that's twice as fast on x86!

http://users.rcn.com/alangrimes/