om
Having no funds to build the robotics lab I was planning. ($40k already
budgeted, ~$10k saved). And having no VR... (you would imagine that
uploaders would jump at the chance to build the 'reality' they want to
spend eternity in...)
Because making progress is becoming ever more urgent, I'm forced to
attempt to move forward on a purely theoretical basis. This is where
most other people on this list go and metaphorically hang themselves.
Their theories become too ungrounded and they start writing the type of
crap I immediately delete. (I'm looking at you, Jim.)
Knowledge representation (KR) is it's own sub-field because it is both
an obvious and major problem. The trend in the field, indeed one that
seems to be supported by psychological evidence, is the use of linear
models. Linear models are popular because the mathematics is fairly well
developed and because immensely powerful linear solvers (ie Nvidia Tesla
processors) are readily available.
To understand where I'm going with this, I need to review my underlying
theory in a nutshell. Basically the brain is CONSTANTLY imagining. When
you are awake it uses feedback from the environment to correct what the
imaginative process is producing -- this is called perception.. When the
imagined environment basically agrees with the signals from the sensory
organs then the top level description that generated the imagined
environment must also bear a close resemblance to a top level
description of the environment. This top level description, in turn,
could be sufficient to support any of the reasonable classic AI approaches.
Implicitly, this means that the AI system must be capable of generating
photo-realistic images, high fidelity audio, etc... It also means that
the limit to what it can understand is the limit to which it can model
(imagine). It also needs the capacity to understand social interactions
and subjects that humans can only approach mathematically. The
limitations of the human cognition are well documented.
So the first thing I need beyond a linear model is a model that can
scale dynamically to meet needs. This fixed-length vector bullshit is
just not going to cut it. =\ Second, It's going to need to be able to
re-structure itself as it's environment and demands change. So not only
will the network be dynamic in size, it will also be dynamic in
structure. A space-faring AI will probably need a module for orbital
mechanics, etc... None of this shit can be pre-programmed. It must be
learned. I would propose that the AI start out with the most minimal
pre-established kernel and develop all the higher functional areas
dynamically.
Finally, I want to get past linear approximations of the non-linearities
of the world. The Big Weenie makes a convincing argument in his book,
"How to Create a Mind", that the brain uses a system of redundant
representations to solve certain problems. For example you might have
many neural representations of some person or object, each handling a
different angle of orientation, etc. But this is obviously inferior to a
truly 3D model.
Yeah, I'm being ambitious here but because I'm NOT AN UPLOADER I am able
to think past the limitations of the brain and contemplate the ultimate
AGI (then build it).
Once I finally figure this shit out, the code for this motherfucker is
probably going to be on the order of Forth. -- really fucking tiny. At
the size that I envision, it will be possible to prove it mathematically
and at that point basically have a technology that I can throw at any
AGI problem and never have to re-visit. =P
Rabbi Yudkowsky likes to masturbate about ideas of recursive
self-improvement as if it were an un-bounded process. My model is that
there will be a few iterations of the software until it is
mathematically proven, at which point it is basically perfect. The
hardware can also evolve through a few iterations until it is the ideal
architecture for running the software. And the scale of the system can
be built out until most nodes are simply idle most of the time for lack
of things to think about. Bottom line is that the rabbi's thought
patterns with regards to expanding intelligence so closely mirror my
thinking about expanding boobs that the parallels can't be discounted.
(nobody's going to read this far into anything I write so I can
basically say anything at this point in the text.)
oops, some people start reading at the bottom to see if the author ever
got to his point and then read backwards to see how the point was
developed... better pad out the bottom of the message now...
So I do have a partial outline of the system I need but I still have a
few math problems to solve until I am confident that I have a system
that can go the distance here.
--
NOTICE: NEW E-MAIL ADDRESS, SEE ABOVE
Powers are not rights.
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
