I want to write an AGI feasibility program which would respond in real time.
I believe that there are some algorithms which are missing in most fundamental
AI paradigms and even though they may be considered in more elaborate plans I
think they tend to get downplayed when competing against a lot of other
detailed subprograms. For instance, some kinds of problems can be solved
effectively with the use of weighted reasoning (such as Bayesian Reasoning) but
some cannot be. Now most people who advocate the use of weighted reasoning as
a basis for their AGI program know that a weighted method can describe a
discrete situation, so they wrongly conclude that that is not a serious problem
for them. (By the way, this leads to the recognition of a more general
problem. Just because a technique has the potential to represent a specific
situation that does not mean that a particular implementation plan will
effectively implement that potential.)
The anomalousness of intelligence does not mean that you cannot start with a
program that is relatively simple and develop something that will actually be
capable of general intelligence. This is somewhat similar to the error of
believing that since inflexible programs cannot deal with every kind of
situation that are needed for artificial general intelligence then that proves
that artificial intelligence (using contemporary computers) must be impossible.
An AGI implementation cannot be like a program in which only active conditions
are implicitly specified by (user) input. More advanced programs can use input
to specify some conditions, some conditionals and some actions on conditions.
Another more advanced programming potential is one in which the actions on
conditions can be derived based on what occurs in IO. So my implementation
would not only rely on restricted axioms where all the possible conditional
actions would be completely specified in the implementation. Of course most
programs have some of all the characteristics that I described. I am saying
that by being aware of these theoretical distinctions between elementary
programming and more advanced programming facilities the AGI programmer should
have more insight into the characteristics that seem to me to be necessary for
an AGI program and have a greater potential to use them effectively. Another
more advanced form of programming can assign (the program can learn to assign)
symbols and names to the actions that it can acquire and then use these names
in an internal or external language. (This means that the distinctions between
conditions, conditional operations and the actions taken on conditions are be
relativistic.)
So my program - which is supposed to be an initial AGI feasibility test - would
work in real time until it didn't. But I am hoping that by dealing with things
like using a trial and error method, and as I mentioned in this message, using
a trial and error method to detect when weighted reasoning can work and when it
does not work, I can get this feasibility test to create a recognizable base
for intelligence that is (recognizably) more broadly expansive than what you
see in contemporary AI/AGI programs.
Jim Bromer
From: [email protected]
Date: Wed, 17 Jul 2013 10:12:12 +0200
Subject: Re: [agi] A Very Simple AGI Project
To: [email protected]
On Tue, Jul 16, 2013 at 11:25 PM, Jim Bromer <[email protected]> wrote:
not worrying about writing something that would be scalable to adult human
level AGI.
That's OK then, Matt is bound to make you honorary member of the "without
actually accomplishing anything" club. Just joking.
Of course all kinds of simple learning have been tried for decades, like I said
mostly without the ambition to solve AGI, very often just to publish a paper
(and as I've noted before a multitude of authors of interesting papers and
dissertations ended up working in more or less unrelated fields). Jan above is
right that a lot of engineering knowledge does come from simple exercises that
we eventually discover if and how we can eventually scale - I should point out
however that intelligence is anomalous, it is not like building a hut first and
a 100-story skyscraper later, it is more like building a 100 dimensional
skyscraper . But what may have been missed by a collective IQ of a million or a
billion? I don't know but in my outline of RiskAI, an intellect that first and
foremost manages risk in its environment trying to survive, I proposed a rather
challenging starting point for AGI: real time intelligence! The basic idea is
that risk becomes infinite if you are too slow, and then again you may always
be too slow for some environments and activities, in which case you stay closer
to your comfort zone where your reaction times are not a handicap, but still
they would have to be relatively fixed and consistent.
Now, Jim, this is a perspective that at least guarantees you that you don't
fall in your complexity/recursive traps. Instead of coding learning first and
waiting for a program to respond later, you first make sure the program
responds, and then build learning around it. I am not going to lie, this can be
quite an engineering challenge, and frankly I think it is an area that will see
many breakthroughs, especially if you look at the "real-time ecosystem", for
example FPGAs and HPC where you could be guaranteed very "thin" computing power
like a million agents each running for some milliseconds. You can of course
arbitrarily choose the response time on your hardware, even 10 minutes or
whatever, but the idea is to stick to whatever limit you chose. Then you can
always claim that some hardware engineering can speed your algorithm 1000x and
make it suitable for ordinary environments.
AT
AGI | Archives
| Modify
Your Subscription
-------------------------------------------
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
