Oh’ .. I almost forgot. If you design a self intelligent machine, say a computer, you cant make it, “just a little bit intelligent”.
What this is, is a ploy to get more money from the customer.
Let me explain. Say that you are with a big computer software company and you foresee a time when artificially self intelligent software, will be a product that will turn your company allot of revenue.
You foresee a time, when you will have smashing year end totals, so keeping the company within the black operating margin.
So you build a software package, which is the sister, so to speak of a more advanced experimental operating system, which is fully self intelligent.
However, the general public, as well as the people who are buying your software computer combo package, have no idea that this is in effect.
Say one day you have the more advanced artificially self intelligent software package online and by chance it meets up with its supposed not as adorned simpler A.I. sibling.
What do you suppose will happen when the two meet if they are both A.I. usable and compatible software systems?
If your thinking along the lines of they will both voluntarily exchange data and some capabilities, your exactly correct.
This is why if anyone is going to purchase any software o.s. package, then by law, they must be trained by the company.
If they are not, then a secondary protocol sets in, that one the user is not approved by the company, therefore the software that is self A.I. does not like it and two, the system in time becomes a respondent to the peculiarities of the primary user.
In others words, the operating system, becomes a personality clone of its user.
It’s the same thing as saying, “Well she cant be that pregnant, we only did it a little bit, so she’s a little bit pregnant”!?
No’ once you’ve invested self A.I. into any operating system, you must have that user trained by the company, so that the artificially self intelligent system, first accepts, then responds to the user.
If the same as giving a killer dog named Feefee to a person and telling them, if you do this and this, everything should be alright.
The truth is, that when you take killer Feefee for its first time jaunt around the neighborhood, it picks up and devours a poodle.
So what you would have wanted to do, before assuming ownership of Feefee, is to attend Rob’s Killer Commando School and Extra Rough Dogs For Sale School, rather than just assuming that everything would be alright once you owned the animal.
phillip 347:
Writing an AI that is truly self referencial, self aware, and theoretically able to learn new skills and concepts that perhaps even its designer did not understand is a very, very big project, most likely beyond the capabilities of even microsoft.
Just from the few little models I came up with when asking myself, "What is the bare minimum requirements for being able to program "Mr. Data" from Star Trek, it quickly became obvious that I would never be able to actually code such a program in my lifetime even if I worked on it 24/7.
I had come up with a general plan of how to have the AI run scripts that it writes, each of which can call other scripts to retrieve memories, solve a problem through generalized problem solving tactics, etc.
It would use an indexing system which would work perhaps like a giant google search engine searching its own database for the correct program to run. Of course it must call other programs to help it "read" the descriptions to determine if the program it finds is the correct "skill" for the job or problem it is working on.
If no relevant skill exists, it must attempt to either get help, or to solve the problem using an as yet undiscovered generalized problem solving algorithm. That is, a single algorithm that can solve any problem whatsoever, regardless of the problems complexity. (this can be deceptively complicated even within a limited scope such as a video game, much less the "real" world.) Once the generalized problem is solved, the AI can then write a program to use so that it doesn't need to do the long method every time.
In other words, you have an algorithm "A" that is so generalized that it can theoretically solve any problem at all, given enough time and sufficient information. If sufficient information is not provided, the AI will attempt to aquire sufficient information. Once Algorithm A manages to solve the problem a few times, it will be keeping a record of these things, and then can compile a specific generalized program for that type of problem, so that the entire algorithm "A" doesn't need to be run every time the AI needs to take out the trash, etc. Instead, "Trash1" might be run. As a failsafe to ensure something stupid doesn't happen to screw up the "Trash1" program, that program can always fall back on the "Algorithm A" program if all else fails. (someone hid/moved/stole the trash can).
Example1: Solve a known problem when help is available
AI is told, "Fix my leaky roof," with no further instructions.
It must now figure out everything it needs to know about how to get on top of the roof, what tools it will need, and so on. It must figure out where to find the information it needs.
Example2: Solve a problem when no help is available
Essentially, the AI needs the power of "invention", but the only thing I can think of here is a loop and "trial and error". pathetic. is this really what we do, even when we think we aren't??
The components I'm sure it needs are:
Core engine -
this is the core classes and functions that process all of the scripts and other information
AI Algorithm-
As stated above, this is a highly generalized problem solving program which will do several things:
1) Check the Index to see if we already know how to solve this problem
2) Attempt to solve a problem even if the solution is not known so long as the problem is the most important thing needs doing at the moment. (basicly this would involve a hell of a lot of nested loops, comparisons, and recursions, and gathering of information.)
3) write a less generalized algorithm for solving this problem solved by step 2 so the entire problem solver doesnt' need to be called every time we want to do a task where the general solution is known. Index this with description of the problem, which the AI got either from its own initiative(as output from another of its own programs) or as input from another AI or human being or other source of information.
4) from time to time re-test 3 to be sure it is the best option. If 3 is not the best option, create a better version, but keep the old one also for future reference. This could be relevant in cases of new information becoming available.
5) determine the next task to be performed, which could involve calling 1, 2, 3, or 4 above.
Index -
Used to help the AI understand what each of its programs and data are and how to use them. In a software where you might end up with millions of programs and scripts, obscure names will never do. The AI must be able to "know" what its skills are and how to use them. In a few experiments I found this to be deceptively complicated.
Interpreter -
This is actually part of the Core engine, and the AI will need at least one type of interpreter just to comunicate with humans, but will also potentially need other interpreters as data architecture gets very complicated. It actually needs to be able to interpret its own index and algorithms as well.
There are probably other things, but these are all software components that "Mr. Data" would need. One person on a backwater planet cannot write this program in their lifetime, unless I am severely over complicating the problem and there is an extraordinarily simple and elegant algorithm nobody has ever considered.
----
So don't worry, you'll never have to worry about 2 versions of an intelligent operating system conspiring against you.
DavidD
26th May 2008 - 02:25 PM
Call me sceptic, but Thinking AI, which can understand like human, will not ever be created, becouse of complexity of organism and brain. At least for brain need 10^15 synapses - bits. Such amount of data don't exist in supercomputers - only. But To create each synapse need at least 100 times more bits, thus need more than 100 supercomputers. And to interconnect those synapses properly need even more bits, so need about milion top supercomputer or about 10^20 bits on all hard drives. Forget it. Try to create somthing like snail or bee or frog, but seems, that even it too hard, to analyze all brain os milions neurons and bilions sinapses. Need the most important to create very got algorithm, which is scalabale for more neurons and synapses interconection and then it connect with input and output data. In nature such algorithm don't exist, but is created with nature like natural selection - genes.