My description of a simple AGI subprogram vs a simple Narrow AI subprogram and a non-AI subprogram is interesting. It can be used to demonstrate a progression of AGI learning which I called judgment-mediated learning. It can be modeled in the most artificial ways which is important for preliminary feasibility tests and to create control backgrounds where new ideas may be introduced carefully.
My definition of a simple AGI subprogram (a judgment-mediated learning environment) naturally supports my own ideas about AGI. But remember, I have not actually written a program which uses my theories in anyway, so this idea of a simple AGI subprogram does not merely validate my own theories in a trivial way. Many people have not believed that my ideas are feasible and they find that my discussions about things like this are either trivial or without merit. Some people have thought that I am just talking about things that they have already thought about or that I am talking about something that will always remain a fantasy. I have always thought that my ideas are just based on common sense but I do not know if they are feasible or not. One problem I have run into is that the complications of programming something like what I am thinking about make the project impossible. So the focus on making the initial program as simple as I can get without getting stuck in Narrow AI or Non-AI is absolutely necessary. Right now my theory of Conceptual Typing is extremely complicated. But there are two points that I can make in defense of the possible feasibility of the idea. One is that when a typing scheme for a comment (or system of concepts) is chosen or sufficiently narrowed, the resulting characteristics are more like a simpler formal computational language. The other is that the system is not open to every kind of variation, it is only open to some relations and although the number of possible combinations of relations might be huge, my expectation is that only a few will be reasonable once the greater complex of relations appropriate for a particular case are taken into account. (It takes a great deal of information related to a subject to really know one simple thing about it). The problem that there is always a possibility for something new to be learned makes it more difficult, but there are a number of variations on the implementation (of my theories) which would make it more versatile. For instance it might be tested, based on what it had previously learned or it might be tested with a limitation on the number or types of new information that it might be learned during a training session. (I expect that an automated constraint on the possible variations that might be considered relative to the analysis within some familiar kind of situation is a fundamental method of learning.) So although there are a few big unsolved problems in my conjecture, the essential features of this theory could be developed according to some constraints which would make it simpler to work with. Part of the simplicity of the model is its versatility. -Jim Bromer On Thu, Aug 15, 2013 at 5:30 PM, Jim Bromer <[email protected]> wrote: > The Turing Test is not good enough to tell us if a particular AGI > program is feasible. We also need to find some way to simplify the > idea of AGI in order to start somewhere. A particular model of AGI > sub-programs may not turn out to be a good indicator of potential > scalability but I feel that I have to try to devise some kind of test > and then test the test out as long as near-human-like AGI is out of > reach. > > There are some things that a human being can't figure out, so AGI is > not going to be perfectly general. But an AGI program also has to be > able to deal with a great variety of situations similar to the way > human beings can. > > And an AGI program also has to be able to come up with some pretty > good insights about a variety of problems or situations. > > The idea of insight extends beyond a narrow solution to variations of > one kind of problem, but it also extends beyond the narrow idea of one > kind of response. So not only does an AGI method have to give > appropriate responses to a variety of problems it also has to be able > to fit that analysis and response into a greater variety of relations. > (This may go without saying, but my point is that it leads to a model > for more fundamental AGI methods.) An AGI program (or sub-program) > has to have some kind of meta awareness. A method of meta awareness > does not have to be (in itself) complicated. This ability to have > sufficient awareness to connect an analysis and appropriate response > into other knowledge constitutes a kind of judgment. And I have > mentioned other qualifications that I think are seen in most human > judgment. > > The feeling that once someone figures it out it will be as obvious as > the Turing Test is not valid. There is no reason for an AGI program > to act human without a lot of additional programming. So tying some > kinds of sub-programs with simple tests for judgment relative to > various domains makes perfect sense. The test won't be tied to one > domain but will be tried with different domains. Although there is > evidence that general knowledge is necessary for specialized > knowledge, my theory which states that a great many relations related > to a subject are necessary to know one simple thing about the subject > does not preclude this condition from occurring in tests for domain > knowledge. - Jim Bromer > > > On Wed, Aug 14, 2013 at 3:06 PM, Mike Archbold <[email protected]> wrote: >> I think I got the gist of what you are saying here... It looks >> interesting; maybe just a couple of comments. >> >> I am inclined to think that it depends upon perspective as far as if >> we consider some component narrow- vs strong-AI. I think the >> distinction is helpful, but not necessarily at some level such as >> function vs. subfunction vs. program vs. entire system etc. It >> strikes me that it is the behavior of the program, software, entity, >> as a *whole* that is in the determining factor of narrow- vs strong-. >> It is hard to imagine as subroutine that has some specific task as >> being general. Perhaps I missed the point above (?). The >> distinguishing characteristic of narrow-ai seems to be that it 1) has >> nothing like a common sense understanding of human consensus reality >> and 2) can only function in a well described domain and often gets >> tangled up on cases which veer to far from the domain. If it isn't >> that, it's strong/general AI (ie., we don't have it ;) >> >> The German philosophers that I've studied distinguish sharply between >> judgement and understanding. This is a good distrinction. The >> hallmark of a judgement is that it could be other than what it is. >> Understanding we've talked about on this list a lot. They both happen >> at once, of course.... >> >> conceptual typing... can you point to a thread where you explain? >> >> Mike >> >> On 8/12/13, Jim Bromer <[email protected]> wrote: >>> My idea of judgment does rely on reason based reasoning. This >>> definition would seem to favor explicit representation. However, I do >>> recognize that we make some decisions that are not based on explicit >>> reasons so I do include implicit or hidden reasons in my definition of >>> judgment. And the fact that we can use poor judgment or poor reasons >>> for making a decision does seem to weaken the theory. But by making >>> AGI learning partially dependent on previous judgment-mediated >>> learning, the idea does hold together even if it cannot be pinned down >>> to an absolute computational definition. >>> >>> How does this idea of judgment-mediated-learning tie into a definition >>> of an AGI function that can be differentiated from a Narrow AI >>> function? My idea of conceptual typing, dynamic creative and rational >>> creative functions and trial and error methods can be combined to >>> explain how novel conceptual typing might be developed as the program >>> is running. So that means that I have an explicit way of dynamically >>> developing new ways of looking at the data as the program is running. >>> In most contemporary AGI models this is not detailed. So, artificial >>> judgment can examine the presumptions behind the conceptual structures >>> that are running as well as develop results that are dependent on >>> them. - Jim Bromer >>> >>> >>> ------------------------------------------- >>> AGI >>> Archives: https://www.listbox.com/member/archive/303/=now >>> RSS Feed: https://www.listbox.com/member/archive/rss/303/11943661-d9279dae >>> Modify Your Subscription: >>> https://www.listbox.com/member/?&; >>> Powered by Listbox: http://www.listbox.com >>> >> >> >> ------------------------------------------- >> AGI >> Archives: https://www.listbox.com/member/archive/303/=now >> RSS Feed: https://www.listbox.com/member/archive/rss/303/24379807-f5817f28 >> Modify Your Subscription: https://www.listbox.com/member/?&; >> Powered by Listbox: http://www.listbox.com ------------------------------------------- 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
