Like most here I am mostly perplexed by this manifesto. I could agree with Babiano that the choice of words here and there makes it sound more pseudoscientific than scientific, but as a "practitioner" I am mostly concerned with the missing implementation details. The manifesto started with the idea that complexity shall be defeated, but I don't see how or where, what I do see is that everything will be just too bloody fluid and based on unknown (infinite?) numbers of experiments that sound complex enough.
I think in part 6 we see how not relying on embodiment may come back and bite you. You may know the joke about the hungry slave who goes to his cruel master requesting food, and all he gets is glass of water after glass of water, and the question "are you ready to eat something". When the slave after x glasses of water answers "No" the master quips "You see, you were not hungry, you were thirsty all along". For any embodied entity food has to mean food and sex has to mean sex (plugging in the wall socket etc), or it will be the very end of the entity, so I would not welcome any shifts to a range of concepts. If such a set of independent, "grounded" concepts is fixed then we probably enter a Hutterian universe where agents try to maximize their grounded and fixed utility functions with evangelical zeal. However we can still guarantee "individuality" emerging from this simultaneous pursuit of many objectives, as it is a kind of "many-body problem" with extreme sensitivity to initial conditions and therefore chaotic behavior, so even if we think alike I may end up collecting water while you end up collecting food. The elephant in the room is anyway how you are going to implement concepts, you mention something about data structures you have in mind, but in my mind they should be the first to get out of your mind, if you are not out of your mind :) . In particular, if you reject the above proposed "basic insticts" how are you going to bootstrap they concept graph? Which will be the first concept? As has been suggested a lot in the last 60 years, a concept could easily be a program, but then it may not so easily receive those influences you want to see happening. For a tremendous amount of real world data an object hierarchy or ontology is a great concept implementation, you may know that men and women exist, but not about the emo tribe and the Nanguza tribe, then you can assume things about the Nanguza women but perhaps not for the Nanguza rabbit whisperer. Instead, would you prefer shifting the "human" concept left and right to accomodate emo girls and Nanguza rabbit whisperers? I wouldn't. Please try to keep the manifesto tighter (it includes some speculative fluff here and there) and branch out into implementation ASAP AT On Tue, Apr 23, 2013 at 5:04 PM, Jim Bromer <[email protected]> wrote: > Logan said: > > By doing some programming, you'll gain some insights into how computers > think. > Also you'll learn about how to think more logically and rationally. > > I hope so. > > Logan said: > generally you need to write and interpreter or compiler, to "understand" > i.e. compile or interpret a statement. > > You need to write something that will "understand" or interpret statements > but the question is how do you do that so that it actually works. My theory > is that it takes many statements to "understand" one statement. Some of > the statements may refer to incidental associated information and some may > refer to information about usage and so on. Furthermore, you need > contextual information about an ongoing conversation and what some of the > consequences of interpreting a sentence in a certain way may be. It is not > a straightforward problem. Anaphora-like relations, for example, can > change the meaning of an apparent object of an indefinite article which > means that a sub-sentence which is exactly the same can refer to a > broad range of a-kind-of-object in one sentence and to a very particular > object in another. It takes many statements, some of which will refer to > how linguistic objects are typically used, to "understand" a single simple > statement. > > I had said: > So this means that it can be very difficult to determine the meaning of a > combination of concepts if the program does not explicitly contain a > reference to that particular combination. > > Logan replied: > That is completely false, it's like saying computer-programming languages > contain references to every particular combination, when in fact you only > need to understand the sub-components. Similar to how you don't need to > know every story in conceivability to listen to a new story and derive > meaning from it. In fact the very process of acquiring new information > falsifies your hypothesis. > > The mysteries of the capabilities of human intelligence do not > automatically falsify hypotheses about the problems of artificial > intelligence on a computer. That is a serious logical error in reasoning. > You cannot transcend the boundaries of two very distinct reference subjects > without recognizing that the argument from one does not necessarily hold > for the other. (In some discussions that would be ok but there is no reason > to believe that my reference to "the program" referred to human abilities.) > > > I agree that the ability to ask questions and search through external > sources of information would allow the program to redirect its search and > help it to avoid search complexities in some cases. > > The simplistic use of generalizations in the 60's did not work to produce > AI, and different kinds of weighted reasoning in the 70s and the 80s did > not work either. Weighted Reasoning can refer to a number of different > paradigms. Putting weights on statements is one kind (John Anderson), > Neural Networks is another and Bayesian Networks is another. > > The simulation I plan to start with will use a constrained language of > 100-200 words. I will start by explicitly directing the program > (algorithmically) to produce the kinds of data structures that I have in > mind for the program, then I will see if I can write the subprograms which > could use those kinds of data relations to determine what an input sentence > is referring to. I will start with something simple and if I make some > progress then I will try something a little more difficult. I plan to > learn a great deal from this process and I expect that my theories about > AGI will become stronger. > > Jim Bromer > > > ------------------------------ > Date: Tue, 23 Apr 2013 07:13:11 -0400 > > Subject: Re: [agi] Summary of My Current Theory For an AGI Program. > From: [email protected] > To: [email protected] > On Mon, Apr 22, 2013 at 4:13 PM, Jim Bromer <[email protected]> wrote: > > Logan,**** > Thanks for your comments. > > I agree of course that concepts and concept integration may be represented > by words and sentences. I was trying to say that many of the > complications that will arise using word-concepts will arise using some > other kinds of referential concepts. One of the reasons that I am > convinced that text-only AGI is a good way to go is because there is such a > potential for expressiveness and the representation of different kinds of > ideas. It is often difficult to express complicated ideas using words > because they are not substitutes for the implementations of the things that > we are talking about. > > > We can implement anything using words, from programs through bridges to > relationships. > > However, that does not mean that they cannot be used as representations of > ideas. I understand what I am talking about even though other people do > not.**** > ** > > > That simply indicates a need to enhance your communication skills. > > ** > I believe that when we acquire a learned habit the parts of the habit may > not be directly understandable but can only be approached indirectly by > referring to something else. For instance a learned action may be > created by a string of action potentials (for a lack of a better name) and > it may be that the only way to detect the parts of the string is by noting > the whole, more complicated action. > > > Ya, many voice to text parsers work like that, however they aren't very > good at understanding new phrases, or different ways of saying things. > Both are necessary, likely with some supervised learning i.e. "what did you > mean by that?" giving a target, for optimal results. > > > Or we may infer the action by some other action or other event that is > roughly correlated with the inferred action. But essentially, when we > are capable of reflection (meta-cognition) we are able to ‘understand’ a > concept potential if we know something more about how to use and integrate > the concept. So by having some kind of understanding of a concept > potential we can consciously try to use it in different ways based on some > kind of reasoning. Now, if are not explicitly aware of the concept > potential there may be a chance that we can infer something about it > indirectly just as we might infer something about an action potential.**** > ** ** > I believe that the theory that it takes many statements to understand one > simple statement has a great deal of value. > > > generally you need to write and interpreter or compiler, to "understand" > i.e. compile or interpret a statement. > > Concepts are relativistic. That means that when a simple concept is used > in association with other concepts the meaning of the concept can vary. > Concepts > are contextual. But there are more problems. Concepts are > interdependent. There is not (necessarily) an independent concept and a > dependent concept in a conceptual function the way there are in a > mathematical function. > > > Actually there are a whole host of such axiomatic concepts. If there > weren't we'd just be wishy washy not really saying anything all the time. > > > So this means that it can be very difficult to determine the meaning of a > combination of concepts if the program does not explicitly contain a > reference to that particular combination. > > > That is completely false, it's like saying computer-programming languages > contain references to every particular combination, when in fact you only > need to understand the sub-components. Similar to how you don't need to > know every story in conceivability to listen to a new story and derive > meaning from it. In fact the very process of acquiring new information > falsifies your hypothesis. > > One way to work with this problem is to rely on generalization systems > in which the systems of generalizations of a collection of concepts can be > used to guide in the decoding of a particular string of concepts which > haven’t been seen before. However, when this was tried in the simplistic > fashion of the discrete text based programs of the 60’s it did not > produce intelligence. > > can you give some examples? > Cause C, fortran and a host of other discrete text string concepts > happened and seem to have produced significant intelligence, i.e. beating > world chess champions among a multitude of other achievements. > > So in the 70’s weighted reasoning became all the rage because it looked > like it might be used to infer subtle differences in the strings that > simple discretion substitution did not. However, this promise did not > hold up either. > > > Those are neuro-nets I infer, and they are merely one statistical tool, > in an arsenal of learning. Multiple forms of learning, in combination with > strong core for knowledge representation is necessary to achieve general > intelligence. > > Neither system have, in themselves, proven sufficient to resolve the > problem. My feeling is that the recognition that it takes many > references to a concept to ‘understand’ that concept is part of the key to > resolving these problems without hoping to rely on a method that suffers > from combinatorial complexity. > > > programming languages and operating systems don't suffer from > combinatorial complexity, or if they do, it is well managed, yet they are > the most generally intelligent things/thought-systems on computers. > > > Another part of the key is to recognize that concept objects may contain > numerous lateral similarities to other concept objects and that these > similarities may run across the dominant categories of a concept object > that is being examined. > > > Jim Bromer > > > > > On Mon, Apr 22, 2013 at 6:01 PM, Jim Bromer <[email protected]> wrote: > > > I think just skimmed through the outline html -- it seems like a good > > start. I wouldn't start writing any code for quite a while yet. It > > seems to me that you need to fight with those issues first. > > > Thanks for the friendly comment, but I am going to push myself to > start coding (experimenting) next month. > > > Great! the sooner the better. > > > Formal methods have to be tried and shaped based on extensive applications > of the methods to real world problems. > > > By doing some programming, you'll gain some insights into how computers > think. > Also you'll learn about how to think more logically and rationally. > > I am thinking of starting with simple simulations to see if I can > eventually find some formal methods (programmable methods) that can work > with the kinds of problems that I will throw at it. > > > What would you be simulating? > > > If I don't make any progress with that then I might try creating a > language which is designed to be extensible via generalizations. > Jim Bromer > > > Didn't you say generalizations failed in the 60's? > Did you know, that much like people, > programming languages, are extensible, > through the use of libraries .i.e. books of information. > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/14050631-7d925eb1> | > Modify<https://www.listbox.com/member/?&;>Your Subscription > <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
