David Noziglia wrote: > First, if the idea is to build computer systems that can do the > same things > that "natural" intelligent systems - people - can do, then the > term "General > Intelligence" is not a true description of either the goal or the > methodology.
To me, "generality" of intelligence is a fuzzy, not binary criterion. Humans clearly have more generality in their intelligence than Deep Blue or Mathematica.... There may be other minds out there somewhere, with intelligence vastly more general than ours. An awful lot of human intelligence is specialized for things like scene understanding and social interaction, but yet we're nowhere near as specialized as "narrow AI" programs like the ones I just mentioned... > The idea that human intelligence is contained or > described by > a single factor - Burt's and Spearman's g - is pretty much a dead letter, > despite trash political manifestos disguised as science like The > Bell Curve. > We no are pretty sure that human intelligence is built out of lots of > different "tools" to accomplish specific tasks like perception, > comprehension, motivation, and calculation. Whether there is a single number to measure *degree* of intelligence, is a different question than whether it's meaningful to speak about the *generality* of a system's intelligence. There are many ways to measure degree of intelligence, and also many ways to formally define generality of intelligence if one goes that route. I continue to think that the generality of a system's intelligence is a meaningful concept. However I stress that it's an intuitive concept, which could be mathematically formalized and empirically measured in a lot of different ways. > So the members of this group are > each in their own way working on ways of modeling intelligent > tasks, and the > work done on these goals is significant and valuable. I have done (and am doing) work on task-specific "narrow AI", and it's very different from the work I'm doing on AGI. The difference is that in narrow AI, one is trying to create programs that can solve some specific human-defined problems. In AGI, we're trying to create computer programs that can solve a wide variety of problems, and can solve new types of problems without having humans explicitly formulate the new problems for them, or modify them to suit the new problems. To achieve this kind of "general intelligence", I believe, one needs to create a system that integrates perception, action, cognition, memory, learning of declarative & procedural knowledge, and a host of other things. I believe one needs a self-organizing system that interacts richly with an environment that contains a variety of interconnected "problems" and "tasks." In practice, work on building this kind of system is VERY DIFFERENT from work on building narrow-AI systems oriented toward particular tasks. This is so, even though there is some overlap in terms of software tools, algorithms and data structures between the two pursuits. > But that's not what you're aiming at. You're also hoping that at > some point > all your task-specific intelligent tools will get hooked together into an > entity that can coordinate all their work, and that can then turn its > electron-fast analysis into a self-analyzing, creative loop that generates > something that passes the Turing Test. It is not the case that we're building a host of task-specific tools and hooking them together. In the Novamente project, we have a coherent overall design for an AGI, and we're coding and testing it -- step by step. It happens that some of the parts of this in-development system have narrow-AI uses which can be commercialized. This is fortunate in terms of pragmatic funding issues, but not so pertinent to the AGI goal itself. > I would suggest, however - and here I know I'm going way too far > - that the > term you have chosen (AGI) for this meta-project has been deliberately > selected to make the enterprise sound scientific and legitimate. That is somewhat true. The term was formulated (or at least, introduced to me) by Shane Legg, when we were looking for a title for our (forthcoming) edited volume on AGI. My proposed title "Real AI", was deemed too confrontational, and AGI was suggested as something that wouldn't ruffle so many feathers. I don't think there's anything wrong with choosing a relatively non-confrontational name. The ambitious and speculative nature of the research is not being hidden from anyone. But if choosing a milder name helps funding sources and journal editors to accept our work as serious, then to my mind it's a worthwhile thing. [At age 35, my idealism has limits, and I'd prefer to focus it on really important things like keeping the AI design itself pure ;) ] > What you > are actually after - and what the arguments are really about - is > something > quite different. > > Building the tools is possible. Coordinating them is possible. But the > next step, which you have mislabeled AGI, is not. > > Because what you are really after should be called not Artificial > Intelligence, but Artificial Consciousness. My own opinion is that when AGI is created, Artificial Consciousness will come along for free. I am a panpsychist; I think that everything, right down to particles, is "conscious" a little bit. Some things are more conscious than others. The apparent partial correlation btw degree of general intelligence and degree of consciousness is a very interesting thing. I have done a lot of deep thinking about this, and written some things about it, e.g. in my rough-draft online manuscript "Unification of Science and Spirit" [which is old, and doesn't fully represent my current views; some bits of it are rather embarrassing to me now, actually.] But I don't think we need to crack that deep philosophical puzzle in order to create an AGI.... To make an even stronger statement: I think we can create a conscious machine without fully understanding consciousness. > That is the key characteristic of being human. It is not > something that can > be built through a reductionist construction of finite Turing Machine > programs. I understand that you, Paul Pruiett, Roger Penrose, Stuart Hameroff, and some other smart, deep-thinking people hold this belief. In fact, I once held that same belief too. But I changed my mind. None of the arguments you or other believers in the non-algorithmic creed have put forth, seem very convincing to me. I have read nearly everything that's been published on this topic, so my view is certainly not ill-informed. This gets into some rather deep philosophical issues... I now view the universe as consisting of three things: patterns, physical reactions, and pure chance. Pure chance could be called a lot of other things: spirit, mystery, elemental randomness, whatever you like.... Charles S. Peirce had a lot to say about this. He called it First, as opposed to Second (the physical universe of reactions), and Third (the mental universe of relationships). Classical physics, quantum physics, quantum gravity and the theory of Turing machines are examples of patterns. So are you and me. So is the letter "a." In Peircean terms, patterns are Third. Consciousness is clearly somehow associated with the pure-chance aspect of being, with Peircean First. Naively, one might think that this intuition about consciousness being associated with chance, is incompatible with the possibility of a conscious computer program. Because computer programs are in theory deterministic, non-chance-based. But reality is a lot deeper than that. I have not yet plumbed all the depths but I have visited a lot of them.... The following comments are speculative but may illustrate the flavor of my thinking in these areas. I cannot in practice predict the behavior of a very complex computer program. I cannot predict the detailed behavior of a Novamente system even now, let alone a Novamente system as it will (hopefully) exist in a few years, with some real AGI going on. When you look at the mathematical definition of "randomness", one observes that the concept is defined *only relevant to an observer*. Now, the theory of CS tells you that for infinitely large entities, the observer-dependence goes away in a sense. But I think the finite definition is the important one. I think that chance is intrinsically a subjective notion. If I can't predict what a Novamente system is going to do, because of basic limitations imposed by the finitude of my brain, then that Novamente system is chance-displaying, to me. I think chance is subjective, and hence, if consciousness is associated with chance, then in a sense whether X is conscious or not may depend on the perspective from which X is being observed.... We then have a Godelian argument that any complex mental system is going to be conscious with respect to itself, its own subjective point of view. Because no really complex system can fully predict itself.... This conclusion fits in naturally with the panpsychic attitude I mentioned above. I do not expect these brief philosophical musings to be convincing to anyone. I give them here mostly just to point out that this is something I've thought about deeply, something I think that is important. If I disagree with you, Paul, Penrose etc., it is not because I haven't reflected deeply on these matters. I don't dismiss consciousness glibly a la Daniel Dennett. I know it's a very real and deep phenomenon, but I don't believe that it's related to quantum physics or quantum gravity in the way that Penrose/Hameroff propose. I think that a future physics theory will clarify the relationship between physical reality and consciousness, but I suspect that this clarification will explain why digital computer programs CAN be conscious, and not the opposite. > I can be as wildly speculative as the next person, and with a lot less > real-world scholarship to base that on. I can say that every truly > significant step in macro-evolutionary history has been brought about by > symbiosis. Along these lines, you should check out the book Symbiogenesis by Werner J. Schwemmler. http://www.amazon.de/exec/obidos/ASIN/0899255892/qid%3D1033565584/sr%3D1-19/ ref%3Dsr%5F1%5F1%5F19/028-4347167-6373313 He puts forth a lot of really interesting speculative theories along these lines. My review of his later book "Basic Cancer Programs" touches on some of these themes as well: www.goertzel.org/papers/CancerPrograms.htm Finally, you say: > That linear binary > programs just may not be capable of developing emergent overlays > upon which > autonomous autopoietic entities can appear. It's a side point, but I don't understand where "linearity" comes in here. Digital computer programs generally represent nonlinear systems, actually.... Or are you referring to the quantum level, where infinite-dimensional linear mappings are used; and contrasting this with quantum gravity theories which mostly involve nonlinear aspects a la general relativity theory? -- Ben G ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/
