Jim: what would constitute a real empirical test ? The [2nd] basic principle of AGI testing is v. simple - and a particular test doesn't have to be defined, though suggestions like I and Benjamin made are always helpful.
The principle is: does this robot have "generalizabilty"? Can it automatically generalize whatever capacity it has been designed with? Crudely: can it "take off"? So if you have a robot that is focussed to begin with on nothing else but handling - a handling/manipulative robot - then it's AGI if it can automatically go on to handle an endless diversity of objects without any additional programming. If it starts by handling small rocks, then it should automatically be able to grasp bricks, bottles, small pyramids, ropes etc and whatever surprise objects are presented to it, (within reasonable boundaries). As with humans and infants, this will be by a process of trial and error, which may include failures but will include sucess after success. Ditto if you have a robot that can locomote on one terrain, then it's AGI if it can automatically go on to handle new kinds of terrain - if it starts with stony ground, it should be able to go on to, say, rocky ground, grassy ground, sandy ground etc. waterbeds - an endless range of new terrains. The same principle would apply "in theory" to a language AGI - if it can talk about navigating one terrain, can it go on to discuss an endless range of new terrains? I say, "in theory" here because the idea of a language AGI in any foreseeable future is farcical - and anyone contemplating it hasn't got much of a clue about the conceptual nature of language. The endless generalization of a faculty and particular activity is what distinguishes humans and animals - we do go on to handle an endless range of new objects and navigate an endless range of new terrains -.. and talk to an endless range of new personalities with new philosophies, attitudes, vocabularies, accents etc. Our capacity to do this is the basis of our acquiring new skills/activities., Our capacity to handle ever new objects, for example, is basic to handle ever new rackets/bats and successively learn tennis/table tennis/baseball/cricket/hockey et al This basic principle is, I think, not something that anyone here could or would argue with. Obviously an AGI must have generalizability. But I doubt whether a single project is aiming directly/immediately for a *testable* version of it. I can virtually guarantee that Ben and Boris et al aren't. The 1st principle of AGI testing is also simple and is inseparable from the 2nd - but will be more controversial. It is creativity. AN AGI must be able to create a given course of action WITHOUT having been specifically programmed for it. It must be able to handle new object after new object, new terrain after new terrain WITHOUT any programming for those specific objects. So you should be able to tell your AGI in one form or other - "pick up that object" - and it will both design and effect the necessary course of action, with no human programming input. This again is absolutely fundamental to how all humans and animals pursue courses of action - we can take "briefs"/brief instructions and flesh out the appropriate course of action. It is also fundamental to Ben's "dog fetch ball" test of old. (As I said, Ben's first intuitions are often good ones. In reality, a dog who fetches a ball always has to create the necessary course of action in a somewhat unfamiliar field. But the actual version of a dog fetching a ball implemented by Ben had nothing to do with AGI). Generalizability and creativity (creating a course of action without specific programming) - those are the fundamental,intertwined, **clearly. testable ** principles of AGI. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
