Stanley, Thank you for asking me these questions! I have thought about them. I want to respond over a few brief replies. The baby does not learn language by trying every possibility and a baby does need good sources to act as a guide in learning. (I mean I agree with those views.) But, how do you get that in a computer program? One of the most significant models of knowledge, (really the only one that I can imagine as viable) is a component model of knowledge. These component parts can be used in 'generating' different outputs. So we do not need to learn a separate name and symbol for every number, we can use 'components' to both represent them and to name them up to our limits of knowledge. But we only need to remember maybe a hundred different labels and how to attach them and we can thereby represent and talk about (samples) of octodecillions of numbers (if you can find the right page on Wikipedia to get the names of large numbers). But, you still have to follow certain rules when using one of those numbers. These rules are themselves expressed in the n-ary system as compressions that are convenient to use with many large numbers. (It is not convenient for us to do calculations on numbers in the octodecillion range but it is convenient for a computer to do calculations on numbers that large.)
Some methods, like neural networks tend to combine the components 'of knowledge' but the successes in the field seem to reflect somewhat narrow AI applications. In other words the more interesting uses in AI are when they are treated more like component methods. My argument against them is that important knowledge (knowledge-stuff) is lost when you methodically trash or mush the component representations of that knowledge. However, there is no evidence -as of yet- that general knowledge can be so conveniently broken down to a convenient number of components as it has been done with the n-ary representations of numbers. So even though our would be AGI programs would use component models of knowledge it still would have to be able to go through heaps of stuff in order to find the best responses for a given situation. In a simple world model an AGI program could do deep searches without any problem. But as the world model becomes more and more nuanced the deep searches would *themselves* become nuanced just because they too are based on component models. The problem is that their are no conveniently compressed component methods that can be used on general AI. Imagination gives us the ability to think outside the box. It is at the least an ingredient of the 'magic sauce'. Once you begin to think deeply about reinforcement methods that are operating on an imaginative AGI program you will find that the simplicity of the method would start failing rather quickly at the least sign of conceptual complexity. Jim Bromer On Wed, Dec 2, 2015 at 11:35 AM, Stanley Nilsen <[email protected]> wrote: > Greetings Jim, > Have some questions and comments about the recent email... > (thread used to be: Scholarpedia article on AGI published) > > On generating possibilities... > > Complexity seems to be the issue that lies underneath much of your > efforts. And, there is no doubt that life has it's complexity. I don't > recall (didn't search past emails) any explanation from you, that ties this > complexity to an AGI implementation. > > I personally have an issue with complexity and that is because it is > over-rated. In many writings there is an implication that AGI type > intelligence will emerge through the construction of a unit that will try > lots of things and get feedback and thereby learn what creates the best > reward. Sort of an evolutionary intelligence. Do we really believe that? > I don't, do you? > > I've seen reference to "search space" and the implication is that the > intelligent unit is going to find an efficient way to navigate this huge > space of possible "stuff." Am I off track in assuming that this may be the > connection with trying to discover the p=np solution? Thus enabling the > traversing this search space more efficiently. Is this how you see a better > intelligence created? > > Perhaps a simple example will suffice to see my contrasting view of the > development of intelligence. Consider a new born baby, or imagine you > create a computer baby that has a sound generating capability. One could > program the computer to generate random sounds and then watch for feedback > at the keyboard or through a listening device. Good luck with that. There > are several ways one might alter the experiment and produce suggestive > results. But, lets save time because we know that a human intelligence, the > baby, does not "simply" try lots of sounds and thereby become a speaking > person. Instead, the baby learns by lots of feedback and, importantly, by > mimicking the sounds that it hears. What does this prove? > > First, it shows that the baby isn't developing an intelligence by randomly > trying every combination. Rather, the baby is "adopting" behaviors that > produce positive feedback from the adult who is nurturing the baby. Enough > about babies... > > The computer intelligence will not evolve from nothing, far from it. It will > need a core set of behaviors to work with. "Thousands of different > approaches" could refer to the thousands of "cores" that are eventually > developed by different researchers. The core will be a big factor in how > the intelligent unit goes about acquiring behaviors to add to it's core. > And, as you might suppose, the unit will be influenced by it's surroundings, > the core isn't everything. > > To net it out... It isn't the "learning" by trial and error that is > significant, it is the adoption process that dwarfs the "scientific" > gathering of truth. The baby AGI doesn't need lots of experiments, it needs > good sources to guide it's adoption process. > > Stan > > On 12/02/2015 07:36 AM, Jim Bromer wrote: >> >> ... >> >> It was useful to me only because it showed me how I might generate >> variations on the abstractions underlying some purpose. So my methodology >> might be used in an actual AI program to generate possibilities, or more >> precisely, generate the basis's for the possibilities. It would not stand as >> a basis for AI itself. Solomonoff's methods do not stand as a basis for AI. >> (Years from now, when there are thousands of different approaches to AI and >> AGI we might still disagree about this.) But it should be clear that I am >> saying that Solomonoff's universal prediction methods are not actually >> standards for viable AI. Compression methods are not a viable basis for AI >> either. They could be used as part of the process but it is pretty far >> fetched to say that any computational technique that could be used in an AI >> or AGI program could be used as a basis for AI. >> I just wanted to follow through since you wrote a reasonable reply to my >> first reply. >> >> > > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/24379807-653794b5 > 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
