I would like to create an initial feasibility test, using a text-based IO, that would show the potential for intelligence across a broad range of subject matters (within that IO modality.) I am not worrying about writing something that would be scalable to adult human level AGI. I believe that there has been something missing in AI/AGI. Someone needs to show how one might create a good base for intelligently acquiring knowledge (ie using both rational and creative methods) which might be scaled up with some future computers system. The sense that narrow AI can be pushed beyond human capabilities in certain human games that once seemed to demand higher general reasoning is a little unexpected and hard to understand without concluding that there must be some very basic AGI ideas that haven't been discovered. I am having problems just working on the program and another fundamental problem is that once a program gets a little beyond the mundane world of contemporary programming it can bog down quickly in complexity. In one of the few AI experiments that I actually tried I found that a reasonable plan for an analytical algorithm crunched to a stop just because there was recursive complexity that I did not easily see. Even after becoming aware of it I had a difficult job getting around it. And I wasn't trying to make the algorithm exhaustive. But, as a result, I now know that recursive complexity is serious problem that is lurking behind any problem that calls for some analysis or for broad searching. Relying on a well established method that has been hammered out across a number of decades can help you achieve a sense of sophistication about the problem, but if that method is clearly not good enough then it is obvious that more insights into the problem are key to making some progress. I believe that if you study the problem seriously, come up with some creative ideas and run enough experiments you may solve some real problems even if you can't solve them all. From the one trivial experiment that I learned that being too careful is a form of carelessness in AGI. This is a surprising result. But, thinking about that result I recognize that the solution to the problem of initial learning is obvious. So from one experiment I learned something valuable and I am ready to take the next step. Jim Bromer From: [email protected] Date: Tue, 16 Jul 2013 18:02:29 +0200 Subject: Re: [agi] A Very Simple AGI Project To: [email protected]
On Tue, Jul 16, 2013 at 4:52 AM, Jim Bromer <[email protected]> wrote: a simple initial feasibility test may be designed around this format as a means of designing a way for a computer program to learn direction from a human user so that it can further discover interesting ways to acquire structured knowledge Jim, We have nothing against intentions to experiment, even though we slightly prefer the results of experiments over the intentions - not that I am not guilty of the same crime, mind you. Now, scalability is the number one risk of wannabe intelligent systems, and you are not going to escape the problem by avoiding the discussion or the formalization. I will kindly remind you that learning systems abound in the 50 year old history of the field, even though I believe the right mix of ambition and resources was not there 99% of the time. Now, what kind of learning you want to do? In my humble opinion it does not get any more simple than template learning or Bayesian learning. In the first you are more in "canned response" territory, in that you can save/remember your entire "lifestream" of inputs and the choose "rewarding" outputs based on similarity metrics (nearest neighbor, whatever) between your actual input and previously rewarding output. In Bayesian learning, which is not at all uncommon biologically despite taking mathematicians thousands of years to formulate it, you probably can account for a slightly more dynamic world (like one in which "I am hungry" only gets you food half of the time or once a day). I guess the Bayesian world would be driven by random number generators that will regularly break away from canned responses. Both "solutions" are mathematically multidimensional, or rather dimensionally cursed, if the problem domain is not a toy then the implied mathematical objects are pretty enormous. Do I think any such model can scale to human-level intelligence? Not really. Do I think it can provide a lifetime of hobbyist entertainment? I sure do. Is there learning of a third kind? I doubt it. I will briefly restate my hunch; human-level intelligence will need a lot of real world statistics that more generally enable loads of heuristics, all at the service of real world simulations, possibly agent-based ones, that are scrutinized quickly and result in appropriate action. AT AGI | Archives | Modify Your Subscription ------------------------------------------- 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
