These specifications are not feasible and therefore not reasonable. I have a feeling that if IBM or Google could've the would've. So more intermediate goals are necessary. Furthermore, many of my specifications would be more like design specifications. I would want to see the program "integrate" ideas in both a creative method and in a structural method (structural reasoning is not opposite of creative reasoning.) So I would need to see conceptual integration occurring in some novel ways - or more toward the specification side of the goal - in ways that use reason-based reasoning in novel ways. (Non-novel reasoning includes logic, probability reasoning and neural networks.) I would want to see some attainment of modest levels (very modest levels) of conceptual complexity. Getting stuck on a low-level local hollows doesn't cut it.
When you are experimenting with new ideas (or you hope you will be trying some new ideas) you cannot specify what you will find before hand. However, your design specifications are absolutely essential to holding yourself to accept the results of your experiments. Without some reasonable specifications it becomes much more easy to delude yourself into thinking that you are making progress even though it is not going just as smoothly as you had hoped for. So I would also add specifications (including design specifications) that represent a failure to achieve some minimum goals. For instance, I would need to see that my AGI program was using reasons to integrate pieces of knowledge in an effective way. There are many ways here where I might be fooled by 'logical' similarities. Since my relational referential network would be designed to do some stuff that neural networks do, only better, then what might look like a reason-based integration of concepts might turn out only to be a learned network association similar to the way neural networks work. So the question then becomes even if this (limited) goal works how can I be sure that it did work. It might turn out that a question like that can only be answered by cross-experimental examination of the effects. For instance I might try a traditional neural network, take a look at how that works and then compare it to my relational referential network to see just how different the results are. And I might then try to use those differences to see, for example, if I can induce a virtual decomposition of the synthesis in a way that a traditional neural network can't do very easily. (I probably wouldn't actually try a traditional neural network out if I got that far, I am just saying that might be one approach to cross-examine what looks like a successful achievement of a limited design goal.) Jim Bromer On Tue, Sep 4, 2012 at 11:49 AM, Matt Mahoney <[email protected]> wrote: > On Tue, Sep 4, 2012 at 11:04 AM, Mike Tintner <[email protected]> > wrote: >> Can you be ahem more specific about your ideal specification? > > A specification describes what a program should do. For AGI, we > generally mean it should be able to do anything that a human could do. > More practically, it would be anything that you might pay a human to > do (because these are goals that investors would be willing to fund). > > A more detailed specification might list specific, intermediate goals > or tests, such as: > 1. Ability to fill in missing words in text with human level accuracy. > 2. Ability to pass college level final exams in most subjects. > 3. Ability to match captions to images. > 4. Ability to classify videos as funny or not. > 5. If equipped with a camera and robot arm, ability to catch and throw a ball. > >> Do any AGI projects meet your criteria? > > I believe all of these problems will eventually be solved. However I > don't believe that anyone on this list has the resources to do it > themselves. > > > -- Matt Mahoney, [email protected] > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/10561250-164650b2 > 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-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
