[I started this thread on the SL4 list - decided that it belongs here. My original question & Ben's reply are at the bottom.]
I agree with Ben about the difficulty of developing products based on (early) AGI: In most cases you will have to do all of the engineering (and research and marketing) of a conventional application *plus* developing & integrating your AGI engine. Once we develop practical applications, I hope to mitigate this difficulty by: a) applying the AGI engine as a layer on top of existing applications; b) having some very powerful adaptive (real-time, incremental) learning algorithms to give an edge; c) selecting aspects of applications that do not require 100% accuracy, reliability, predictability. The various data mining (text, genomics/ proteomics, data etc.) applications suggested are far from what our AGI project is all about - and from what I think is important for general intelligence. I know that Ben disagrees (to some extent). I believe that crucial AGI abilities center around online, interactive selection-perception-action learning (of static and temporal knowledge and skills). It needs to combine unsupervised, self-supervised and supervised methods. Large statistical number crunching, language, and symbolic logic are distractions in my opinion - especially if data is preselected, essentially static, and batch processed. Our early applications are likely to be at the level of animal (say, dog) cognition, but initially of course at extremely low data/ resolution rates (for both perception and action/ dexterity). Notwithstanding this animal reference, I do think that adaptively learning a PC user's habits (by monitoring mouse, screen, etc.), plus being taught specific 'tricks' could be an early AGI application. The (difficult) trick is to chose applications that leverage inherent strengths of artificial systems without shifting focus from core AGI requirements. The other problem, that I know that Ben is well aware of, is that for *any given* application it is almost always easier to hard-code or custom engineer a solution, rather than using general intelligence abilities and letting the system learn. Peter --------- Ben's post from SL4 ------------ > I'm interested in any and all potential early applications for AGI - both to evaluate the performance of our a2i2 system, and for possible implementation. > > Any ideas? > > Peter First, I have a general observation about early AGI applications. Then I'll try to answer your question.. An advanced AGI will be able to observe a new application area, and figure out how to apply itself to that area, and connect itself to other appropriate software programs, etc. Until we're at that stage, setting up a concrete application of a proto-AGI software system is a *lot of work*. There are two factors here: 1) setting up any software application is a lot of work 2) Creating an AGI-based application can actually be more work than just setting up a narrow-AI-based software application, even not counting the work involved in creating the AGI system itself. The reason for 2 is as follows. AGI systems are created for autonomy and flexibility and nondeterminism, whereas in a software application context, one often needs different virtues instead: repeatable efficient behavior in particular contexts. There is of course no intrinsic reason a software system can't have both AGI virtues and software application virtues... but in practice, early-stage AGI systems often aren't created with software application virtues in mind. We specifically architected Novamente so that it could support both the autonomy & flexibility & nondeterminism required for AGI, AND also so that one could create highly constrained & efficient Novamente-based software apps. But not all AGI's will be this way; Webmind AI Engine, for instance, did not have this property at all, and so building practical apps on top of it was like pulling teeth (and we wound up primarily pulling objects out of it to use in narrow-AI apps, rather than actually using Webmind AI Engine in practical apps). So, my experience is, it's possible to make a simple prototype application of an AGI system in a particular area by having a strong programmer work on it for 6 months or so. But to actually make an AGI-based product that can be sold or used in some domain, is a huge amount of work, even more than building an analogous product without AGI involved. Another bit of general wisdom, which you surely know already: The most important thing is to pick an app area where you have a lot of domain expertise.... Now, having gotten that blather out of the way, what are good app areas for AGI systems? We're now working in bioinformatics. There are loads of subfields here... we're working mostly on genomics & proteomics data analysis. At Webmind Inc., we worked in computational finance, and information retrieval (document categorization, document retrieval). Big uses for AGI here.... And in the info. retrieval case, one can often outperform existing apps without any real NLP in one's system, just treating documents as patterned character-sequences ... Gaming has been mentioned by Reason; I think gaming is an OK application if one has a very fast AGI or if one has a gaming idea & biz model that supports running game-interacting AGI's on a server farm... System administration is a wide open market... in desperate need of smart automation tools... Knowledge management is a huge market. So many diverse DB's out there, with incompatible DB schemas, waiting for a system to come along and reconcile them ... Robot control is another big area. Basic robot arm control issues are handled by standard engineering methods. But no one knows how to deal with mobile robotics yet, hardly at all.... The problem with mobile robotics is decision making based on data integration. See Albus's work at NIST, on a robot-controlled tank. Very interesting. this area is begging for AGI. Of course, the military applications may be disturbing... Data mining in general is in great need of AGI. It is hard to make a LOT of money as a datamining company, but plenty of small companies seem to survive in this way, often via a handful of contracts with large firms. The biomedical area incorporates aspects of robot control automation, datamining, and bioinformatics, knowledge management, and system administration.... And you get to deal with a huge bureaucracy too ;) Well, I could go on further, but there's work to do ;) If anyone wants to explore one of these topics further, I'll be happy to do so. Or, Eliezer, if you feel this thread wanders too far from SL4's intended focus, we could move this to the AGI list. -- Ben ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/