Peter Voss wrote:
> 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; Unfortunately, this is usually hard to do, because most existing applications are so poorly architected that it's hard to make them interact usefully with innovative new software. But where possible, it's a good idea, agreed. > b) having some very powerful adaptive (real-time, incremental) learning > algorithms to give an edge; No question there! Proto-AGI-based applications can certainly outperform narrow-AI-based applications in some areas. The problem is that in many market niches, "performance" (in the sense of intelligent performance) doesn't have all that much business value... Finding a market niche where greater software IQ leads to real competitive advantage, is not as easy as one might think... > c) selecting aspects of applications that do not require 100% accuracy, > reliability, predictability. Agree also. > 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. I agree that datamining is far from AGI. And I agree that AGI has to involve all the things you mention. However, if you have a sufficiently well-formulated conceptual & software design, it can be possible to build datamining products that use *parts* of your AGI system in non-AGI-ish ways. And this can be a way to get parts of your AGI system built, and to some extent tested. There are definitely many aspects of an AGI that are not required for datamining apps, and can't meaningfully be developed in that context. However, AGI does involve * integration of multiple types of information * integration of multiple learning algorithms and it seems that this sort of integration, taken outside the AGI experiential-learning context, is VERY VALUABLE in many practical datamining areas (including bioinformatics). In Novamente, it happens to be possible to use parts of the system in ways that make use of the system's ability for integrating multiple information types and learning algorithms, but do NOT make use of the system's intended ability for experiential learning, perception-action-cognition coordination, etc. This is how we're building datamining apps. Regarding "distraction", I think that nearly ANY practical commercial application is going to end up being a significant distraction. Let's say you got a contract to build an intelligent mobile robot for the Navy, in conjunction with a robotics hardware maker. Great stuff! But most likely, at some point, the Navy's particular requirements are going to end up differing from what you want to do for AGI. They'll need some specific functionality working thus well by thus and thus date, and to achieve it with a high degree of certainty, you'll need to hack something together using narrow-AI methods.... etc. etc. The reality is, commercial development is not R&D, and will always involve some "distraction" from R&D, though it can also be a learning experience in many different ways. > 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). Mobile robotics sounds like the best application area for you, actually. The sensors and actuators generally suck, so low data resolution rates are the norm anyway... And the competition really isn't much good at all. The main customer is the government. I imagine if you found someone with robotics experience and got them to collaborate with you, you could get some DARPA funding for A2I2 based mobile robot control... > 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. Yep -- it's easier -- but won't necessarily yield as much intelligence. The key then is to find market niches where more intelligence is worth more money. All the niches I listed in my prior post, have this property to varying degrees in various specialized ways. -- Ben G ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/
