Hi Ben, I think there are some important points pertaining to the early achievement of AGI here -
Ben> My feeling on dog-level intelligence is that the *cognition* aspects of dog-level intelligence are really easy, but the perception and action components are significantly difficult and subtle. 1.) I thought it was well established that by now that perception & action are *integral* aspects of human and/or dog-level intelligence (as appropriate to AGI). You can dramatically reduce the 'resolution'/ capacity (especially for proof-of-concept prototypes), but AGI without them seems misguided. 2.) If 'dog-level intelligence' is so simple, why has no-one come near to achieving it? I see it as a crucial sub-set of higher-level intelligence (and thus our shared AGI ambitions). Ben> In other words, once a dog's brain has produced abstract patterns not tied to particular environmental stimuli, the stuff it does with these patterns is probably not all that fancy. But the dog's brain is really good at recognizing and enacting complex patterns, and doing this recognizing & enacting in a coordinated way. Producing abstract patterns from stimuli and being 'really good at recognizing and enacting complex patterns' is a core AGI requirement - the basis for all higher-level ability. Ben> Peter Voss's (www.adaptiveai.com) approach to AI aims to emulate biological evolution on Earth, in the sense that it wants to start with a dog-level brain (very roughly speaking) and then incrementally build more cognition on top of this. This is a reasonable approach, to be sure. I would not characterize our approach as 'emulating biological evolution'. I believe that roughly dog-level intelligence is the right level to aim at, because it includes much of the fundamental cognition needed for AGI while eliminating the 'distractions' of language, abstract thinking, and formal logic. (I see these as inappropriate problems to focus on at this stage - especially if they *not* perception based. The cart before the horse - for numerous reasons.) Ben> But if I had to make a guess, I'd say this approach should probably begin with robotics, with real sensors and actuators and a system embodied in a real physical environment. I am skeptical that simplistic simulated worlds provide enough richness to support development of robust dog-level intelligence... as perception and action oriented as dog intelligence is... I don't see any problem with using virtual environments for testing & proving basic abilities - one can get an enormous amount of complexity out of them these days. But in any case, our framework (and actual testing) seamlessly integrates virtual and real-world perception/ action. Peter ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
