It has been about 10 years since I completed my PhD in AI. I have been busy building decision support systems for industry - in building, predictive models and customer decision support systems for research institutions here in Australia, within Oracle Corporation and recently to support a start-up I set up with technology that analyses customer health insurance needs and then makes a buying recommendation, a company called iSelect. I have kept half an eye and ear on what is happening in the AI field and had been building an 'opinion' that AI has disappeared from general view because it is increasingly becoming embedded in software applications and because customers/users in general have started to expect or demand a certain level of intelligent behaviour from the systems they interact with. I have only recently noticed the use of the term AGI and could guess as to what it meant and the reasons behind it - I have always told non -AI friends that it is much more difficult to build an AI system that understands jokes or walks than it is to build a system that solves a deep but very specialised problem. I therefore am in agreement that at some hierarchical level of complexity AGI sits at a higher level than the narrow, focussed AI.
However and here is the but: as part of a roadmap I think a large part of AGI can manifest itself as emergent behaviour that is complex and general if the lower level, more focussed AI systems of this world :
(i) could talk to each other i.e. share data, meaning, intent, semantics - this is analogous to Feigenbaum's "if the books in a library could talk to each other what many new things can we learn";
(ii) could incrementally learn.
I suppose what I am trying to say is that from a practical and a commercial point of view a Roadmap for AGI should have as a key component the ability for AI systems to talk to each other and a standardised language for describing objectives and outcomes rather than focus on building from the ground up the mother of all systems that is ultra-general - I think history and biology has shown us that this is not the most efficient way - the more efficient way is to do it incrementally in small steps to solve small (and maybe but not necessarily) complex problems (that give near term benefits) but in a way that allows interaction among these systems such that at the right moment we will witness emergent behaviour that is far beyond the initial, individual designers' dreams.
To this end let us work on a roadmap that allows the development of such a framework and universality of (artificial) thought - again to turn to bio-mimicry, the resultant intelligence on planet Earth is based on such a universal framework of DNA.
I will be interested in other people's thoughts on this...
best regards, David Urpani (david@urpani.com.au)