On Sat, Dec 20, 2008 at 8:01 AM, Derek Zahn <derekz...@msn.com> wrote:
> Ben: > > > Right. My intuition is that we don't need to simulate the dynamics > > of fluids, powders and the like in our virtual world to make it adequate > > for teaching AGIs humanlike, human-level AGI. But this could be > > wrong. > > I suppose it depends on what kids actually learn when making cakes, > skipping rocks, and making a mess with play-dough. Some might say that if > they get conservation of mass and newton's law then they skipped all the > useless stuff! > OK, but those "some" probably don't include any preschool teachers or educational theorists. That hypothesis is completely at odds with my own intuition from having raised 3 kids and spent probably hundreds of hours helping out in daycare centers, preschools, kindergartens, etc. Apart from naive physics, which is rather well-demonstrated not to be derived in the human mind/brain from basic physical principles, there is a lot of learning about planning, scheduling, building, cooperating ... basically, all the stuff mentioned in our AGI Preschool paper. Yes, you can just take a "robo-Cyc" type approach and try to abstract, on your own, what is learned from preschool activities and code it into the AI: code in Newton's laws, axiomatic naive physics, planning algorithms, etc. My strong prediction is you'll get a brittle AI system that can at best be tuned into adequate functionality in some rather narrow contexts. > > But in the case where we are trying to roughly follow stages of human > development with goals of producing human-like linguistic and reasoning > capabilities, I very much fear that any significant simplification of the > universe will provide an insufficient basis for the large sensory concept > set underlying language and analogical reasoning (both gross and fine). > Literally, I think you're throwing the baby out with the bathwater. But, as > you say, this could be wrong. > Sure... that can't be disproven right now, of course. We plan to expand the paper into a journal paper where we argue against this obvious objection more carefully -- basically arguing why the virtual-world setting provides enough detail to support the learning of the critical cognitive subcomponents of human intelligence. But, as with anything in AGI, even the best-reasoned paper can't convince a skeptic. > > > It's really the only critique I have of the AGI preschool idea, which I do > like because we can all relate to it very easily. At any rate, if it turns > out to be a valid criticism the symptom will be that an insufficiently rich > set of concepts will develop to support the range of capabilities needed and > at that point the simulations can be adjusted to be more complete and > realistic and provide more human sensory modalities. I guess it will be > disappointing if building an adequate virtual world turns out to be as > difficult and expensive as building high quality robots -- but at least it's > easier to clean up after cake-baking. > Well, it's completely obvious to me, based on my knowledge of virtual worlds and robotics, that building a high quality virtual world is orders of magnitude easier than making a workable humanoid robot. *So* much $$ has been spent on humanoid robotics before, by large, rich and competent companies, and they still suck. It's just a very hard problem, with a lot of very hard subproblems, and it will take a while to get worked through. On the other hand, making a virtual world such as I envision, is more than a spare-time project, but not more than the project of making a single high-quality video game. It's something that any one of these big Japanese companies could do with a tiny fraction of their robotics budgets. The issue is a lack of perceived cool value and a lack of motivation. Ben ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com