Ben, This is GREAT. We are finally distilling some of the most important "details"...
On Tue, Sep 4, 2012 at 3:25 PM, Ben Goertzel <[email protected]> wrote: > >> Via theoretical understanding of the type of system one is building... > > > > > > But, for emergent properties, this is equivalent to having done the math > > to do this, which we have previously agreed is a BIG job probably > requiring > > years of work at a research institute. > >>> > > No, I think it will be possible to cause the needed emergent properties to > arise > via building a system based on a combination of mathematical and > qualitative > theory, I don't think there was much of value presently in existence - certainly not enough to provide significant AGI design guidance. It would seem that out of all of the somewhat reasonable ways that things might work, that we are in an EXTREMELY tiny part of the search space. If not, highly intelligent animals would have been around >100 million years ago. and then letting it self-adapt via embodied social experience > If GI were a significant fraction of the prospective search space, I would agree with you. However, the evidence is quite the contrary. It took ~150 million years of evolution AFTER the dinosaurs to evolve significant GI, during which time MANY other things evolved. Given the apparent value of GI, this tells me that our level of GI is an incredibly tiny fraction of the prospective search space - FAR too tiny to ever find in a lifetime. Probably about the size of a grain of sand, somewhere on the Earth. Given our presently poor level of "expertise" combined with our presently short lifetimes, I just can't imagine anyone stumbling onto such a tiny part of the search space, at least not anyone that **I** know. Were you thinking of anyone in particular? How would YOU go about looking for such a tiny needle in such a gigantic haystack? > > Having a full math theory would be great, but I don't think it's going > to be necessary > I see math as a gradually converging process as things get verified via simulation and wet lab. Somewhere in this process we will pass human-level GI, but I doubt whether there will EVER be a "full math theory" that goes as far as can be gone, with one intriguing possibility... You are probably familiar with how random waveforms form the basis of spread spectrum communication that is MUCH more power-efficient than other forms of electronic communication? You are probably familiar with how random selection is essential in solving game theory problems? These basic realizations have allowed their respective fields to "jump to hyperspace" in one quick jump for each field. There may be some similar/parallel realization connected with GI, that someone can leverage as Einstein leveraged the Lorentz transformations into his Theory of Relativity, to jump right over the extreme complexities that have gone into our own brains, like the ~200 different types of neurons. To demonstrate a parallel, looking back on prior modulation techniques once spread spectrum has been well understood, it becomes obvious that maximizing complexity is the key, and you can't get any more complex than a random bit stream. Without a random bit stream, improvements had involved ever increasing complexity, much as our central nervous system has. It would be an EXTREMELY interesting conversation to try to identify such a theory. The search would probably start by presuming its existence, and then trying to prove some paradox based on its existence. The flaws found in the "proof" would point the way to identifying the theory. However, lacking such an all-encompassing theory, I just can't see you finding SUCH a tiny sweet spot in SUCH a large search space, at least not until long after you have been uploaded/downloaded. Oops, we need to solve these problems BEFORE we can upload/download you, so that won't work. Hence, if I had bet my own career on AGI (your problem, not mine), I would be actively hosting this conversation in the hopes of figuring out enough to bracket it in enough for some clever mathematician to finally realize what it is. Oh, isn't that we are doing here on this forum?!!! Not exactly, but close. The difference between what I see happening here and what appears to be needed, is that present AGIers seem to be trying to develop incremental improvements in methodology, where a radically revolutionary replacement is probably needed, to avoid having to find that tiny little spot in the very large search space. I have been looking at some prospective approaches for cracking this particular nut. Applying the Scientific Method by starting to list all apparent possibilities (I think this list is quite finite): 1. Perhaps all "logic" can be viewed as a "process control problem", and perhaps there is some math that is fundamental to self-adaptive process control beyond PID control, etc. If so, I would think that the ultimate process control system MUST necessarily be intelligent? This makes sense to me, but so far I haven't found any recruits to this line of thinking. 2. Perhaps all "logic" can be viewed as game theory problems. It isn't that we may have "an" adversary, but we live in a world with a vast hierarchy of competing adversaries. Hence, rather than AND and OR logic, perhaps the fundamental unit of computation is the solution to small game theory problems (that much like conventional logic also provides probabilistic results)? Von Neumann (also the father of the stored program computer) and Morganstern (who together first authored *The Theory of Games and Economic Behavior*) would just LOVE this theory. 3. How about some more theories? If there is a possible solution to save Ben's career, and if we are going to find it, then we absolutely MUST include some sort of description of it in order to identify it. Otherwise, Ben and OpenCog will be about as well remembered as the Perceptron. Then, we can start throwing stones at these candidates, until hopefully one survives the onslaught and gives Ben what he so desperately needs. So, start throwing... Steve ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
