Hi Eric, According to my belief, that I also claim to have published a strong case
for, we have such a theory, in which the common principle underlying intelligence is that of Occam programs, that are computationally hard to extract.
(I don't mean a program in the Occam language, a program
constructed according to an extrapolated Occam's razor.) Also according to this belief, "understanding" is comprised of having such an Occam program that exploits underlying structure in order to generalize. According to this belief, unfortunately, the Occam program underlying our intelligence is itself unlikely to admit any more compact Occam program understanding it, and thus may be inherently not understandable.
I don't quite agree with this perspective, though my view is pretty close. I also find it useful to view understanding in terms of algorithmic information, but I think that "finding the shortest program capable of computing X" is not a good way of conceptualizing "understanding X." Rather, I think that "finding the fuzzy set of programs capable of compressing X, relative to one's knowledge base K" is a better perspective. For a complex X, there will be many different programs capable of compressing X. Just finding the shortest program for computing X does not necessarily give a complete understanding of X. Ben's comments, and to some extent his approach to AGI of building
code and then hoping when run it will produce a complex set of patterns that do stuff seems somewhat related to this,
Yes, it's related.... The Novamente system can be viewed as attempting to find a bunch of compressing programs in relevant datasets, esp. in datasets of the form "carrying out action A in context C will lead to achievement of goal G." This is not necessarily the most useful way to view the system in practice, but it is a correct way. And of course, in accordance with the "no free lunch theorem", the idea is that it should be good at finding compressing programs in datasets of the above form that actually occur in the practical life of an embodied agent, not in mathematically general datasets of the above form. except for
some reason he stipulates that human intelligence is understandable. I'm not clear on why he thinks human level intelligence is "understandable", or even what he means by this.
I've tried to clarify this above. What I mean is that humans are able to detect many meaningful patterns (read: patterns = compressing programs, if you like) in human behaviors ... and once brain scans work better, I bet we will be able to detect many very meaningful, significant patterns emergent btw human behaviors and the output of brain scanners... OTOH, for a massively superhuman AI, the quantity of patterns we will be able to detect in this way, may be far far less ---- because most of the significant patterns in its behavior and state may have an algorithmic information far beyond the capacity of our brains. -- Ben G ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
