RE: Funding AI
Tim May wrote: Except I'll add that I don't agree physics is stumped by most complex systems. Physics doesn't try to explain messy and grungy situations, nor should it. Turbulence is a special case, and I expect progress will be made, especially using math (which is why Navier-Stokes issues are on the same list with other math problems for the prize money). I guess this comes down to the semantics of the word physics. If you want to define physics to exclude all complex systems besides turbulent fluids, that's your right. But I don't like your definition, personally. What about protein folding? What about potential quantum effects in water macromolecules in the brain? What about bioelectromagnetic fields, as studied by Russian researchers extensively over the last 50 years? Etc. etc. etc. I feel like you're taking a whole lot of things that contemporary physics can't deal with because of its conceptual shortcomings, and classifying them as not physics, in order to make physics look more successful than it is. Of course, physics has been dramatically successful in some areas, but let's not overlook its weaknesses. Quantum gravity is not the only area it's tried and failed to touch. And I have a suspicion that the same mathematical/conceptual breakthrough that allows complex systems to be rigorously studied, will also help with the quantum gravity problem. This suspicion is in line with the intuition of plenty of smart physicists, including John Wheeler with his whole It From Bit concept (which portrays physical law itself as the result of complex self-organization). Funding is the key issue. Someday I'll write a thing for this list about successes vs. failures in terms of auto-funding each successive stage of a complex technological path. In a nutshell, the electronics/computer industry was essentially self-funding for the past 50 years, with the products of 1962, for example, paying for the work that led to the 1965 products. Same thing with aviation. ... It is unlikely that the path to AI will be successful if there are not numerous intermediate successes and ways to make a _lot_ of money. My tip to all AI workers is to look for those things. (This is more than just banal advice about try to make money, I am hoping. I have seen too many tech enthusiasts clamoring for moon shots to fund what they think is needed...)) The AGI may come from the distant great-great grandchild of financial AI systems. I've worked on financial AI systems myself. It's hard to argue with a statement as loose as distant great-great grandchild ... but I think financial AI is not on the shortest path to AGI right now. I agree that it's important for incremental progress toward AGI to be financially viable, and in fact my own current work involves building toward real AGI, partly via building bioinformatics applications (which ARE financially viable in the short run). However, I also would point out that AGI research is different from many other kinds of research, in that the primary research tools are very inexpensively available. All you need are computers. yeah, you may need a shitload of RAM, but it's a very different situation from other sciences: you don't need a cyclotron, a chip fabrication plant, a microarrayer, a PCR machine, etc. etc. There is real potential for real progress to be made on a shoestring. Funding is valuable, but less critical than in a lot of other areas. Fundamental physics has the same inexpensiveness, to an extent -- a theoretical breakthrough could be made by a guy sitting alone in his attic unfunded. But to verify the currently fashionable theories requires insanely expensive equipment, which is a real obstacle to progress -- a type of obstacle that's not nearly so severe in AI right now. ben
RE: Applied vs. Theoretical
Tim May wrote: As I hope I had made clear in some of my earlier posts on this, mostly this past summer, I'm not making any grandiose claims for category theory and topos theory as being the sine qua non for understanding the nature of reality. Rather, they are things I heard about a decade or so ago and didn't look into at the time; now that I have, I am finding them fascinating. Some engineering/programming efforts already make good use of the notions [see next paragraph] and some quantum cosmologists believe topos theory is the best framework for partial truths. The lambda calculus is identical in form to cartesian closed categories, program refinement forms a Heyting lattice and algebra, much work on the fundamentals of computation by Dana Scott, Solovay, Martin Hyland, and others is centered around this area, etc. FWIW, I studied category theory carefully years ago, and studied topos theory a little... and my view is that they are both very unlikely to do more than serve as a general conceptual guide for any useful undertaking. (Where by useful undertaking I include any practical software project, or any physics theory hoping to make empirical predictions). My complaint is that these branches of math are very, very shallow, in spite of their extreme abstractness. There are no deep theorems there. There are no surprises. There are abstract structures that may help to guide thought, but the theory doesn't tell you much besides the fact that these structures exist and have some agreeable properties. The universe is a lot deeper than that Division algebras like quaternions and octonions are not shallow in this sense; nor are the complex numbers, or linear operators on Hilbert space Anyway, I'm just giving one mathematician's intuitive reaction to these branches of math and their possible applicability in the TOE domain. They *may* be applicable but if so, only for setting the stage... and what the main actors will be, we don't have any idea... -- Ben Goertzel
Re: Applied vs. Theoretical
From Osher Doctorow [EMAIL PROTECTED], Sunday Dec. 1, 2002 1243 Sorry for keeping prior messages in their entirety in my replies. Let us consider the decision of category theory to use functors and morphisms under composition and objects and commuting diagrams as their fundamentals. Because of the functor-operator-linear transformation and similar properties, composition and its matrix analog multiplication automatically take precedence over anything else, and of course so-called matrix division when inverses are defined - that is to say, matrix inversion and multiplication. It was an airtight argument, it was foolproof by all that preceded it from the time of the so-called Founding Fathers in mathematics and physics, and it was wrong - well, wrong in a competitive sense with addition-subtraction rather than multiplication-division. There is, of course, nothing really wrong with different models, and at some future time maybe the multiplication-division model will yield more fruit than the addition-subtracton models. And, of course, each model uses the other model secondarily to some extent - nobody excludes subtraction from the usual categories or multiplication from the subtractive models. What do I mean when I say it was relatively wrong, then, in the above sense [question-mark]. Consider the following subtraction-addition results - in fact, subtraction period. 1. Discriminates the most important Lukaciewicz and Rational Pavelka fuzzy multivalued logics from the other types which are divisive or identity in their implications. 2. Discriminates the most important Rare Event Type [RET] or Logic-Based Probability [LBP] which describes the expansion-contraction of the universe as a whole, expansion of radiation from a source, biological growth, contraction of galaxies, etc., from Bayesian and Independent Probability-Statistics which are divisive/identity function/multiplicative. 3. Discriminates the proximity function across geometry-topology from the distance-function/metric, noting that the proximity function is enormously easier to use and results in simple expressions. It sounds or reads nice, but the so-called topper or punch line to the story is that ALL THREE subtractive items above have the form f[x, y] = 1 plus y - x. ALL THREE alternative division-multiplication forms have the form f[x, y] = y/x or y or xy. Category theory has ABSOLUTELY NOTHING to say about all this. So where are division and multiplication mainly used [question mark]. It turns out that they are used in medium to zero [probable] influence situations, while subtraction is used in high to very high influence situations. Come to your own conclusions, so to speak. Osher Doctorow - Original Message - From: Tim May [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Sunday, December 01, 2002 10:44 AM Subject: Applied vs. Theoretical
Re: Everything need a little more than 0 information
Hal Finney wrote: That would be true IF you include descriptions that are infinitely long. Then the set of all descriptions would be of cardinality c. If your definition of a description implies that each one must be finite, then the set of all of them would have cardinality aleph-zero. What Russell wrote was that the set of all descriptions could be computed in c time on an ordinary Universal Turing Machine. My question is, does it make sense to speak of a machine computing for c steps; it seems like asking for the cth integer. The descriptions in the Schmidhuber ensemble are infinite in length. At this stage, I see no problem in talking about machines computing c steps, but obviously others (such as Schmidguber) I know would disagree. Its like asking for the cth real number, rather than the cth integer, if you like. I'm not sure what the connection is with this non-standard model of computation and others such as Malament-Hogarth machines (sp?) Cheers A/Prof Russell Standish Director High Performance Computing Support Unit, Phone 9385 6967, 8308 3119 (mobile) UNSW SYDNEY 2052 Fax 9385 6965, 0425 253119 () Australia[EMAIL PROTECTED] Room 2075, Red Centrehttp://parallel.hpc.unsw.edu.au/rks International prefix +612, Interstate prefix 02