[agi] Re: Solomonoff Machines – up close and personal
Hi Ed, So is the real significance of the universal prior, not its probability value given in a given probability space (which seems relatively unimportant, provided is not one or close to zero), but rather the fact that it can model almost any kind of probability space? It just takes a binary string as input. If you can express your problem as one in which a binary string represents what has been observed so far, and a continuation of this string represents what happens next, then Solomonoff induction can deal with it. So you don't have to pick the space. You do however have to take your problem and represent it as binary data and feed it in, just as you do when you put any kind of data into a computer. The power of the universal prior comes from the fact that it takes all computable distributions into account. In a sense it contains all well defined hypotheses about what the structure in the string could be. This is a point that is worth contemplating for awhile. If there is any structure in there and this structure can be described by a program on a computer, even a probabilistic one, then it's already factored into the universal prior and the Solomonoff predictor is already taking it into account. How does the Kolmogorov complexity help deal with this problem? The key thing that Kolmogorov complexity provides is that it assigns a weighting to each hypothesis in the universal prior that is inversely proportional to the complexity of the hypothesis. This means that the Solomonoff predictor respects, in some sense, the principle of Occam's razor. That is, a priori, simpler things are considered more likely than complex ones. ED## ??Shane??, what are the major ways programs are used in a Solomonoff machine? Are they used for generating and matching patterns? Are they used for generating and creating context specific instantiations of behavioral patterns? Keep in mind that Solomonoff induction is not computable. It is not an algorithm. The role that programs play is that they are used to construct the universal prior. Once this is done, the Solomonoff predictor just takes the prior and conditions on the observed string so far to work out the distribution over the next bit. That's all. Lukasz## The programs are generally required to exactly match in AIXI (but not in AIXItl I think). ED## ??Shane??, could you please give us an assist on this one? Is exact matching required? And if so, is this something that could be loosened in a real machine? Exact pattern matching is required in the sense that if a hypothesis says that something cannot happen, and it does, then that hypothesis is effectively discarded. A real machine might have to loosen this, and many other things. Note that nobody I know is trying to build a real AGI machine based on Solomonoff's model. Isn't there a large similarity between a Solomonoff machine that could learn a hierarchy of pattern representing programs and Jeff Hawking's hierarchical learning (as represented in the Serre paper). One could consider the patterns at each level of the higherarchy as sub-routines. The system is designed to increase its representational efficiency by having representational subroutines available for use by multiple different patterns at higher compositional levels. To the extent that a MOSES-type evolutionary system could be set to work making such representations more compact, it would become clear how semi-Solomonoff machines could be made to work in the practical world. In think the point is that if you can do really really good general sequence prediction (via something impractical like Solomonoff induction, or practical like the cortex) then you're a long way towards being able to build a pretty impressive AGI. Some of Hutter's students are interested in the latter. The def of Solomonoff induction on the web and even in Shane Legg's paper Solomonoff induction make it sound like it is merely Bayesian induction, using the picking of priors based on Kolmogorov complexity. Yes, that's all it is. But statements made by Shane and Lukasz appears to imply that a Solomonoff machine uses programming and programming size as a tool for pattern representation, generalization, learning, inference, and more. All these programs are weighted into that universal prior. So I think (but I could well be wrong) I know what that means. Unfortunately I am a little fuzzy about whether NCD would take what information, what-with-what or binding information, or frequency information sufficiently into account to be an optimal measure of similarity. Is this correct? NCD is just a computable approximation. The universal similarity metric (in the Li and Vitanyi book that I cited) gives the pure incomputable version. The pure version basically takes all effective similarity metrics into account when working out how similar two things are. So if you have some concept of similarity that you're
Re: [agi] Re: What best evidence for fast AI?
Excellent post, and I hope that I may come across enough time to give it a more thorough reading. Is it possible that at the moment our working with 'intelligence' is just like flapping in an attempt to fly? It seems like the concept of intelligence is a good way to preserve the nonabsurdity of the future. - Bryan - 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=8660244id_secret=63962021-7b03c5
RE: [agi] What best evidence for fast AI?
Ben said -- the possibility of dramatic, rapid, shocking success in robotics is LOWER than in cognition That's why I tell people the value of manual labor will not be impacted as soon by the AGI revolution as the value of mind labor. Ed Porter -Original Message- From: Benjamin Goertzel [mailto:[EMAIL PROTECTED] Sent: Saturday, November 10, 2007 5:29 PM To: agi@v2.listbox.com Subject: Re: [agi] What best evidence for fast AI? I'm impressed with the certainty of some of the views expressed here, nothing like I get talking to people actually building robots. - Jef Robotics involves a lot of difficulties regarding sensor and actuator mechanics and data-processing. Whether these need to be solved to create AGI is a matter of much contention. Some, like Rodney Brooks, think so. Others, like me, doubt it -- though I think embodiment does have a lot to offer an AGI system, hence my current focus on virtual embodiment... Still, in spite of the hurdles, the solvability of the problems facing humanoid robotics w/in the next few decades seems pretty clear to me --- if sufficient resources are devoted to the problem (and it's not clear they will be). I think that, compared to fundamental progress in AGI cognition, -- our certitude in dramatic robotics progress can be greater, under assumptions of adequate funding -- the possibility of dramatic, rapid, shocking success in robotics is LOWER than in cognition -- 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/? http://v2.listbox.com/member/?; - 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=8660244id_secret=63975170-cc0347
Re: [agi] What best evidence for fast AI?
On 11/11/07, Edward W. Porter [EMAIL PROTECTED] wrote: Ben said -- the possibility of dramatic, rapid, shocking success in robotics is LOWER than in cognition That's why I tell people the value of manual labor will not be impacted as soon by the AGI revolution as the value of mind labor. Both valid points -- emphasizing possibility leading to dramatic, shocking success -- but this does not invalidate the (in my opinion) greater near-term *probability* of accelerating development and practical deployment of robotics and its broad impact on society. Robotics (like all physical technologies) will hit a ceiling defined by intelligence. Machine intelligence surpassing human capabilities in general will be far more dramatic, rapid, and shocking than any previous technology. But we do not yet have a complete, verifiable theory, let alone a practical design. - Jef - 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=8660244id_secret=63984519-51ebc9
Re: [agi] What best evidence for fast AI?
But we do not yet have a complete, verifiable theory, let alone a practical design. - Jef To be more accurate, we don't have a practical design that is commonly accepted in the AGI research community. I believe that I *do* have a practical design for AGI and I am working hard toward getting it implemented. This practical design is based on a theory that is fairly complete, but not easily verifiable using current technology. The verification, it seems, will come via actually getting the AGI built! -- 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=8660244id_secret=63987650-f9a81b
Re: [agi] question about algorithmic search
YKY (Yan King Yin) wrote: I have the intuition that Levin search may not be the most efficient way to search programs, because it operates very differently from human programming. I guess better ways to generate programs can be achieved by imitating human programming -- using techniques such as deductive reasoning and planning. This second method may be faster than Levin-style searching, especially for complex programming problems, yet technically it is still a search algorithm. My questions are: Is deductive-style programming more efficient than Levin-search? If so, why is it faster? YKY Deduction can only be used a very constrained circumstances. In such circumstances, it's exponentially slow (or super-exponentially?) with the number of cases to be handled. I don't know anything about Levin searches, but heuristic searches are much faster at finding things in large search spaces than is deduction, even if deduction can be brought to bear (which is unusual). OTOH, if deduction can be brought to bear, then it is guaranteed to find the most correct solution. Heuristic searches stop with something that's good enough, and rarely will do an exhaustive search. That said, why do you think that people generally operate deductively? This is something that some people have been trained to do with inferior accuracy. I still don't know anything about Levin searches, but people don't search for things deductively except in unusual circumstances, so that it's not deductive is not saying that it doesn't do things the way that people do. (I think that people do a kind of pattern matching...possibly several different kinds. Actually, I think that even when people are doing something that can be mapped onto the rules of deduction, what they're actually doing is matching against learned patterns.) One reason that computers are so much better than people at logic is that that's what they were built to do. People weren't and aren't. But whenever one steps outside the bounds of logic and math, computers really start showing how little capability they actually have compared to people. But computers will do what they are told to do with incredible fidelity. (Another part of how they were designed. So they can even emulate heuristic algorithms...slowly. You just don't notice most of what you are doing thinking. Only a small fraction that can easily be serialized (plus a few random snap-shots with low fidelity [lossy compression?]). - 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=8660244id_secret=63993815-f7d737
Re: [agi] Connecting Compatible Mindsets
Bryan Bishop wrote: On Saturday 10 November 2007 14:10, Charles D Hixson wrote: Bryan Bishop wrote: On Saturday 10 November 2007 13:40, Charles D Hixson wrote: OTOH, to make a go of this would require several people willing to dedicate a lot of time consistently over a long duration. A good start might be a few bibliographies. http://liinwww.ira.uka.de/bibliography/ - Bryan Perhaps you could elaborate? I can see how those contributing to the proposed wiki who also had access to a comprehensive mathcomp-sci library might find that useful, but I don't see it as a good way to start. Bibliography + paper archive, then. http://arxiv.org/ (perhaps we need one for AGI) It seems to me that better way would be to put up a few pages with (snip) Yes- that too would be useful. create. For this kind of a wiki reliability is probably crucial, so Or deadly considering the majority of AI reputation comes from I *think* that guy over there, the one in the corner, might be doing something interesting. - Bryan Reputation in *this* context means a numeric score that is attached to the user account at the wiki. How it gets modified is crucial, but it must be seen as fair by the user community. Everybody (except the founders sysAdmins) should start equal. A decent system is to start everyone at 0.1 and have all scores range between (1, 0) .. a doubly open interval. At discrete steps along the way new moderation capabilities should become available. If your score drops much below 0.1, your account becomes deactivated. It seems to me that a good system would increase the score for every article posted and accepted...but it seems dubious that all postings should be considered equal. Perhaps individual pages could be voted on, and that vote used to weigh the delta to the account. There should also be a bonus for continued participation, at even the reader level. Etc. LOTS of details. Also, some systems have proven vulnerable to manipulation via the creation of large numbers of throwaway accounts. This would need to be guarded against. (This is part of the rationale for increased weight for continued *active* participation, at even the reader level. Dormant accounts should not accrue status, and neither should hyperactive accounts.) OTOH, considering the purpose of this wiki, perhaps there should be a section which is open for bots, and in this section hyperactive might well have a very different meaning. If you're planning on implementing this, these are just some ideas to think about. Personally I've never administered a wiki, and don't have access to a reasonable host if I wanted to. Also, I don't know Perl (though I understand that some are written in Python or Ruby). - 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=8660244id_secret=63999495-e194e4
[agi] Re: What best evidence for fast AI?
At 05:48 PM 11/10/2007, Eliezer S. Yudkowsky wrote: The anchor that I start with is my rough estimate of how long whole brain emulation will take, and so I'm most interesting in comparing AGI to that anchor. The fact that people are prone to take these estimate questions as attitude surveys is all the more reason to seek concrete arguments, rather than yet more attitudes. If you want to compare AGI *relative* to whole brain emulation - unanchoring the actual time and hence tossing any pretense of futuristic prophecy out the window - then that's a whole separate story. Well to the extent that I do think we have grounds for rough estimates of emulation dates, comparative estimates for AGI would allow date estimates for AGI as well. I would begin by asking if there was ever, in the whole history of technology, a single case where someone *first* duplicated a desirable effect by emulating biology at a lower level of organization, without understanding the principles of that effect's production from that low level of organization. I know of no important cases, but we do often emulate non-biological systems this way, when they are complex and we mainly care about computed I/O behavior. We record musics and movies, and we port software. We also often reverse-engineer physical devices by copying complex designs we don't fully understand. Organizations also often copy procedures from other organizations they don't understand. I agree that a lack of biological examples should give us pause, but have we ever really wanted to reproduce the I/O behavior of complex biological software before? Looking at history, we find two lessons: 1) Extremely mysterious-seeming desirable natural phenomena are eventually understood and duplicated by engineering; 2) Because they have ceased to be mysterious by the time they are duplicated, humans design them by engineering backward from the desired results, rather than by exactly emulating the lower levels of organization of a black box in Nature whose mysteriousness remains intact even as it is emulated. Cars don't emulate horse biochemistry, sonar doesn't emulate bat biochemistry, compasses don't emulate pigeon biochemistry, suspension bridges don't emulate spider biochemistry, dams don't emulate beaver building techniques, and *certainly* none of these things emulate biology *without understanding why the resulting product works*. But again, these aren't examples of trying to reproduce complex computed I/O behavior. Robin Hanson [EMAIL PROTECTED] http://hanson.gmu.edu Research Associate, Future of Humanity Institute at Oxford University Associate Professor of Economics, George Mason University MSN 1D3, Carow Hall, Fairfax VA 22030- 703-993-2326 FAX: 703-993-2323 This list is sponsored by AGIRI: http://www.agiri.org/emailTo unsubscribe or change your options, please go to:http://v2.listbox.com/member/?member_id=8660244_secret=64006530-809d9a
Re: [agi] What best evidence for fast AI?
Richard, Even Ben Goertzel, in a recent comment, said something to the effect that the only good reason to believe that his model is going to function as advertised is that *when* it is working we will be able to see that it really does work: The above paragraph is a distortion of what I said, and misrepresents my own thoughts and beliefs. I think that, after the Novamente design and the ideas underlying it are carefully studied by a suitably trained individiual, the hypothesis that it will lead to a human-level AI comes to seem plausible. But, there is no solid proof, it's in part a matter of educated intuition. The following quote which you gave is accurate: Ben Goertzel wrote: This practical design is based on a theory that is fairly complete, but not easily verifiable using current technology. The verification, it seems, will come via actually getting the AGI built! This is a million miles short of a declaration that there are no hard problems left in AI. Whether there are hard problems left in AI, conditional on the assumption that the Novamente design is workable, comes down to a question of semantic interpretation. In the completion of the detailed-design and implementation of the Novamente system, there are around a half-dozen research problems on the PhD thesis level to be solved. This means there is some hard thinking left, yet if the Novamente design is correct, it pertains some well-defined and well-delimited technical questions, which seem very likely to be solvable. As an example, there is the task of generalizing the MOSES algorithm (see metacog.org) to handle general programmatic constructs at the nodes of its internal program trees. Of course this is a hard problem, yet it's a well-defined computer science problem which (after a lot of things) doesn't seem likely to be hiding any deep gotchas. But this is research and development -- not pure development -- so one never knows for sure... -- Ben - 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=8660244id_secret=64055433-fe7f04
Re: [agi] What best evidence for fast AI?
Benjamin Goertzel wrote: Richard, Even Ben Goertzel, in a recent comment, said something to the effect that the only good reason to believe that his model is going to function as advertised is that *when* it is working we will be able to see that it really does work: The above paragraph is a distortion of what I said, and misrepresents my own thoughts and beliefs. When pressed, you always resort to a phrase equivalent to the one you give below: I think that, after the Novamente design and the ideas underlying it are carefully studied by a suitably trained individiual, the hypothesis that it will lead to a human-level AI comes to seem plausible When you look carefully at this phrasing, its core is a statement that the best reason to believe that it will work is the *intuition* of someone who studies the design ... and you state that you believe that anyone who is suitably trained, who studies it, will have the same intuition that you do. This is all well and good, but it contains no metric, no new analysis of the outstanding problems that we can all scrutinize and assess. I would consider an appeal to the intuition of suitably trained individuals to be very much less than a good reason to believe that the model is going to function as advertised. Thus: if someone wanted volunteers to fly in their brand-new aircraft design, but all they could do to reassure people that it was going to work were the intuitions of suitably trained individuals, then most rational people would refuse to fly - they would want more than intuitions. In this light, my summary would not be a distortion of your position at all, but only a statement about whether an appeal to intuition counts as a good reason to believe. And, of course, there are some suitably trained individuals who do not share your intuitions, even given the limited access they have to your detailed design. I respect your optimism, and applaud your single-minded commitment to the project: if it is going to work, that is the way to get it done. I certainly wish you luck with it. Richard Loosemore I think that, after the Novamente design and the ideas underlying it are carefully studied by a suitably trained individiual, the hypothesis that it will lead to a human-level AI comes to seem plausible. But, there is no solid proof, it's in part a matter of educated intuition. The following quote which you gave is accurate: Ben Goertzel wrote: This practical design is based on a theory that is fairly complete, but not easily verifiable using current technology. The verification, it seems, will come via actually getting the AGI built! This is a million miles short of a declaration that there are no hard problems left in AI. Whether there are hard problems left in AI, conditional on the assumption that the Novamente design is workable, comes down to a question of semantic interpretation. In the completion of the detailed-design and implementation of the Novamente system, there are around a half-dozen research problems on the PhD thesis level to be solved. This means there is some hard thinking left, yet if the Novamente design is correct, it pertains some well-defined and well-delimited technical questions, which seem very likely to be solvable. As an example, there is the task of generalizing the MOSES algorithm (see metacog.org http://metacog.org) to handle general programmatic constructs at the nodes of its internal program trees. Of course this is a hard problem, yet it's a well-defined computer science problem which (after a lot of things) doesn't seem likely to be hiding any deep gotchas. But this is research and development -- not pure development -- so one never knows for sure... -- Ben 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/?; http://v2.listbox.com/member/?; - 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=8660244id_secret=64076724-00fae4
Re: [agi] What best evidence for fast AI?
Edward W. Porter wrote: Richard, Geortzel claims his planning indicates it is rougly 6 years x 15 excellent, hard-working programmers, or 90 man years to getting his architecture up an running. I assume that will involve a lot of “hard” mental work. By “hard problem” I mean a problem for which we don’t have what seems -- within the Novemente model -- to be a way for handling it at, at least, a roughly human-level. We won’t have proof that the problem is not hard until we actually get the part of the system that deals with that problem up and running successfully. Until then, you have every right to be skeptical. But you also have the right, should you so choose, to open your mind up to the tremendous potential of the Novamente approach. RICHARD What would be the solution of the grounding problem? ED Not hard. As one linguist said “Words are defined by the company they keep”. Kinda like I am guessing Google sets work, but at more different levels in the gen/comp pattern hierarchy and with more cross inferencing between different google-set seeds. The same goes not only for words, but for almost all concepts and sub-concepts. Grounding is made out of a life-time of experience recording such associations and the dynamic reactivation of those associations both in the subconscious and conscious in response to current activations. RICHARD What would be the solution of the problem of autonomous, unsupervised learning of concepts? ED Not hard! Read Novamente (or for a starter my prior summaries of it). That’s one of its main focus. RICHARD Can you find proofs that inference control engines will not show divergent behavior under heavy load (i.e. will they degrade gracefully when forced to provide answers in real time)? ED Not totally clear. Brain level hardware will really help here, but what is six orders of magnitude to the potential of combinatorial explosion in dynamic activations of something as large and high-dimensional as world knowledge?. This issue falls under the getting-it-all-to-work-together-well-automatically heading, which I said is non-trivial. But Novamente directs a lot of attention to this problems, by among other approaches (a) using long and short term importance metrics to guide computational resource allocation, (b) having a deep memory of which computational patterns have proven appropriate in prior similar circumstances, (c) having a gen/comp hierarchy of such prior computational patterns which allows them to be instantiated in a given case in a context appropriate way, and (d) providing powerful inferencing mechanisms that go way beyond those commonly used in most current AIs. I am totally confident we could get something very useful out of the system even if it was not as well tuned as a human brain. There as all sorts of ways you could dampen the potential not only for combinatorial explosion, but also for instability. We probably would start it out with a lot of such damping, but over time give it more freedom to control its own parameters. RICHARD Are there solutions to the problems of flexible, abstract analogy building? Language learning? ED Not hard! A Novamente class machine would be like Hofstadter’s CopyCat on steroids when it comes to making analogies. The gen/comp hierarchy of patterns would not only apply to all the concepts that fall directly within what we think of as NL, but also to the system’s world-knowledge, itself, of which such NL concepts and their contexts would be a part. This includes knowledge about its own life-history, behavior, and the feedback it has received. Thus, it would be fully capable of representing and matching concepts at the level humans do when understanding and communicating with NL. The deep contextual grounding contained within such world knowledge and the ability to make inferences from it in real time would largely solve the hard disambiguation problems in natural language recognition, and allow language generation to be performed rapidly in a way that is appropriate to all the levels of context that humans use when speaking. RICHARD Pragmatics? ED Not hard! Follows from the above answer. Understanding of pragmatics would result from the ability to dynamically generalize from prior similar statements in prior similar contexts, of what those prior contexts contained. RICHARD Ben Goertzel wrote: Goertzel This practical design is based on a theory that is fairly complete, but not easily verifiable using current technology. The verification, it seems, will come via actually getting the AGI built! ED You and Ben are totally correct. None of this will be proven until it has actually been shown to work. But significant pieces of it have already been shown to work. I think Ben believes it will work, as do I, but we both agree it will not be “verifiable” until it actually does.
Re: [agi] What best evidence for fast AI?
Richard, Thus: if someone wanted volunteers to fly in their brand-new aircraft design, but all they could do to reassure people that it was going to work were the intuitions of suitably trained individuals, then most rational people would refuse to fly - they would want more than intuitions. Yeah, sure. I wouldn't trust the Novamente design's AGI potential, at this stage, nearly enough to allow the life of one of my kids to depend on it. But I trust cars and airplanes in this manner every day. Novamente is a promising-looking RD project, not a proven technology; that's obvious. In this light, my summary would not be a distortion of your position at all, but only a statement about whether an appeal to intuition counts as a good reason to believe. Just to be clear: the whole design doesn't have to be taken in one big gulp of mysterious intuition. There are plenty of well-substantiated aspects, substantiated by math or by prototype experiments or functionalities of various system components. But there are some aspects whose ability to deliver the desired functionality is not yet well substantiated, also. And, of course, there are some suitably trained individuals who do not share your intuitions, even given the limited access they have to your detailed design. So far, no one who has taken the time to carefully study the detailed design has come forward and told me I think that ain't gonna work. Varying levels of confidence have been expressed; and most of all, the opinion has been expressed that the design is complicated and even though the whole thing seems to make a lot of sense, there are a heck of a lot of details to be resolved. I respect your optimism, and applaud your single-minded commitment to the project: if it is going to work, that is the way to get it done. I certainly wish you luck with it. Thanks! Ben - 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=8660244id_secret=64085576-1e462a
