Re: [agi] What Must a World Be That a Humanlike Intelligence May Develop In It?
2009/1/9 Ben Goertzel b...@goertzel.org: This is an attempt to articulate a virtual world infrastructure that will be adequate for the development of human-level AGI http://www.goertzel.org/papers/BlocksNBeadsWorld.pdf goertzel.org seems to be down. So I can't refresh my memory of the paper. Most of the paper is taken up by conceptual and requirements issues, but at the end specific world-design proposals are made. This complements my earlier paper on AGI Preschool. It attempts to define what kind of underlying virtual world infrastructure an effective AGI preschool would minimally require. In some ways this question is under defined. It depends what the learning system is like. If it is like a human brain it would need a sufficiently (lawfully) changing world to stimulate its neural plasticity (rain, seasons, new buildings, death of pets, growth of its own body). That is a never ending series of connectible but new situations to push the brain in different directions. Cat's eyes deprived of stimulation go blind, so a brain in an unstimulating environment might fail to develop. So I would say that not only are certain dynamics important but there should also be a large variety of externally presented examples. Consider for example learning electronics, the metaphor of rivers and dams is often used to teach it, but if the only example of fluid dynamics you have come across is a flat pool of beads, then you might not get the metaphor. Similarly a kettle boiling dry might be used to teach about part of the water cycle. There may be lots of other subconscious analogies of these sorts that have to be made when we are young that we don't know about. It would be my worry when implementing a virtual world for AI development. If it is not like a human brain (in this respect), then the question is a lot harder. Also are you expecting the AIs to make tools out of the blocks and beads? Will --- 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=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] What Must a World Be That a Humanlike Intelligence May Develop In It?
2009/1/13 Ben Goertzel b...@goertzel.org: Yes, I'm expecting the AI to make tools from blocks and beads No, i'm not attempting to make a detailed simulation of the human brain/body, just trying to use vaguely humanlike embodiment and high-level mind-architecture together with computer science algorithms, to achieve AGI I wasn't suggesting you were/should. The comment about ones own changing body was simply one of the many examples of things that happen in the world that we have to try and cope with and adjust to, making our brains flexible and leading to development rather than stagnation. As we don't have a formal specification for all the mind agents in opencog it is hard to know how it will actually learn. The question is how humanlike do you have to be for the problem of lack of varied stimulation to lead to developmental problems. If you emphasised that you were going to make the world the AI exist in alive, that is not just play pens for the AI/humans to do things and see results of those things but some sort of consistent ecology, I would be happier. Humans managed to develop fairly well before there was such thing as structured pre-school, the replication of that sort of system seems more important for AI growth, as humans still develop there as well as structured teacher lead pre-school. Since I can now get to the paper some further thoughts. Concepts that would seem hard to form in your world is organic growth and phase changes of materials. Also naive chemistry would seem to be somewhat important (cooking, dissolving materials, burning: these are things that a pre-schooler would come into contact more at home than in structured pre-school). Will --- 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=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] Hypercomputation and AGI
2008/12/29 Ben Goertzel b...@goertzel.org: Hi, I expanded a previous blog entry of mine on hypercomputation and AGI into a conference paper on the topic ... here is a rough draft, on which I'd appreciate commentary from anyone who's knowledgeable on the subject: http://goertzel.org/papers/CognitiveInformaticsHypercomputationPaper.pdf I'm still a bit fuzzy about your argument. So I am going to ask a question to hopefully clarify things somewhat. Couldn't you use similar arguments to say that we can't use science to distinguish between finite state machines and Turing machines? And thus question the usefulness of Turing Machines for science? As if you are talking about a finite data sets these can always be represented by a compressed giant look up table. Will --- 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=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Hypercomputation and AGI
2008/12/30 Ben Goertzel b...@goertzel.org: It seems to come down to the simplicity measure... if you can have simplicity(Turing program P that generates lookup table T) simplicity(compressed lookup table T) then the Turing program P can be considered part of a scientific explanation... Can you clarify what type of language this is in? You mention L-expressions however that is not very clear what that means. lambda expressions I'm guessing. If you start with a language that has infinity built in to its fabric, TMs will be simple, however if you started with a language that only allowed FSM to be specified e.g. regular expressions, you wouldn't be able to simply specify TMs, as you need to represent an infinitely long tape in order to define a TM. Is this analogous to the argument at the end of section 3? It is that bit that is the least clear as far as I am concerned. Will --- 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=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Taleb on Probability
You can read the full essay online here http://www.edge.org/3rd_culture/taleb08/taleb08.1_index.html Will 2008/11/8 Mike Tintner [EMAIL PROTECTED]: REAL LIFE IS NOT A CASINO By Nassim Nicholas Taleb On New Years day I received a a prescient essay from Nassim Taleb, author of The Black Swan, as his response to the 2008 Edge Question: What Have You Change Your Mind About? In Real Life Is Not A Casino, he wrote: I've shown that institutions that are exposed to negative black swans-such as banks and some classes of insurance ventures-have almost never been profitable over long periods. The problem of the illustrative current subprime mortgage mess is not so much that the quants and other pseudo-experts in bank risk-management were wrong about the probabilities (they were) but that they were severely wrong about the different layers of depth of potential negative outcomes. Taleb had changed his mind about his belief in the centrality of probability in life, and advocating that we should express everything in terms of degrees of credence, with unitary probabilities as a special case for total certainties and null for total implausibility. Critical thinking, knowledge, beliefs-everything needed to be probabilized. Until I came to realize, twelve years ago, that I was wrong in this notion that the calculus of probability could be a guide to life and help society. Indeed, it is only in very rare circumstances that probability (by itself) is a guide to decision making. It is a clumsy academic construction, extremely artificial, and nonobservable. Probability is backed out of decisions; it is not a construct to be handled in a stand-alone way in real-life decision making. It has caused harm in many fields. The essay is one of more than one hundred that have been edited for a new book What Have You Changed Your Mind About? (forthcoming, Harper Collins, January 9th). agi | Archives | Modify Your Subscription --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Occam's Razor and its abuse
2008/10/28 Ben Goertzel [EMAIL PROTECTED]: On the other hand, I just want to point out that to get around Hume's complaint you do need to make *some* kind of assumption about the regularity of the world. What kind of assumption of this nature underlies your work on NARS (if any)? Not directed to me, but my take on this interesting question. The initial architecture would have limited assumptions about the world. Then the programming in the architecture would for the bias. Initially the system would divide up the world into the simple (inanimate) and highly complex (animate). Why should the system expect animate things to be complex? Because it applies the intentional stance and thinks that they are optimal problem solvers. Optimal problems solvers in a social environment tend to high complexity, as there is an arms race as to who can predict the others, but not be predicted and exploited by the others. Thinking, there are other things like me out here, when you are a complex entity entails thinking things are complex, even when there might be simpler explanations. E.g. what causes weather. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
On architecture was Re: [agi] On programming languages
2008/10/24 Mark Waser [EMAIL PROTECTED]: But I thought I'd mention that for OpenCog we are planning on a cross-language approach. The core system is C++, for scalability and efficiency reasons, but the MindAgent objects that do the actual AI algorithms should be creatable in various languages, including Scheme or LISP. *nods* As you know, I'm of the opinion that C++ is literally the worst possible choice in this context. However... ROTFL. OpenCog is dead-set on reinventing the wheel while developing the car. They may eventually create a better product for doing so -- but many of us software engineers contend that the car could be more quickly and easily developed without going that far back (while the OpenCog folk contend that the current wheel is insufficient). Perhaps we don't need wheels? Perhaps we need a machine that can retrofit different propulsion systems as an when they are needed We don't seem to be getting anywhere of much with wheeled prototypes, towards generality anyway. (To be clear, the specific wheels in this case are things like memory management, garbage collection, etc. -- all those things that need to be written in C++ and are baked into more modern languages and platforms). I'd go further back and throw out the dumb VMM, at least eventually. Who wants a robot that while it is catching something you threw to it, pauses for half a second due to it having to move information between hard disk and memory? The whole edifice of most operating system/programming language isn't very suited for real time operation. We have real time kernels and systems to deal with that (which .Net is not one of AFAIK). Although to be fair my throwing out the architecture is not based on the real-time system argument, if you have any sort of experimental self-modifying code, you really want an architecture with vastly more nuanced security capabilities so prevent accidents spreading too far. You can go to a POLA architecture like one of the capability security ones (E, keykos), yet they all require a user to manage security rather than allowing systems to control what the code does. In brief my long term road map: 1) VM with security and real time potential 2) High level languages to make use of the features of the languages 3) Write code to solve problems and rewrite code. We are interested in generality of intelligence, we must be prepared to go back to the roots of generality in computing. AI to me has been a series of premature optimisations. People saying, I'm going to create a system to solve problem X, with no thought into how a system that solves X can become one that solves Y. There is always a human in the loop to program the next generation, we need to break that cycle to one where the systems can look after themselves. I can't see a way to retrofit current systems to allow them to try out a new kernel and revert to the previous one if the new one is worse and malicious, without a human having to be involved. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Value of philosophy
2008/10/20 Mike Tintner [EMAIL PROTECTED]: (There is a separate, philosophical discussion, about feasibility in a different sense - the lack of a culture of feasibility, which is perhaps, subconsciously what Ben was also referring to - no one, but no one, in AGI, including Ben, seems willing to expose their AGI ideas and proposals to any kind of feasibility discussion at all - i.e. how can this or that method solve any of the problem of general intelligence? This is because you define GI to be totally about creativity, analogy etc. Now that is part of GI, but no means all. I'm a firm believer in splitting tasks down and people specialising in those tasks, so I am not worrying about creativity at the moment, apart from making sure that any architecture I build doesn't constrain people working on it with the types of creativity they can produce. Many useful advances in computer technology (operating systems, networks including the internet) have come about by not assuming too much about what will be done with them. I think the first layer of a GI system can be done the same way. My self-selected speciality is resource allocation (RA). There are some times when certain forms of creativity are not a good option, e.g. flying a passenger jet. When shouldn't humans be creative? How should creativity and X other systems be managed? Looking at opencog the RA is not baked into the arch so I have doubts about how well it would survive in its current state under recursive self-change. It will probably be reasonable for what the opencog team is doing at the moment, but getting low-level arch wrong or not fit for the next stage is a good way to waste work. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: [agi] Re: Defining AGI
2008/10/19 Dr. Matthias Heger [EMAIL PROTECTED]: The process of outwardly expressing meaning may be fundamental to any social intelligence but the process itself needs not much intelligence. Every email program can receive meaning, store meaning and it can express it outwardly in order to send it to another computer. It even can do it without loss of any information. Regarding this point, it even outperforms humans already who have no conscious access to the full meaning (information) in their brains. The only thing which needs much intelligence from the nowadays point of view is the learning of the process of outwardly expressing meaning, i.e. the learning of language. The understanding of language itself is simple. I'd disagree, there is another part of dealing with language that we don't have a good idea of how to do. Deciding whether to assimilate it and if so how. If I specify in a language to a computer that it should do something, it will do it no matter what (as long as I have sufficient authority). Telling a human to do something, e.g. wave your hands in the air and shout, the human will decide to do that based on how much it trusts you and whether they think it is a good idea. Generally a good idea in a situation where you are attracting the attention of rescuers, otherwise likely to make you look silly. I'm generally in favour of getting some NLU into AIs mainly because a lot of the information we have about the world is still in that form, so an AI without access to that information would have to reinvent it, which I think would take a long time. Even mathematical proofs are still somewhat in natural language. Other than that you could work on machine language understanding where information was taken in selectively and judged on its merits not its security credentials. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Twice as smart (was Re: RSI without input...) v2.1))
2008/10/18 Ben Goertzel [EMAIL PROTECTED]: 1) There definitely IS such a thing as a better algorithm for intelligence in general. For instance, compare AIXI with an algorithm called AIXI_frog, that works exactly like AIXI, but inbetween each two of AIXI's computational operations, it internally produces and then deletes the word frog one billion times. Clearly AIXI is better than AIXI_frog, according to many reasonable quantitative intelligence measures. Hi Ben, First off, the quantitative measure of intelligence of both systems would be the same! They both can't exist :-p Brief definition: A system is intelligent in world X if it achieves it's goals. Over most worlds, maybe AIXI is more intelligent. Over all worlds I'd say definately no. The theory behind AIXI doesn't account for death, there is nothing that the system can do to the environment that makes the system stop computing. One world that could easily exist is one where very fast computers where cracked down on, and eliminated with extreme prejudice. A system that slowed down it's aparrent ability to process might not incur the wrath of the anti seedAI police. If you reject that scenario I can construct a real world one, with me a debugger and the dreaded kill command, where I can make the froggy AI more able to do whatever it is trying to do, because it still exists. Arbitrary sure, but humans are strange and arbitrary. So it might be the right thing for the non_froggy AI to change itself to a froggy AI to better achieve it's goals in the long term. This could be considered an improvement, but won't help it improve more quickly in the future. 2) More relevantly, there is definitely such a thing as a better algorithm for intelligence about, say, configuring matter into various forms rapidly. Or you can substitute any other broad goal here. Let me put it this way as system designers we are often tasked with choosing between accuracy and speed. Say the problem is generating realistic pictures. Sure we could render everything with radiosity, but you will get far far fewer frames in a given time period than with ray casting even with todays hardware. If you want the best possible picyure you use radiosity and ray tracing, if you want real time motion pictures you would just project the polygons onto a screen with a zbuffer. I see no reason why AIs won't have to make the same decision between speed and accuracy for whatever is best for their intelligence in the future. Therefore the better algorithm is context dependent. In terms of reconfiguring matter, if you are sending an expensive probe into the reaches of space you would probably want to model everything down to the last angstrom. Less so if you were just making a towel quickly. These require different ways of thinking. Or would your idea of best would be one that could do both? But then you might be wasting computer resources, if you were never asked to do one or the other of these types of tasks. 3) Anyway, I think it's reasonable to doubt my story about how RSI will be achieved. All I have is a plausibility argument, not a proof. What got my dander up about Matt's argument was that he was claiming to have a debunking of the RSI ... a proof that it is impossible or infeasible. I do not think he presented any such thing; I think he presented an opinion in the guise of a proof It may be a reasonable opinion but that's very different from a proof. He defines RSI too tightly to make much use in the real world. However similar definitions have been done for intelligence. And then things proved about overly tight definition. It worries me that we don't have a way of settling disputes like these. Whether RSI is likely is an important fact for the future of humanity, surely we should be able to pick out some thread of reality that we can experiment on without building full AI. Theories of intelligence should guide us enough to say, If we exist in a world where RSI is unlikely, X should also be unlikely and vice versa. As the possibility of faster than light travel was squashed without have to try travelling faster than the speed of light. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Twice as smart (was Re: RSI without input...) v2.1))
2008/10/17 Ben Goertzel [EMAIL PROTECTED]: The difficulty of rigorously defining practical intelligence doesn't tell you ANYTHING about the possibility of RSI ... it just tells you something about the possibility of rigorously proving useful theorems about RSI ... More importantly, you haven't dealt with my counterargument that the posited AGI that is qualitatively intellectually superior to humans in every way would a) be able to clone itself N times for large N b) have the full knowledge-base and infrastructure of human society at its disposal Surely these facts will help it to self-improve far more quickly than would otherwise be the case... I'm not thinking about this so abstractly, really. I'm thinking, qualitatively, that 1-- The members of this list, collectively, could solve algorithmic problems that a team of one million people with IQ 100 would not be able to solve in a feasible period of time 2-- an AGI that was created by, say, the members of this list, would be architected based on **our** algorithms 3-- so, if we could create an AGI that was qualitatively intellectually superior to **us** (even if only moderately so), this AGI (or a team of such) could probably solve algorithmic problems that one million of **us** would not be able to solve in a feasible period of time 4--thus, this AGI we created would be able to create another AGI that was qualitatively much smarter than **it** 5--etc. I don't buy the 5 step plan, either. For a few reasons. Apologies for the rather disjointed nature of this message, it is rather late, and I want to finish it before I am busy again. I don't think there is such thing as an better algorithm for intelligence, there are algorithms suited to certain problems. Human intelligences seem to adapt their main reasoning algorithms in an experimental self-changing fashion at a sub concious level. Different biases are appropriate for different problems, including at the meta-level. See deceptive functions from genetic algorithms for examples. And deceptive functions can always appear in the world, as humans can create whatever problems are needed to fool the other agents around them. What Intelligence generally measures in day to day life is the ability to adopt other peoples mental machinery, for your own purposes. It gives no guarantee of finding new solutions to problems. The search spaces are so huge that you can easily lose yourself, trying to hit a tiny point. You might have the correct biases to get to point A, but that doesn't mean you have the right biases to get to point B. True innovation is very hard. It is not hard to be bayesian optimal if you know what data you should be looking at to solve a problem, it is knowing what data is pertinent. This is not always obvious and requires trial, error and the correct bias to limit this to reasonable time scales. Copying yourself doesn't get you different biases. You would all try the same approach to start with, or if you purposefully set it so that you didn't you would all still rate certain things/approaches as very unlikely to be any good, when they might well be what you need to do. Will --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
2008/10/14 Terren Suydam [EMAIL PROTECTED]: --- On Tue, 10/14/08, Matt Mahoney [EMAIL PROTECTED] wrote: An AI that is twice as smart as a human can make no more progress than 2 humans. Spoken like someone who has never worked with engineers. A genius engineer can outproduce 20 ordinary engineers in the same timeframe. Do you really believe the relationship between intelligence and output is linear? I'm going to use this post as a place to grind one of my axes, apologies Terren. The relationship between processing power and results is not necessarily linear or even positively correlated. And as an increase in intelligence above a certain level requires increased processing power (or perhaps not? anyone disagree?). When the cost of adding more computational power, outweighs the amount of money or energy that you acquire from adding the power, there is not much point adding the computational power. Apart from if you are in competition with other agents, that can out smart you. Some of the traditional views of RSI neglects this and thinks that increased intelligence is always a useful thing. It is not very There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed. Will Pearson --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Hi Terren, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. What I meant was at processing power/memory Z, there is an problem solving ability Y which is the maximum. To increase the problem solving ability above Y you would have to increase processing power/memory. That is when cognitive efficiency reaches one, in your terminology. Efficiency is normally measured in ratios so that seems natural. There are things you can't model with limits of processing power/memory which restricts your ability to solve them. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Not quite what I meant. Will --- 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=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] COMP = false
2008/10/4 Colin Hales [EMAIL PROTECTED]: Hi Will, It's not an easy thing to fully internalise the implications of quantum degeneracy. I find physicists and chemists have no trouble accepting it, but in the disciplines above that various levels of mental brick walls are in place. Unfortunately physicists and chemists aren't usually asked to create vision!... I inhabit an extreme multidisciplinary zone. This kind of mental resistance comes with the territory. All I can say is 'resistance is futile, you will be assimilated' ... eventually. :-) It's part of my job to enact the necessary advocacy. In respect of your comments I can offer the following: I started off doing chemistry at Uni, but I didn't like all the wet experiments. There are things like the bonds in graphite sheets that are degenerate, but that is of a completely different nature to the electrical signals in the brain. You are exactly right: humans don't encounter the world directly (naive realism). Nor are we entirely operating from a cartoon visual fantasy(naive solipsism). You are also exactly right in that vision is not 'perfect'. It has more than just a level of indirectness in representation, it can malfunction and be fooled - just as you say. In the benchmark behaviour: scientific behaviour, we know scientists have to enact procedures (all based around the behaviour called 'objectivity') which minimises the impact of these aspects of our scientific observation system. However, this has nothing to say about the need for an extra information source. necessary for there is not enough information in the signals to do the job. This is what you cannot see. It took me a long while to discard the tendency to project my mental capacity into the job the brain has when it encounters a retinal data stream. In vision processing using computing we know the structure of the distal natural world. We imagine the photon/CCD camera chip measurements to be the same as that of the retina. It looks like a simple reconstruction job. I've never thought computer vision to be simple... But it is not like that at all. It is impossible to tell, from the signals in their natural state in the brain, whether they are about vision or sound or smell. They all look the same. So I did not completely reveal the extent of the retinal impact/visual scene degeneracy in my post. The degeneracy operates on multiple levels. Signal encoding into standardised action potentials is another level. The locations that the signals travel through would be a strong indication of what they are about. It also seems likely that the different signals would have different statistics. For example somehow the human brain can learn to get visual data from the tongue with a brainport. http://vision.wicab.com/index.php I'm not entirely sure what you are getting at, do you think we are in superposition with the environment? Would you expect a camera + signals going through your tongue to preserve that? Maybe I can just paint a mental picture of the job the brain has to do. Imagine this: You have no phenomenal consciousness at all. Your internal life is of a dreamless sleep. Except ... for a new perceptual mode called Wision. Looming in front of you embedded in a roughly hemispherical blackness is a gigantic array of numbers. The numbers change. Now: a) make a visual scene out of it representing the world outside: convert Wision into Vision. b) do this without any information other than the numbers in front of you and without assuming you have any a-priori knowledge of the outside world. That is the job the brain has. Resist the attempt to project your own knowledge into the circumstance. You will find the attempt futile. The brain starts with at least some structure that has implicit knowledge of the outside world (just as bones shows the genome stores information of what is strong in the world). The Blank slate does not seem a viable hypothesis. There are no numbers in the brain or even in a computer it is all electric signals distributed spatially, temporally and with different statistics that allow them to be distinguished. I'd be curious to read your thoughts in a bit more of a structured format, but I can't get a grasp of what you are trying to say at the moment, it seems degenerate with other signals :P Will --- 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=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] COMP = false
Hi Colin, I'm not entirely sure that computers can implement consciousness. But I don't find your arguments sway me one way or the other. A brief reply follows. 2008/10/4 Colin Hales [EMAIL PROTECTED]: Next empirical fact: (v) When you create a turing-COMP substrate the interface with space is completely destroyed and replaced with the randomised machinations of the matter of the computer manipulating a model of the distal world. All actual relationships with the real distal external world are destroyed. In that circumstance the COMP substrate is implementing the science of an encounter with a model, not an encounter with the actual distal natural world. No amount of computation can make up for that loss, because you are in a circumstance of an intrinsically unknown distal natural world, (the novelty of an act of scientific observation). . But humans don't encounter the world directly, else optical illusions wouldn't exist, we would know exactly what was going on. Take this site for example. http://www.michaelbach.de/ot/ It is impossible by physics to do vision perfectly without extra information, but we do not do vision by any means perfectly, so I see no need to posit an extra information source. Will --- 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=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
[agi] Waiting to gain information before acting
I've started to wander away from my normal sub-cognitive level of AI, and have been thinking about reasoning systems. One scenario I have come up with is the, foresight of extra knowledge, scenario. Suppose Alice and Bob have decided to bet $10 on the weather in the 10 days time in alaska whether it is warmer or colder than average, it is Bobs turn to pick his side. He already thinks that it is going to be warmer than average (p 0.6) based on global warming and prevailing conditions. But he also knows that the weather in russia 5 day before is a good indicator of the conditions, that is he has a p 0.9 that if the russian weather is colder than average on day x alaskan weather will be colder than average on day x+5 and likewise for warmer. He has to pick his side of the bet 3 days before the due date so he can afford to wait. My question is, are current proposed reasoning systems able to act so that Bob doesn't bet straight away, and waits for the extra information from Russia before making the bet? Lets try some backward chaining. Make money - Win bet - Pick most likely side - Get more information about the most likely side The probability that a warm russia implies a warm alaska, does not intrinsically indicate that it gives you more information, allowing you to make a better bet. So, this is where I come to a halt, somewhat. How do you proceed the inference from here, it would seem you would have to do something special and treat every possible event that increases your ability to make a good guess on this bet as implying you have got more information (and some you don't?). You also would need to go with the meta-probability or some other indication of how good an estimate is, so that more information could be quantified. There are also more esoteric examples of waiting for more information, for example suppose Bob doesn't know about the russia-alaska connections but knows that a piece of software is going to be released that improves weather predictions in general. Can we still hook up that knowledge somehow? Will Pearson --- 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=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] self organization
2008/9/16 Terren Suydam [EMAIL PROTECTED]: Hi Will, Such an interesting example in light of a recent paper, which deals with measuring the difference between activation of the visual cortex and blood flow to the area, depending on whether the stimulus was subjectively invisible. If the result can be trusted, it shows that blood flow to the cortex is correlated with whether the stimulus is being perceived or not, as opposed to the neural activity, which does not change... see a discussion here: http://network.nature.com/groups/bpcc/forum/topics/2974 In this case then the reward that the cortex receives in the form of nutrients is based somehow on feedback from other parts of the brain involved with attention. It's like a heuristic that says, if we're paying attention to something, we're probably going to keep paying attention to it. Maier A, Wilke M, Aura C, Zhu C, Ye FQ, Leopold DA. Nat Neurosci. 2008 Aug 24. [Epub ahead of print], Divergence of fMRI and neural signals in V1 during perceptual suppression in the awake monkey. Interesting, I'll have to check it out. Thanks. I really need to keep up with brain research a little better. Will --- 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=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] self organization
2008/9/15 Vladimir Nesov [EMAIL PROTECTED]: I guess that intuitively, argument goes like this: 1) economy is more powerful than individual agents, it allows to increase the power of intelligence in individual agents; 2) therefore, economy has an intelligence-increasing potency; 3) so, we can take stupid agents, apply the economy potion to them and get powerful intelligence as a result. I like economies. They are not a magic bullet, but part of the solution at the low level. Let me see if I can explain why, note please that I am explaining why economies might be used in human type fallible systems, very improbably fallible systems (the sort required for friendliness) seem improbable to me. Take the problem of the human brain blood flow. You only have a certain amount of blood to flow to each part of the brain. Now there are three obvious things that you can do: 1) Have a centralised bit of the brain that makes choices about how to distribute the blood flow. This in turn would need a large blood flow to it. It would also need to be vastly complex and know what was going on in all bits of the brain so as it changed it could make sensible decisions about how to direct the blood flow. It might end up taking a large proportion of the blood flow, and if any errors were in it they would not self-correct. 2) Each bit of the brain indicates how much blood it needs at the moment. However if any bit of the brain thinks it needs more blood than it actually needs it could sit their stuck in a loop wasting lots of oxygen. 3) Each bit of the brain pays a bit of non-forgable credit for the blood flow they get. They get credit by participating in an economy with the amygdala as the ultimate money source. If they pay more credit than they get, they tend to lose credit and go bankrupt, so they can't foul up the system any more. Now this is a very hypothetical view of the brain. But I think the answer to how the brain decides to allocate blood flow would be more similar to the third case, expecting errors but self-correcting and taking up minimal resources. However despite it being nothing to do with bayesian reasoning or rational decision making, if we didn't have a good way of allocating blood flow in our brains we really couldn't do very much of use at all (as blood would be directed to the wrong parts at the wrong times). Decentralised economies of dumb things can be somewhat useful. See for example Learning Classifier Systems. Personally I would prefer to create economies of things as smart as our best systems. They would then work on solving different parts of how to win. Will Pearson --- 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=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Does prior knowledge/learning cause GAs to converge too fast on sub-optimal solutions?
2008/9/8 Benjamin Johnston [EMAIL PROTECTED]: Does this issue actually crop up in GA-based AGI work? If so, how did you get around it? If not, would you have any comments about what makes AGI special so that this doesn't happen? Does it also happen in humans? I'd say yes, therefore it might be a problem we can't avoid but only mitigate by having communities of intelligences sharing ideas so that they can shake each other out of their maxima assuming they settle in different ones (different search landscapes and priors help with this). The community might reach a maxima as well, but the world isn't constant so good ideas might always be good, changing the search landscapes, meaning a maxima my not be a maxima any longer. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] A NewMetaphor for Intelligence - the Computer/Organiser
2008/9/5 Mike Tintner [EMAIL PROTECTED]: MT:By contrast, all deterministic/programmed machines and computers are guaranteed to complete any task they begin. Will:If only such could be guaranteed! We would never have system hangs, dead locks. Even if it could be made so, computer systems would not always want to do so. Will, That's a legalistic, not a valid objection, (although heartfelt!).In the above case, the computer is guaranteed to hang - and it does, strictly, complete its task. Not necessarily, the task could be interrupted at that process stopped or paused indefinately. What's happened is that you have had imperfect knowledge of the program's operations. Had you known more, you would have known that it would hang. If it hung because of mult-process issues, you would need perfect knowledge of the environment to know the possible timing issues as well. Were your computer like a human mind, it would have been able to say (as you/we all do) - well if that part of the problem is going to be difficult, I'll ignore it or.. I'll just make up an answer... or by God I'll keep trying other ways until I do solve this.. or... .. or ... Computers, currently, aren't free thinkers. Computers aren't free thinkers, but it does not follow from an inability to switch, cancel, pause and restart or modify tasks. All of which they can do admirably. They just don't tend to do so, because they aren't smart enough (and cannot change themselves to be so) to know when it might be appropriate for what they are trying to do, so it is left up to the human operator to do so. I'm very interested in computers that self-maintain, that is reduce (or eliminate) the need for a human to be in the loop or know much about the internal workings of the computer. However it doesn't need a vastly different computing paradigm it just needs a different way of thinking about the systems. E.g. how can you design a system that does not need a human around to fix mistakes, upgrade it or maintain it in general. As they change their own system I will not know what they are going to do, because they can get information from the environment about how to act. This will me it a 'free thinker' of sorts. Whether it will be enough to get what you want, is an empirical matter, as far as I am concerned. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] A NewMetaphor for Intelligence - the Computer/Organiser
2008/9/6 Mike Tintner [EMAIL PROTECTED]: Will, Yes, humans are manifestly a RADICALLY different machine paradigm- if you care to stand back and look at the big picture. Employ a machine of any kind and in general, you know what you're getting - some glitches (esp. with complex programs) etc sure - but basically, in general, it will do its job. What exactly is a desktop computers job? Humans are only human, not a machine. Employ one of those, incl. yourself, and, by comparison, you have only a v. limited idea of what you're getting - whether they'll do the job at all, to what extent, how well. Employ a programmer, a plumber etc etc.. Can you get a good one these days?... VAST difference. If you find a new computer that I do not know how it has been programmed (whether it has linux/windows and what version). You also lack knowledge of what it is going to do. Aibo is a computer as well! It follows a program. And that's the negative side of our positive side - the fact that we're 1) supremely adaptable, and 2) can tackle those problems that no machine or current AGI - (actually of course, there is no such thing at the mo, only pretenders) - can even *begin* to tackle. Our unreliability . That, I suggest, only comes from having no set structure - no computer program - no program of action in the first place. (Hey, good idea, who needs a program?) You equate set structure with computer program. A computer program is not set! There is set structure of some sorts in the brain, at the neural level anyway. so you would have to be more precise in what you mean by lack of set structure. Wait, program of action? You don't think computer programs are like lists of things to do in the real world, do you? That is just something cooked up by the language writers to make things easier to deal with, a computer program is really only about memory manipulation. Some of the memory locations might be hooked up to the real world, but at the end of the day the computer treats it all as semanticless memory manipulations. Since what controls the memory manipulations are themselves in memory, they to can be manipulated! Here's a simple, extreme example. Will, I want you to take up to an hour, and come up with a dance, called the Keyboard Shuffle. (A very ill-structured problem.) How about you go learn about self-modifying assembly language, preferably with real-time interrupts. That would be a better use of the time, I think. Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] A NewMetaphor for Intelligence - the Computer/Organiser
2008/9/5 Mike Tintner [EMAIL PROTECTED]: By contrast, all deterministic/programmed machines and computers are guaranteed to complete any task they begin. If only such could be guaranteed! We would never have system hangs, dead locks. Even if it could be made so, computer systems would not always want to do so. Have you every had a programmed computer system say to you. This program is not responding, do you wish to terminate it. There is no reason in principle why the decision to terminate the program couldn't be made automatically. (Zero procrastination or deviation). Multi-tasking systems deviate all the time... Very different kinds of machines to us. Very different paradigm. (No?) We commonly talk about single program systems because they are generally interesting, and can be analysed simply. My discussion on self-modifying systems ignored the interrupt driven multi-tasking nature of the system I want to build, because that makes analysis a lot more hard. I will still be building an interrupt driven, multi tasking system. Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] A NewMetaphor for Intelligence - the Computer/Organiser
2008/9/4 Mike Tintner [EMAIL PROTECTED]: Terren, If you think it's all been said, please point me to the philosophy of AI that includes it. A programmed machine is an organized structure. A keyboard (and indeed a computer with keyboard) are something very different - there is no organization to those 26 letters etc. They can be freely combined and sequenced to create an infinity of texts. That is the very essence and manifestly, the whole point, of a keyboard. Yes, the keyboard is only an instrument. But your body - and your brain - which use it, are themselves keyboards. They consist of parts which also have no fundamental behavioural organization - that can be freely combined and sequenced to create an infinity of sequences of movements and thought - dances, texts, speeches, daydreams, postures etc. In abstract logical principle, it could all be preprogrammed. But I doubt that it's possible mathematically - a program for selecting from an infinity of possibilities? And it would be engineering madness - like trying to preprogram a particular way of playing music, when an infinite repertoire is possible and the environment, (in this case musical culture), is changing and evolving with bewildering and unpredictable speed. To look at computers as what they are (are you disputing this?) - machines for creating programs first, and following them second, is a radically different way of looking at computers. It also fits with radically different approaches to DNA - moving away from the idea of DNA as coded program, to something that can be, as it obviously can be, played like a keyboard - see Dennis Noble, The Music of Life. It fits with the fact (otherwise inexplicable) that all intelligences have both deliberate (creative) and automatic (routine) levels - and are not just automatic, like purely programmed computers. And it fits with the way computers are actually used and programmed, rather than the essentially fictional notion of them as pure turing machines. And how to produce creativity is the central problem of AGI - completely unsolved. So maybe a new approach/paradigm is worth at least considering rather than more of the same? I'm not aware of a single idea from any AGI-er past or present that directly addresses that problem - are you? You can't create a program out of thin air. So you have to have some sort of program to start with. You probably want to change the initial program in some way as well as perhaps adding more programming. This leads you to recursive self-change and its subset RSI, which is a very tricky business even if you don't think it is going to go FOOM and take over the world. So this very list has been discussing in abstract terms the very thing you want it to be discussing! Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Recursive self-change: some definitions
2008/9/2 Ben Goertzel [EMAIL PROTECTED]: Yes, I agree that your Turing machine approach can model the same situations, but the different formalisms seem to lend themselves to different kinds of analysis more naturally... I guess it all depends on what kinds of theorems you want to formulate... What I am interested in is if someone gives me a computer system that changes its state is some fashion, can I state how powerful that method of change is likely to be? That is what the exact difference between a traditional learning algorithm and the way I envisage AGIs changing their state. Also can you formalise the difference between a humans method of learning how to learn, and boot strapping language off language (both examples of a strange loop), and a program inspecting and changing its source code. I'm also interested in recursive self changing systems and whether you can be sure they will stay recursive self changing systems, as they change. This last one especially with regard to people designs systems with singletons in mind. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
2008/8/28 Valentina Poletti [EMAIL PROTECTED]: Got ya, thanks for the clarification. That brings up another question. Why do we want to make an AGI? To understand ourselves as intelligent agents better? It might enable us to have decent education policy, rehabilitation of criminals. Even if we don't make human like AGIs the principles should help us understand ourselves, just as optics of the lens helped us understand the eye and aerodynamics of wings helps us understand bird flight. It could also gives us more leverage, more brain power on the planet to help solve the planets problems. This is all predicated on the idea that fast take off is pretty much impossible. It is possible then all bets are off. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
[agi] Recursive self-change: some definitions
I've put up a short fairly dense un-referenced paper (basically an email but in a pdf to allow for maths) here. http://codesoup.sourceforge.net/RSC.pdf Any thoughts/ feed back welcomed. I'll try and make it more accessible at some point, but I don't want to spend too much time on it at the moment. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Recursive self-change: some definitions
2008/9/2 Ben Goertzel [EMAIL PROTECTED]: Hmmm.. Rather, I would prefer to model a self-modifying AGI system as something like F(t+1) = (F(t))( F(t), E(t) ) where E(t) is the environment at time t and F(t) is the system at time t Are you assuming the system knows the environment totally? Or did you mean the input the system gets from the environment? Would you have to assume the environment was deterministic as well in order to construct a hyperset? Unless you can construct a hyperset tree kind of thing, with branches for each possible environmental state? This is a hyperset equation, but it seems to nicely and directly capture the fact that the system is actually acting on and modifying itself... I'll use _ to indicate subscript for now. I think s_n+1 = g_s_n(x) encompasses the same idea of self-modification, as the function that g performs on x is determined by the state if you consider g to be a UTM and s to be a program it becomes a bit clearer. Consider g() and f() to be the hardware or physics of the system. Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Goedel machines ..PS
2008/8/29 Ben Goertzel [EMAIL PROTECTED]: About recursive self-improvement ... yes, I have thought a lot about it, but don't have time to write a huge discourse on it here One point is that if you have a system with N interconnected modules, you can approach RSI by having the system separately think about how to improve each module. I.e. if there are modules A1, A2,..., AN ... then you can for instance hold A1,...,A(N-1) constant while you think about how to improve AN. One can then iterate through all the modules and improve them in sequence. (Note that the modules are then doing the improving of each other.) I'm not sure what you are getting at here... Is modification system implemented in a module (Ai)? If so how would evaluate whether a modification Ai, call it AI' did a better job? What I am trying to figure out is whether the system you are describing could change to one which modules A1 to A10 were modified twice as often as the other modules? Can it change itself so it could remove a module altogether, or duplicate a module and specialise each of the modules to a different purpose? Will Pearson Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Goedel machines ..PS
2008/8/30 Ben Goertzel [EMAIL PROTECTED]: On Sat, Aug 30, 2008 at 10:06 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/8/29 Ben Goertzel [EMAIL PROTECTED]: About recursive self-improvement ... yes, I have thought a lot about it, but don't have time to write a huge discourse on it here One point is that if you have a system with N interconnected modules, you can approach RSI by having the system separately think about how to improve each module. I.e. if there are modules A1, A2,..., AN ... then you can for instance hold A1,...,A(N-1) constant while you think about how to improve AN. One can then iterate through all the modules and improve them in sequence. (Note that the modules are then doing the improving of each other.) I'm not sure what you are getting at here... Is modification system implemented in a module (Ai)? If so how would evaluate whether a modification Ai, call it AI' did a better job? The modification system is implemented in a module (subject to modification), but this is a small module, which does most of its work by calling on other AI modules (also subject to modification)... Isn't it an evolutionary stable strategy for the modification system module to change to a state where it does not change itself?1 Let me give you a just so story and you can tell me whether you think it likely. I'd be curious as to why you don't. Let us say the AI is trying to learn a different language (say french with its genders), so the system finds it is better to concentrate its change on the language modules as these need the most updating. So a modification to the modification module that completely concentrates the modifications on the language module should be the best at that time. But then it would be frozen forever and once the need to vary the language module was past it wouldn't be able to go back to modifying other modules. Short sighted I know, but I have yet to come across an RSI system that isn't either short sighted or limited to what it can prove. Will 1 Assuming there is no pressure on it for variation. --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Goedel machines ..PS
2008/8/30 Ben Goertzel [EMAIL PROTECTED]: Isn't it an evolutionary stable strategy for the modification system module to change to a state where it does not change itself?1 Not if the top-level goals are weighted toward long-term growth Let me give you a just so story and you can tell me whether you think it likely. I'd be curious as to why you don't. Let us say the AI is trying to learn a different language (say french with its genders), so the system finds it is better to concentrate its change on the language modules as these need the most updating. So a modification to the modification module that completely concentrates the modifications on the language module should be the best at that time. But then it would be frozen forever and once the need to vary the language module was past it wouldn't be able to go back to modifying other modules. Short sighted I know, but I have yet to come across an RSI system that isn't either short sighted or limited to what it can prove. You seem to be assuming that subgoal alienation will occur, and the long-term goal of dramatically increasing intelligence will be forgotten in favor of the subgoal of improving NLP. But I don't see why you make this assumption; this seems an easy problem to avoid in a rationally-designed AGI system, although not so easy in the context of human psychology. Have you implemented a long term growth goal atom yet? Don't they have to specify a specific state? Or am I reading http://opencog.org/wiki/OpenCogPrime:GoalAtom wrong? Also do you have any information on how the top level goal will play a part in assigning a fitness in Moses? How can you evaluate how good a change to a module will be for long term growth, without allowing the system to run for a long time and measure its growth? Will --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Goedel machines ..PS
2008/8/30 Ben Goertzel [EMAIL PROTECTED]: Don't they have to specify a specific state? Or am I reading http://opencog.org/wiki/OpenCogPrime:GoalAtom wrong? They don't have to specify a specific state. A goal could be some PredicateNode P expressing an abstract evaluation of state, programmed in Combo (a general purpose programming language)... So it could be a specific set of states? To specify long term growth as a goal, wouldn't you need to be able to do an abstract evaluation of how the state *changes* rather than just the current state? Also do you have any information on how the top level goal will play a part in assigning a fitness in Moses? That comes down to the basic triad Context Procedure == Goal The aim of the Ai mind is to understand the context it's in, then learn or select a procedure that it estimates (infers) will have a high probability of helping it achieve its goal in the relevant context. MOSES is a procedure learning algorithm... This is described in the chapter on goal-oriented cognition in the OCP wikibook... Searching for goal in the wikibook got me a whole lot of pages, none of them with goal in the title. Is there any way to de-wiki the titles so that a search for goal would pick up http://opencog.org/wiki/OpenCogPrime:SchemaContextGoalTriad in its title? Goal picks up way too many text searches. I'll have a read of it. How can you evaluate how good a change to a module will be for long term growth, without allowing the system to run for a long time and measure its growth? By inference... ... at least, that's the theory ;-) What are your expected false positive rates for classifying a change to the modification module that leads to long term growth? Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: RSI (was Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)))
2008/8/29 j.k. [EMAIL PROTECTED]: On 08/28/2008 04:47 PM, Matt Mahoney wrote: The premise is that if humans can create agents with above human intelligence, then so can they. What I am questioning is whether agents at any intelligence level can do this. I don't believe that agents at any level can recognize higher intelligence, and therefore cannot test their creations. The premise is not necessary to arrive at greater than human intelligence. If a human can create an agent of equal intelligence, it will rapidly become more intelligent (in practical terms) if advances in computing technologies continue to occur. An AGI with an intelligence the equivalent of a 99.-percentile human might be creatable, recognizable and testable by a human (or group of humans) of comparable intelligence. That same AGI at some later point in time, doing nothing differently except running 31 million times faster, will accomplish one genius-year of work every second. Will it? It might be starved for lack of interaction with the world and other intelligences, and so be a lot less productive than something working at normal speeds. Most learning systems aren't constrained by lack of processing power for how long it takes them to learn things (AIXI excepted), but by the speed of running an experiment. Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: RSI (was Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)))
2008/8/29 j.k. [EMAIL PROTECTED]: On 08/29/2008 01:29 PM, William Pearson wrote: 2008/8/29 j.k.[EMAIL PROTECTED]: An AGI with an intelligence the equivalent of a 99.-percentile human might be creatable, recognizable and testable by a human (or group of humans) of comparable intelligence. That same AGI at some later point in time, doing nothing differently except running 31 million times faster, will accomplish one genius-year of work every second. Will it? It might be starved for lack of interaction with the world and other intelligences, and so be a lot less productive than something working at normal speeds. Yes, you're right. It doesn't follow that its productivity will necessarily scale linearly, but the larger point I was trying to make was that it would be much faster and that being much faster would represent an improvement that improves its ability to make future improvements. The numbers are unimportant, but I'd argue that even if there were just one such human-level AGI running 1 million times normal speed and even if it did require regular interaction just like most humans do, that it would still be hugely productive and would represent a phase-shift in intelligence in terms of what it accomplishes. Solving one difficult problem is probably not highly parallelizable in general (many are not at all parallelizable), but solving tens of thousands of such problems across many domains over the course of a year or so probably is. The human-level AGI running a million times faster could simultaneously interact with tens of thousands of scientists at their pace, so there is no reason to believe it need be starved for interaction to the point that its productivity would be limited to near human levels of productivity. Only if it had millions of times normal human storage capacity and memory bandwidth, else it couldn't keep track of all the conversations, and sufficient bandwidth for ten thousand VOIP calls at once. We should perhaps clarify what you mean by speed here? The speed of the transistor is not all of what makes a system useful. It is worth noting that processor speed hasn't gone up appreciably from the heady days of Pentium 4s with 3.8 GHZ in 2005. Improvements have come from other directions (better memory bandwidth, better pipelines and multi cores). The hard disk is probably what is holding back current computers at the moment. Will Pearson --- 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/?; Powered by Listbox: http://www.listbox.com --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] The Necessity of Embodiment
2008/8/25 Terren Suydam [EMAIL PROTECTED]: --- On Sun, 8/24/08, Vladimir Nesov [EMAIL PROTECTED] wrote: On Sun, Aug 24, 2008 at 5:51 PM, Terren Suydam wrong. This ability might be an end in itself, the whole point of building an AI, when considered as applying to the dynamics of the world as a whole and not just AI aspect of it. After all, we may make mistakes or be swayed by unlucky happenstance in all matters, not just in a particular self-vacuous matter of building AI. I don't deny the possibility of disaster. But my stance is, if the only approach you have to mitigate disaster is being able to control the AI itself, well, the game is over before you even start it. It seems profoundly naive to me that anyone could, even in principle, guarantee a super-intelligent AI to renormalize, in whatever sense that means. Then you have the difference between theory and practice... just forget it. Why would anyone want to gamble on that? You may be interested in goedel machines. I think this roughly fits the template that Eliezer is looking for, something that reliably self modifies to be better. http://www.idsia.ch/~juergen/goedelmachine.html Although he doesn't like explicit utility functions, the provably better is something he want. Although what you would accept as axioms for the proofs upon which humanity fate rests I really don't know. Personally I think strong self-modification is not going to be useful, the very act of trying to understand the way the code for an intelligence is assembled will change the way that some of that code is assembled. That is I think that intelligences have to be weakly self modifying, in the same way bits of the brain rewire themselves locally and subconciously, so to, AI will need to have the same sort of changes in order to keep up with humans. Computers at the moment can do lots of things better that humans (logic, bayesian stats), but are really lousy at adapting and managing themselves so the blind spots of infallible computers are always exploited by slow and error prone, but changeable, humans. Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)
2008/8/23 Matt Mahoney [EMAIL PROTECTED]: Valentina Poletti [EMAIL PROTECTED] wrote: I was wondering why no-one had brought up the information-theoretic aspect of this yet. It has been studied. For example, Hutter proved that the optimal strategy of a rational goal seeking agent in an unknown computable environment is AIXI: to guess that the environment is simulated by the shortest program consistent with observation so far [1]. By my understanding, I would qualify this as Hutter proved that the *one of the* optimal strategies of a rational error-free goal seeking agent, which has no impact on the environment beyond its explicit output, in an unknown computable environment is AIXI: to guess that the environment is simulated by the shortest program consistent with observation so far Will Pearson --- 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=8660244id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] The Necessity of Embodiment
2008/8/11 Mike Tintner [EMAIL PROTECTED]: Will: thought you meant rational as applied to the system builder :P Consistency of systems is overrated, as far as I am concerned. Consistency is only important if it ever the lack becomes exploited. A system that alter itself to be consistent after the fact is sufficient. Do you remember when I wrote this? http://www.mail-archive.com/agi@v2.listbox.com/msg07233.html What parts of it suggest a fixed and totalitarian system to you? WIll, I didn't still don't quite understand your ideas there. You need to give some examples of how they might apply to particular problems.The fact that a program/set of programs can change v. radically - and even engage opposite POV's - doesn't necessarily mean it isn't still a totalitarian system. My ideas, at present, don't as such apply to particular problems. They apply to the shaping of the system. It would make as much sense as asking how setting up the method of voting in a country applied to solving the national debt. Or the monetary system applied to how to transport a person from A to B. Now at some point I or someone else will have to try to solve the practical problems, But if the system allows the analogues of vote rigging or free money in some fashion then even if I set up the system with the right non-totalitarian methods it could still all go horribly wrong. So the shaping system is more fundamental and needs to be solved first. Will Pearson --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] The Necessity of Embodiment
2008/8/10 Mike Tintner [EMAIL PROTECTED]: Just as you are in a rational, specialist way picking off isolated features, so, similarly, rational, totalitarian thinkers used to object to the crazy, contradictory complications of the democratic, conflict system of decisionmaking by contrast with their pure ideals. And hey, there *are* crazy and inefficient features - it's a real, messy system. But, as a whole, it works better than any rational, totalitarian, non-conflict system. Cog sci can't yet explain why, though, can it? (You guys, without realising it, are all rational, totalitarian systembuilders). All? I'm a rational economically minded system builder, thank you very much. I can't answer questions you want answered, like how will my system reason with imagination precisely because I am not a totalitarian. If you wish to be non-totalitarian you have set up a system in a certain way and let the dynamics set up potentially transform the system into something that can reason as you want. Theoretically the system could be set up to reason as you want straight away. But setting up a baby level system seems orders of magnitude easier than expecting it solve problems straight away. In which exact knowledge of the inner workings of mature imagination is not required. The more you ask for early results of systems, the more you are likely to get totalitarians building your machines. Because they can get results quick. Will Pearson --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
[agi] Definition of Pattern?
Is there a mathematical wiki-pedia sized definition of the a Goertzelian pattern out there? It would make assessing the underpinnings of Open Cog Prime easier. Will Pearson --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] The exact locus of the supposed 'complexity'
2008/8/3 Richard Loosemore [EMAIL PROTECTED]: I probably don't need to labor the rest of the story, because you have heard it before. If there is a brick wall between the overall behavior of the system and the design choices that go into it - if it is impossible to go from 'I want the system to behave like [that]' to 'therefore I need to make [this] choice of design at the low level' - then all the stuff about using intuition to sense the right design would go out the window. This is why the conversation yesterday about what John Conway actually did when he came up with Game of Life was so important: the documentary evidence suggests that what he and his team did was just blind search. Other people have tried to assert that he used mathematical intuition. The complex systems community would say that in almost all projects like the one Conway undertook, there would be absolutely no choice whatsoever but to do a blind search. Might it be worth setting people a challenge? Set people the task of building a complex system with a certain property or maybe a few (nothing too bad, perhaps selecting a rule number from something akin to Wolframs numbering). They give reasons why they picked the rules they did and see if they do better than a RNG at picking the correct number. You appear to be going against a strong intuition here, so giving people a practical experiment they can play on themselves might be worthwhile. Will Pearson --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
What does it do? useful in AGI? Re: [agi] US PATENT ISSUED for the TEN ETHICAL LAWS OF ROBOTICS
2008/7/22 Mike Archbold [EMAIL PROTECTED]: It looks to me to be borrowed from Aristotle's ethics. Back in my college days, I was trying to explain my project and the professor kept interrupting me to ask: What does it do? Tell me what it does. I don't understand what your system does. What he wanted was input-function-output. He didn't care about my fancy data structure or architecture goals, he just wanted to know what it DID. I have come across this a lot. And while it is a very useful heuristic for sniffing out bad ideas that don't do anything I also think it is harmful to certain other endeavours. Imagine this hypothetical conversation between Turing and someone else (please ignore all historical inaccuracies). Sceptic: Hey Turing, how is it going. Hmm, what are you working on at the moment? Turing: A general purpose computing machine. Sceptic: I'm not really sure what you mean by computing. Can you give me an example of something it does? Turing: Well you can use it calculate differential equations Sceptic: So it is a calculator, we already have machines that can do that. Turing: Well it can also be a chess player. Sceptic: Wait, what? How can something be a chess player and a calculator? Turing: Well it isn't both at the same time, but you can reconfigure it to do one then the other. Sceptic: If you can reconfigure something, that means it doesn't intrinsically do one or the other. So what does the machine do itself? Turing: Well, err, nothing. I think the quest for general intelligence (if we are to keep any meaning in the word general), will have be hindered by trying to pin down what candidate systems do, in the same way general computing would be. I think the requisite question in AGI to fill the gap formed by not allowing this question, is, How does it change? Will --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: Location of goal/purpose was Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/14 Terren Suydam [EMAIL PROTECTED]: Will, --- On Fri, 7/11/08, William Pearson [EMAIL PROTECTED] wrote: Purpose and goal are not intrinsic to systems. I agree this is true with designed systems. And I would also say of evolved systems. My fingers purpose could equally well be said to be for picking ticks out of the hair of my kin or for touch typing. E.g. why do I keep my fingernails short, so that they do not impede my typing. The purpose of gut bacteria is to help me digest my food. The purpose of part of my brain is to do differentiation of functions, because I have . The designed system is ultimately an extension of the designer's mind, wherein lies the purpose. Oddly enough that is what I want the system to be. Rather an extension of my brain. Of course, as you note, the system in question can serve multiple purposes, each of which lies in the mind of some other observer. The same is true of your system, even though its behavior may evolve. Your button is what tethers its purpose to your mind. On the other hand, we can create simulations in which purpose is truly emergent. To support emergence our design must support large-scale, (global) interactions of locally specified entities. Conway's Game of Life is an example of such a system - what is its purpose? To provide an interesting system for researchers to research cellular automata? ;) I think I can see your point, It has no practical purpose as such. Just a research purpose. It certainly wasn't specified. And neither am I specifying the purpose of mine! I'm quite happy to hook up the button to something I press when I feel like it. I could decide the purpose of the system was to learn and be good at backgammon one day, in which case my presses would reflect that, or I could decide the purpose of the system was to search the web. If you want to think of a good analogy for how emergent I want the system to be. Imagine someone came along to one of your life simulations and interfered with the simulation to give some more food to some of the entities that he liked the look of. This wouldn't be anything so crude as to specify the fitness or artificial breeding, but it would tilt the scales in the favour of entities that he liked all else being equal. Would this invalidate the whole simulation because he interfered and bought some of his purpose into it? If so, I don't see why. The simplest answer is probably that it has none. But what if our design of the local level was a little more interesting, such that at the global level, we would eventually see self-sustaining entities that reproduced, competed for resources, evolved, etc, and became more complex over a large number of iterations? Then the system itself still wouldn't have a practical purpose. For a system Y to have a purpose, you have to have be able to say part X is like it is for Y to perform its function. Internal state corresponding to the entities might be said to have purpose, but not the system as a whole. Whether that's possible is another matter, but assuming for the moment it was, the purpose of that system could be defined in roughly the same way as trying to define the purpose of life itself. We have to be careful here. What meaning of the word life are you using? 1) The biosphere + evolution 2) And individuals exsistance. The first has no purpose. You can never look at the biosphere and figure out what bits are for what in the grander scheme of things, or ask yourself what mutations are likely to be thrown up to better achieve its goal. That we have some self-regulation on the Gaian scale is purely anthropic, biospheres without it would likely have driven themselves to a state not able to support lives. An individual entity has a purpose, though. So to that extent the purposeless can create the purposeful. So unless you believe that life was designed by God (in which case the purpose of life would lie in the mind of God), the purpose of the system is indeed intrinsic to the system itself. I think I would still say it didn't have a purpose. If I get your meaning right. Will --- 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=8660244id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] Is clustering fundamental?
2008/7/6 Abram Demski [EMAIL PROTECTED]: In fact, adding hidden predicates and entities in the case of Markov logic makes the space of models Turing-complete (and even bigger than that if higher-order logic is used). But if I am not mistaken the clustering used in the paper I refer to is not that powerful. So the question is: is clustering in general powerful enough for AGI? Is it fundamental to how minds can and should work? I would say very important, but not fundamental. Consider the square/rectangle problem. You are given a number of pairs of numbers and you want to somehow say that pairs with the same numbers are in one class (of squares) and pairs of different numbers are rectangles. Imagine you have to learn to eat squares but not rectangles. However most cluster methods, while they could represent the cluster of squares, would require a lot of samples to get a long thin cluster running up the x=y line. If there was another dimension called z which was 1 when x equalled y, clustering would be very easy. Where does z come from? And why not z = 1 if x-9 = y^2? Finding decent dimensions to cluster on is a tricky problem. So I think the process by which the dimensions of clustering problems are created is more fundamental than clustering itself. Will Pearson --- 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=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: Formal proved code change vs experimental was Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/3 Steve Richfield [EMAIL PROTECTED]: William and Vladimir, IMHO this discussion is based entirely on the absence of any sort of interface spec. Such a spec is absolutely necessary for a large AGI project to ever succeed, and such a spec could (hopefully) be wrung out to at least avoid the worst of the potential traps. And if you want the interface to be upgradeable, or alterable what then? This conversation was based on the ability to change as much of the functional and learning parts of the systems as possible. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Terren, Remember when I said that a purpose is not the same thing as a goal? The purpose that the system might be said to have embedded is attempting to maximise a certain signal. This purpose presupposes no ontology. The fact that this signal is attached to a human means the system as a whole might form the goal to try and please the human. Or depending on what the human does it might develop other goals. Goals are not the same as purposes. Goals require the intentional stance, purposes the design. To the extent that purpose is not related to goals, it is a meaningless term. In what possible sense is it worthwhile to talk about purpose if it doesn't somehow impact what an intelligent actually does? Does the following make sense? The purpose embedded within the system will be try and make the system not decrease in its ability to receive some abstract number. The way I connect up the abstract number to the real world will the govern what goals the system will likely develop (along with the initial programming). That is there is some connection, but it is tenuous and I don't have to specify an ontology. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/3 Terren Suydam [EMAIL PROTECTED]: --- On Wed, 7/2/08, William Pearson [EMAIL PROTECTED] wrote: Evolution! I'm not saying your way can't work, just saying why I short cut where I do. Note a thing has a purpose if it is useful to apply the design stance* to it. There are two things to differentiate between, having a purpose and having some feedback of a purpose built in to the system. I don't believe evolution has a purpose. See Hod Lipson's TED talk for an intriguing experiment in which replication is an inevitable outcome for a system of building blocks explicitly set up in a random fashion. In other words, purpose is emergent and ultimately in the mind of the beholder. See this article for an interesting take that increasing complexity is a property of our laws of thermodynamics for non-equilibrium systems: http://biology.plosjournals.org/perlserv/?request=get-documentdoi=10.1371/journal.pbio.0050142ct=1 In other words, Darwinian evolution is a special case of a more basic kind of selection based on the laws of physics. This would deprive evolution of any notion of purpose. Evolution doesn't have a purpose, it creates things with purpose. Where purpose means it is useful to apply the design stance on it, e.g. ask what an eye on a frog is for. It is the second I meant, I should have been more specific. That is to apply the intentional stance to something successfully, I think a sense of its own purpose is needed to be embedded in that entity (this may only be a very crude approximation to the purpose we might assign something looking from an evolution eye view). Specifying a system's goals is limiting in the sense that we don't force the agent to construct its own goals based on it own constructions. In other words, this is just a different way of creating an ontology. It narrows the domain of applicability. That may be exactly what you want to do, but for AGI researchers, it is a mistake. Remember when I said that a purpose is not the same thing as a goal? The purpose that the system might be said to have embedded is attempting to maximise a certain signal. This purpose presupposes no ontology. The fact that this signal is attached to a human means the system as a whole might form the goal to try and please the human. Or depending on what the human does it might develop other goals. Goals are not the same as purposes. Goals require the intentional stance, purposes the design. Also your way we will end up with entities that may not be useful to us, which I think of as a negative for a long costly research program. Will Usefulness, again, is in the eye of the beholder. What appears not useful today may be absolutely critical to an evolved descendant. This is a popular explanation for how diversity emerges in nature, that a virus or bacteria does some kind of horizontal transfer of its genes into a host genome, and that gene becomes the basis for a future adaptation. When William Burroughs said language is a virus, he may have been more correct than he knew. :-] Possibly, but it will be another huge research topic to actually talk to the things that evolve in the artificial universe, as they will share very little background knowledge or ontology with us. I wish you luck and will be interested to see where you go but the alife route is just to slow and resource intensive for my liking. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Thu, Jul 3, 2008 at 12:59 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 9:09 PM, William Pearson [EMAIL PROTECTED] wrote: They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. So, this process performs optimization, A has a goal that it tries to express in form of A'. What is the problem with the algorithm that A uses? If this algorithm is stupid (in a technical sense), A' is worse than A and we can detect that. But this means that in fact, A' doesn't do its job and all the search pressure comes from program B that ranks the performance of A or A'. This generate-blindly-or-even-stupidly-and-check is a very inefficient algorithm. If, on the other hand, A happens to be a good program, then A' has a good change of being better than A, and anyway A has some understanding of what 'better' means, then what is the role of B? B adds almost no additional pressure, almost everything is done by A. How do you distribute the optimization pressure between generating programs (A) and checking programs (B)? Why do you need to do that at all, what is the benefit of generating and checking separately, compared to reliably generating from the same point (A alone)? If generation is not reliable enough, it probably won't be useful as optimization pressure anyway. The point of A and A' is that A', if better, may one day completely replace A. What is very good? Is 1 in 100 chances of making a mistake when generating its successor very good? If you want A' to be able to replace A, that is only 100 generations before you have made a bad mistake, and then where do you go? You have a bugged program and nothing to act as a watchdog. Also if A' is better than time A at time t, there is no guarantee that it will stay that way. Changes in the environment might favour one optimisation over another. If they both do things well, but different things then both A and A' might survive in different niches. I suggest you read ( http://sl4.org/wiki/KnowabilityOfFAI ) If your program is a faulty optimizer that can't pump the reliability out of its optimization, you are doomed. I assume you argue that you don't want to include B in A, because a descendant of A may start to fail unexpectedly. Nope. I don't include B in A because if A' is faulty it can cause problems to whatever is in the same vmprogram as it, by overwriting memory locations. A' being a separate vmprogram means it is insulated from the B and A, and can only have limited impact on them. I don't get what your obsession is with having things all be in one program is anyway. Why is that better? I'll read knowability of FAI again, but I have read it before and I don't think it will enlighten me. I'll come back to the rest of your email once I have done that. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/3 Vladimir Nesov [EMAIL PROTECTED]: On Thu, Jul 3, 2008 at 10:45 AM, William Pearson [EMAIL PROTECTED] wrote: Nope. I don't include B in A because if A' is faulty it can cause problems to whatever is in the same vmprogram as it, by overwriting memory locations. A' being a separate vmprogram means it is insulated from the B and A, and can only have limited impact on them. Why does it need to be THIS faulty? If there is a known method to prevent such faultiness, it can be reliably implemented in A, so that all its descendants keep it, unless they are fairly sure it's not needed anymore or there is a better alternative. Because it is dealing with powerful stuff, when it gets it wrong it goes wrong powerfully. You could lock the experimental code away in a sand box inside A, but then it would be a separate program just one inside A, but it might not be able to interact with programs in a way that it can do its job. There are two grades of faultiness. frequency and severity. You cannot predict the severity of faults of arbitrary programs (and accepting arbitrary programs from the outside world is something I want the system to be able to do, after vetting etc). I don't get what your obsession is with having things all be in one program is anyway. Why is that better? I'll read knowability of FAI again, but I have read it before and I don't think it will enlighten me. I'll come back to the rest of your email once I have done that. It's not necessarily better, but I'm trying to make explicit in what sense is it worse, that is what is the contribution of your framework to the overall problem, if virtually the same thing can be done without it. I'm not sure why you see this distinction as being important though. I call the vmprograms separate because they have some protection around them, but you could see them as all one big program if you wanted. The instructions don't care whether we call the whole set of operations a program or not. This, from one point of view, is true at least while it is being simulated the whole VM is one program inside a larger system. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Formal proved code change vs experimental was Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Sorry about the long thread jack 2008/7/3 Vladimir Nesov [EMAIL PROTECTED]: On Thu, Jul 3, 2008 at 4:05 PM, William Pearson [EMAIL PROTECTED] wrote: Because it is dealing with powerful stuff, when it gets it wrong it goes wrong powerfully. You could lock the experimental code away in a sand box inside A, but then it would be a separate program just one inside A, but it might not be able to interact with programs in a way that it can do its job. There are two grades of faultiness. frequency and severity. You cannot predict the severity of faults of arbitrary programs (and accepting arbitrary programs from the outside world is something I want the system to be able to do, after vetting etc). You can't prove any interesting thing about an arbitrary program. It can behave like a Friendly AI before February 25, 2317, and like a Giant Cheesecake AI after that. Whoever said you could? The whole system is designed around the ability to take in or create arbitrary code, give it only minimal access to other programs that it can earn and lock it out from that ability when it does something bad. By arbitrary code I don't mean random, I mean stuff that has not formally been proven to have the properties you want. Formal proof is too high a burden to place on things that you want to win. You might not have the right axioms to prove the changes you want are right. Instead you can see the internals of the system as a form of continuous experiments. B is always testing a property of A or A', if at any time it stops having the property that B looks for then B flags it as buggy. I know this doesn't have the properties you would look for in a friendly AI set to dominate the world. But I think it is similar to the way humans work, and will be as chaotic and hard to grok as our neural structure. So as likely as humans are to explode intelligently. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Theoretic estimation of reliability vs experimental
2008/7/3 Vladimir Nesov [EMAIL PROTECTED]: On Thu, Jul 3, 2008 at 9:36 PM, William Pearson [EMAIL PROTECTED] wrote: Sorry about the long thread jack 2008/7/3 Vladimir Nesov [EMAIL PROTECTED]: On Thu, Jul 3, 2008 at 4:05 PM, William Pearson [EMAIL PROTECTED] wrote: Because it is dealing with powerful stuff, when it gets it wrong it goes wrong powerfully. You could lock the experimental code away in a sand box inside A, but then it would be a separate program just one inside A, but it might not be able to interact with programs in a way that it can do its job. There are two grades of faultiness. frequency and severity. You cannot predict the severity of faults of arbitrary programs (and accepting arbitrary programs from the outside world is something I want the system to be able to do, after vetting etc). You can't prove any interesting thing about an arbitrary program. It can behave like a Friendly AI before February 25, 2317, and like a Giant Cheesecake AI after that. Whoever said you could? The whole system is designed around the ability to take in or create arbitrary code, give it only minimal access to other programs that it can earn and lock it out from that ability when it does something bad. By arbitrary code I don't mean random, I mean stuff that has not formally been proven to have the properties you want. Formal proof is too high a burden to place on things that you want to win. You might not have the right axioms to prove the changes you want are right. Instead you can see the internals of the system as a form of continuous experiments. B is always testing a property of A or A', if at any time it stops having the property that B looks for then B flags it as buggy. The point isn't particularly about formal proof, but more about any theoretic estimation of reliability and optimality. If you produce an artifact A' and theoretically estimate that probability of it working correctly is such that you don't expect it to fail in 10^9 years, you can't beat this reliability with a result of experimental testing. Thus, if theoretic estimation is possible (and it's much more feasible for purposefully designed A' than for arbitrary A'), experimental testing has vanishingly small relevance. This, I think, is a wild goose chase, hence why I am not following it. Why won't the estimation system will run out of steam, like Lenats Automated Mathematician? I know this doesn't have the properties you would look for in a friendly AI set to dominate the world. But I think it is similar to the way humans work, and will be as chaotic and hard to grok as our neural structure. So as likely as humans are to explode intelligently. Yes, one can argue that AGI of minimal reliability is sufficient to jump-start singularity (it's my current position anyway, Oracle AI), but the problem with faulty design is not only that it's not going to be Friendly, but that it isn't going to work at all. By what principles do you think humans develop their intellects? I don't seem to be made processes that probabilistically guarantee that I will work better tomorrow than I did today. How do you explain developing echolocation or specific areas specialised for reading braille in blind people? Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Sorry about the late reply. snip some stuff sorted out 2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: On Tue, Jul 1, 2008 at 2:02 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: If internals are programmed by humans, why do you need automatic system to assess them? It would be useful if you needed to construct and test some kind of combination/setting automatically, but not if you just test manually-programmed systems. How does the assessment platform help in improving/accelerating the research? Because to be interesting the human specified programs need to be autogenous, as in Josh Storr Hall's terminology, which means self-building. Capable of altering the stuff they are made of. In this case machine code equivalent. So you need the human to assess the improvements the system makes, for whatever purpose the human wants the system to perform. Altering the stuff they are made of is instrumental to achieving the goal, and should be performed where necessary, but it doesn't happen, for example, with individual brains. I think it happens at the level of neural structures. I.e. I think neural structures control the development of other neural structures. (I was planning to do the next blog post on this theme, maybe tomorrow.) Do you mean to create population of altered initial designs and somehow select from them (I hope not, it is orthogonal to what modification is for in the first place)? Otherwise, why do you still need automated testing? Could you present a more detailed use case? I'll try and give a fuller explanation later on. This means he needs to use a bunch more resources to get a singular useful system. Also the system might not do what he wants, but I don't think he minds about that. I'm allowing humans to design everything, just allowing the very low level to vary. Is this clearer? What do you mean by varying low level, especially in human-designed systems? The machine code the program is written in. Or in a java VM, the java bytecode. This still didn't make this point clearer. You can't vary the semantics of low-level elements from which software is built, and if you don't modify the semantics, any other modification is superficial and irrelevant. If it's not quite 'software' that you are running, and it is able to survive the modification of lower level, using the terms like 'machine code' and 'software' is misleading. And in any case, it's not clear what this modification of low level achieves. You can't extract work from obfuscation and tinkering, the optimization comes from the lawful and consistent pressure in the same direction. Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. Does this clear things up? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Terren Suydam [EMAIL PROTECTED]: Mike, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Why is it going too far? I agree with you that we can reconstruct thinking, to a point. I notice you didn't say we can completely reconstruct how humans think about problems. Why not? We have two primary means for understanding thought, and both are deeply flawed: 1. Introspection. Introspection allows us to analyze our mental life in a reflective way. This is possible because we are able to construct mental models of our mental models. There are three flaws with introspection. The first, least serious flaw is that we only have access to that which is present in our conscious awareness. We cannot introspect about unconscious processes, by definition. This is a less serious objection because it's possible in practice to become conscious of phenomena there were previously unconscious, by developing our meta-mental-models. The question here becomes, is there any reason in principle that we cannot become conscious of *all* mental processes? The second flaw is that, because introspection relies on the meta-models we need to make sense of our internal, mental life, the possibility is always present that our meta-models themselves are flawed. Worse, we have no way of knowing if they are wrong, because we often unconsciously, unwittingly deny evidence contrary to our conception of our own cognition, particularly when it runs counter to a positive account of our self-image. Harvard's Project Implicit experiment (https://implicit.harvard.edu/implicit/) is a great way to demonstrate how we remain ignorant of deep, unconscious biases. Another example is how little we understand the contribution of emotion to our decision-making. Joseph Ledoux and others have shown fairly convincingly that emotion is a crucial part of human cognition, but most of us (particularly us men) deny the influence of emotion on our decision making. The final flaw is the most serious. It says there is a fundamental limit to what introspection has access to. This is the an eye cannot see itself objection. But I can see my eyes in the mirror, says the devil's advocate. Of course, a mirror lets us observe a reflected version of our eye, and this is what introspection is. But we cannot see inside our own eye, directly - it's a fundamental limitation of any observational apparatus. Likewise, we cannot see inside the very act of model-simulation that enables introspection. Introspection relies on meta-models, or models about models, which are activated/simulated *after the fact*. We might observe ourselves in the act of introspection, but that is nothing but a meta-meta-model. Each introspectional act by necessity is one step (at least) removed from the direct, in-the-present flow of cognition. This means that we can never observe the cognitive machinery that enables the act of introspection itself. And if you don't believe that introspection relies on cognitive machinery (maybe you're a dualist, but then why are you on an AI list? :-), ask yourself why we can't introspect about ourselves before a certain point in our young lives. It relies on a sufficiently sophisticated toolset that requires a certain amount of development before it is even possible. 2. Theory. Our theories of cognition are another path to understanding, and much of theory is directly or indirectly informed by introspection. When introspection fails (as in language acquisition), we rely completely on theory. The flaw with theory should be obvious. We have no direct way of testing theories of cognition, since we don't understand the connection between the mental and the physical. At best, we can use clever indirect means for generating evidence, and we usually have to accept the limits of reliability of subjective reports. My plan is go for 3) Usefulness. Cognition is useful from an evolutionary point of view, if we try to create systems that are useful in the same situations (social, building world models), then we might one day stumble upon cognition. To expand on usefulness in social contexts, you have to ask yourself what the point of language is, why is it useful in an evolutionary setting. One thing the point of language is not, is fooling humans that you are human, which makes me annoyed at all the chatbots that get coverage as AI. I'll write more on this later. This by the way is why I don't self-organise purpose. I am pretty sure a specified purpose (not the same thing as a goal, at all) is needed for an intelligence. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription:
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 2:48 PM, William Pearson [EMAIL PROTECTED] wrote: Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. What are the criteria that VM applies to vmprograms? If VM just shortcircuits the economic pressure of agents to one another, it in itself doesn't specify the direction of the search. The human economy works to efficiently satisfy the goals of human beings who already have their moral complexity. It propagates the decisions that customers make, and fuels the allocation of resources based on these decisions. Efficiency of economy is in efficiency of responding to information about human goals. If your VM just feeds the decisions on themselves, what stops the economy from focusing on efficiently doing nothing? They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. A' has some bugs such that the human notices something wrong with the system, she gives less credit on average each time A' is helping out rather than A. Now A and A' both have to bid for the chance to help program B which is closer to the outputting (due to the programming of B), B pays a proportion of the credit it gets back. Now the credit B gets will be lower when A' is helping, than when A is helping. So A' will get less in general than A. There are a few scenarios, ordered from quickest acting to slowest. 1 ) B keeps records of who helps him and sees that A' is not helping him as well as the average, so no longer lets A' bid. A' resources get used when it can't keep up bidding for them. 2) A' continues bidding a lot, to outbid A. However the average amount A' gets is less than it gets back from B. A' bankrupts itself and other programs use its resources. 3) A' doesn't manage to outbid A' after a fair few trials, so gets the same fate as it does in scenario 1) If you start with a bunch of stupid vmprograms, you won't get anywhere. It can just go to nothingness, you do have to design them fairly well, just in such a way that that design can change later. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Abram Demski [EMAIL PROTECTED]: How do you assign credit to programs that are good at generating good children? I never directly assign credit, apart from the first stage. The rest of the credit assignment is handled by the vmprograms, er, programming. Particularly, could a program specialize in this, so that it doesn't do anything useful directly but always through making highly useful children? As the parent controls the code of its offspring, it could embed code in its offspring to pass a small portion of the credit they get back to it. They would have to be careful how much to skim off so the offspring could still thrive. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 9:09 PM, William Pearson [EMAIL PROTECTED] wrote: They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. So, this process performs optimization, A has a goal that it tries to express in form of A'. What is the problem with the algorithm that A uses? If this algorithm is stupid (in a technical sense), A' is worse than A and we can detect that. But this means that in fact, A' doesn't do its job and all the search pressure comes from program B that ranks the performance of A or A'. This generate-blindly-or-even-stupidly-and-check is a very inefficient algorithm. If, on the other hand, A happens to be a good program, then A' has a good change of being better than A, and anyway A has some understanding of what 'better' means, then what is the role of B? B adds almost no additional pressure, almost everything is done by A. How do you distribute the optimization pressure between generating programs (A) and checking programs (B)? Why do you need to do that at all, what is the benefit of generating and checking separately, compared to reliably generating from the same point (A alone)? If generation is not reliable enough, it probably won't be useful as optimization pressure anyway. The point of A and A' is that A', if better, may one day completely replace A. What is very good? Is 1 in 100 chances of making a mistake when generating its successor very good? If you want A' to be able to replace A, that is only 100 generations before you have made a bad mistake, and then where do you go? You have a bugged program and nothing to act as a watchdog. Also if A' is better than time A at time t, there is no guarantee that it will stay that way. Changes in the environment might favour one optimisation over another. If they both do things well, but different things then both A and A' might survive in different niches. I would also be interested in why you think we have programmers and system testers in the real world. Also worth noting is most optimisation will be done inside the vmprograms, this process is only for very fundamental code changes, e.g. changing representations, biases, ways of creating offspring. Things that cannot be tested easily any other way. I'm quite happy for it to be slow, because this process is not where the majority of quickness of the system will rest. But this process is needed for intelligence else you will be stuck with certain ways of doing things when they are not useful. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/6/30 Terren Suydam [EMAIL PROTECTED]: Hi Will, --- On Mon, 6/30/08, William Pearson [EMAIL PROTECTED] wrote: The only way to talk coherently about purpose within the computation is to simulate self-organized, embodied systems. I don't think you are quite getting my system. If you had a bunch of programs that did the following 1) created new programs, by trial and error and taking statistics of variables or getting arbitrary code from the outside. 2) communicated with each other to try and find programs that perform services they need. 3) Bid for computer resources, if a program loses its memory resources it is selected against, in a way. Would this be sufficiently self-organised? If not, why not? And the computer programs would be as embodied as your virtual creatures. They would just be embodied within a tacit economy, rather than an artificial chemistry. It boils down to your answer to the question: how are the resources ultimately allocated to the programs? If you're the one specifying it, via some heuristic or rule, then the purpose is driven by you. If resource allocation is handled by some self-organizing method (this wasn't clear in the article you provided), then I'd say that the system's purpose is self-defined. I'm not sure how the system qualifies. It seems to be half way between the two definitions you gave. The programs can have special instructions in that bid for a specific resource with as much credit as they want (see my recent message replying to Vladimir Nesov for more information about banks, bidding and credit). The instructions can be removed or not done, the amount of credit bid can be changed. The credit is given to some programs by a fixed function, but they have instructions they can execute (or not) to give it to other programs forming an economy. What say you, self-organised or not? As for embodiment, my question is, how do your programs receive input? Embodiment, as I define it, requires that inputs are merely reflections of state variables, and not even labeled in any way... i.e. we can't pre-define ontologies. The embodied entity starts from the most unstructured state possible and self-structures whatever inputs it receives. Bits and bytes from the outside world, or bits and bytes from reading other programs programing and data. No particular ontology. That said, you may very well be doing that and be creating embodied programs in this way... if so, that's cool because I hadn't considered that possibility and I'll be interested to see how you fare. It is going to take a while. Virtual machine writing is very unrewarding programming. I have other things to do right now, I'll get back to the rest of the message in a bit. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/6/30 Terren Suydam [EMAIL PROTECTED]: Ben, I agree, an evolved design has limits too, but the key difference between a contrived design and one that is allowed to evolve is that the evolved critter's intelligence is grounded in the context of its own 'experience', whereas the contrived one's intelligence is grounded in the experience of its creator, and subject to the limitations built into that conception of intelligence. For example, we really have no idea how we arrive at spontaneous insights (in the shower, for example). A chess master suddenly sees the game-winning move. We can be fairly certain that often, these insights are not the product of logical analysis. So if our conception of intelligence fails to explain these important aspects, our designs based on those conceptions will fail to exhibit them. An evolved intelligence, on the other hand, is not limited in this way, and has the potential to exhibit intelligence in ways we're not capable of comprehending. I'm seeking to do something half way between what you suggest (from bacterial systems to human alife) and AI. I'd be curious to know whether you think it would suffer from the same problems. First are we agreed that the von Neumann model of computing has no hidden bias to its problem solving capabilities. It might be able to do some jobs more efficiently than other and need lots of memory to do others but it is not particularly suited to learning chess or running down a gazelle. Which means it can be reprogrammed to do either. However it has no guide to what it should be doing, so can become virus infested or subverted. It has a purpose but we can't explicitly define it. So let us try and put in the most minimal guide that we can so we don't give it a specific goal, just a tendency to favour certain activities or programs. How to do this? Form and economy based on reinforcement signals, those that get more reinforcement signals can outbid the others for control of system resources. This is obviously reminiscent of tierra and a million and one other alife system. The difference being is that I want the whole system to exhibit intelligence. Any form of variation is allowed, from random to getting in programs from the outside. It should be able to change the whole from the OS level up based on the variation. I agree that we want the systems we make to be free of our design constraints long term, that is eventually correct all the errors and oversimplifications or gaps we left. But I don't see the need to go all the way back to bacteria. Even then you would need to design the system correctly in terms of chemical concentrations. I think both would count as the passive approach* to helping solve the problem, yours is more indirect than is needed I think. Will Pearson * http://www.mail-archive.com/agi@v2.listbox.com/msg11399.html --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Hello Terren A Von Neumann computer is just a machine. It's only purpose is to compute. When you get into higher-level purpose, you have to go up a level to the stuff being computed. Even then, the purpose is in the mind of the programmer. What I don't see is why your simulation gets away from this, where as my architecture doesn't. Read the linked post in the previous message, if you want to understand more about the philosophy of the system. The only way to talk coherently about purpose within the computation is to simulate self-organized, embodied systems. I don't think you are quite getting my system. If you had a bunch of programs that did the following 1) created new programs, by trial and error and taking statistics of variables or getting arbitrary code from the outside. 2) communicated with each other to try and find programs that perform services they need. 3) Bid for computer resources, if a program loses its memory resources it is selected against, in a way. Would this be sufficiently self-organised? If not, why not? And the computer programs would be as embodied as your virtual creatures. They would just be embodied within a tacit economy, rather than an artificial chemistry. And I applaud your intuition to make the whole system intelligent. One of my biggest criticisms of traditional AI philosophy is over-emphasis on the agent. Indeed, the ideal simulation, in my mind, is one in which the boundary between agent and environment is blurry. In nature, for example, at low-enough levels of description it is impossible to find a boundary between the two, because the entities at that level are freely exchanged. You are right that starting with bacteria is too indirect, if your goal is to achieve AGI in something like decades. It would certainly take an enormous amount of time and computation to get from there to human-level AI and beyond, perhaps a hundred years or more. But you're asking, aren't there shortcuts we can take that don't limit the field of potential intelligence in important ways. If you take this attitude you would have to ask yourself whether implementing your simulation on a classical computer is not cutting off the ability to create intelligence. Perhaps quantum affects are important in whether a system can produce intelligence. Protein folding probably wouldn't be the same. You have to at some point simplify. I'm going to have my system have as many degrees of freedom to vary as a stored program computer (or as near as I can make it). Whilst having the internal programs self-organise and vary in ways that would make a normal stored program computer become unstable. Any simulations you do on a computer cannot have any more degrees of freedom. For example, starting with bacteria means we have to let multi-cellular organisms evolve on their own in a virtual geometry. That project alone is an enormous challenge. So let's skip it and go right to the multi-cellular design. The trouble is, our design of the multi-cellular organism is limiting. Alternative designs become impossible. What do you mean by design here? Do you mean an abstract multicellular cell model or do you mean design as in what Tom Ray (you do know Tierra right, I can use this as a common language?) did with his first self replicator, by creating an artificial genome. I can see problems with the first in restricting degrees of freedom, but the second, the degrees of freedom are still there to be acted on by the pressures of variation within the system. Even though Tom Ray built a certain type of replicator, they still managed to replicate in other ways, the one I can remember is stealing other peoples replication machinery as parasites. Lets say you started with an artificial chemistry. You could then design within that chemistry a replicator, then test that replicator. See if the variation is working okay. Then design a multicellular variant, by changing its genome. It could still slip back to single cellularity and find a different way to multicellularity. The degrees of freedom do not go away the second a human starts to design something (else genetically modified foods would not be such a thorny issue), you just got to allow the forces of variation to be able to act upon them. The question at that point is, are we excluding any important possibilities for intelligence if we build in our assumptions about what is necessary to support it, on a low-level basis. In what ways is our designed brain leaving out some key to adapting to unforeseen domains? Just apply a patch :P Or have an architecture that is capable of supporting a self-patching system. I have no fixed design for an AI myself. Intelligence means winning, winning requires flexibility. One of the basic threads of scientific progress is the ceaseless denigration of the idea that there is something special about humans. Pretending that we can solve AGI by mimicking top-down high-level human
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: On Mon, Jun 30, 2008 at 10:34 PM, William Pearson [EMAIL PROTECTED] wrote: I'm seeking to do something half way between what you suggest (from bacterial systems to human alife) and AI. I'd be curious to know whether you think it would suffer from the same problems. First are we agreed that the von Neumann model of computing has no hidden bias to its problem solving capabilities. It might be able to do some jobs more efficiently than other and need lots of memory to do others but it is not particularly suited to learning chess or running down a gazelle. Which means it can be reprogrammed to do either. However it has no guide to what it should be doing, so can become virus infested or subverted. It has a purpose but we can't explicitly define it. So let us try and put in the most minimal guide that we can so we don't give it a specific goal, just a tendency to favour certain activities or programs. It is a wrong level of organization: computing hardware is the physics of computation, it isn't meant to implement specific algorithms, so I don't quite see what you are arguing. I'm not implementing a specific algorithm I am controlling how resources are allocated. Currently architecture does whatever the kernel says, from memory allocation to irq allocation. Instead of this my architecture would allow any program to bid credit for a resource. The one that bids the most wins and spends its credit. Certain resources like output memory space, (i.e if the program is controlling the display or an arm or something) allow the program to specify a bank, and give the program income. A bank is a special variable that can't be edited by programs normally but can be spent. The bank of an outputing program will be given credit depending upon how well the system as whole is performing . If it is doing well the amount of credit it gets would be above average, poorly it would be below. After a certain time the resources will need to be bid for again. So credit is coming into the system and continually being sunk. The system will be seeded with programs that can perform rudimentarily well. E.g. you will have programs that know how to deal with visual input and they will bid for the video camera interupt. They will then sell their services for credit (so that they can bid for the interrupt again), to a program that correlates visual and auditory responses. Who sell their services to a high level planning module etc, on down to the arm that actually gets the credit. All these modules are subject to change and re-evaluation. They merely suggest one possible way for it to be used. It is supposed to be ultimately flexible. You could seed it with a self-replicating neural simulator that tried to hook its inputs and outputs up to other neurons. Neurons would die out if they couldn't find anything to do. How to do this? Form and economy based on reinforcement signals, those that get more reinforcement signals can outbid the others for control of system resources. Where do reinforcement signals come from? What does this specification improve over natural evolution that needed billions of years to get here (that is, why do you expect any results in the forseable future)? Most of the internals are programmed by humans, and they can be arbitrarily complex. The feedback comes from a human, or from a utility function although those are harder to define. The architecture simply doesn't restrict the degrees of freedom that the programs inside it can explore. This is obviously reminiscent of tierra and a million and one other alife system. The difference being is that I want the whole system to exhibit intelligence. Any form of variation is allowed, from random to getting in programs from the outside. It should be able to change the whole from the OS level up based on the variation. What is your meaning of `intelligence'? I now see it as merely the efficiency of optimization process that drives the environment towards higher utility, according to whatever criterion (reinforcement, in your case). In this view, how does I'll do the same, but with intelligence differ from I'll do the same, but better? Terran's artificial chemistry as whole could not be said to have a goal. Or to put it another way applying the intentional stance to it probably wouldn't help you predict what it did next. Applying the intentional stance to what my system does should help you predict what it does. This means he needs to use a bunch more resources to get a singular useful system. Also the system might not do what he wants, but I don't think he minds about that. I'm allowing humans to design everything, just allowing the very low level to vary. Is this clearer? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: On Tue, Jul 1, 2008 at 1:31 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: It is a wrong level of organization: computing hardware is the physics of computation, it isn't meant to implement specific algorithms, so I don't quite see what you are arguing. I'm not implementing a specific algorithm I am controlling how resources are allocated. Currently architecture does whatever the kernel says, from memory allocation to irq allocation. Instead of this my architecture would allow any program to bid credit for a resource. The one that bids the most wins and spends its credit. Certain resources like output memory space, (i.e if the program is controlling the display or an arm or something) allow the program to specify a bank, and give the program income. A bank is a special variable that can't be edited by programs normally but can be spent. The bank of an outputing program will be given credit depending upon how well the system as whole is performing . If it is doing well the amount of credit it gets would be above average, poorly it would be below. After a certain time the resources will need to be bid for again. So credit is coming into the system and continually being sunk. The system will be seeded with programs that can perform rudimentarily well. E.g. you will have programs that know how to deal with visual input and they will bid for the video camera interupt. They will then sell their services for credit (so that they can bid for the interrupt again), to a program that correlates visual and auditory responses. Who sell their services to a high level planning module etc, on down to the arm that actually gets the credit. All these modules are subject to change and re-evaluation. They merely suggest one possible way for it to be used. It is supposed to be ultimately flexible. You could seed it with a self-replicating neural simulator that tried to hook its inputs and outputs up to other neurons. Neurons would die out if they couldn't find anything to do. Well, yes, you implement some functionality, but why would you contrast it with underlying levels (hardware, OS)? Like Java virtual machine, your system is a platform, and it does some things not handled by lower levels, or, in this case, by any superficially analogous platforms. Because I want it done in silicon at some stage. It is also assumed to be the whole system, that is no other significant programs on it. Machines that run lisp natively have been made, this makes the most sense as the whole computer. Rather than as a component. How to do this? Form and economy based on reinforcement signals, those that get more reinforcement signals can outbid the others for control of system resources. Where do reinforcement signals come from? What does this specification improve over natural evolution that needed billions of years to get here (that is, why do you expect any results in the forseable future)? Most of the internals are programmed by humans, and they can be arbitrarily complex. The feedback comes from a human, or from a utility function although those are harder to define. The architecture simply doesn't restrict the degrees of freedom that the programs inside it can explore. If internals are programmed by humans, why do you need automatic system to assess them? It would be useful if you needed to construct and test some kind of combination/setting automatically, but not if you just test manually-programmed systems. How does the assessment platform help in improving/accelerating the research? Because to be interesting the human specified programs need to be autogenous, as in Josh Storr Hall's terminology, which means self-building. Capable of altering the stuff they are made of. In this case machine code equivalent. So you need the human to assess the improvements the system makes, for whatever purpose the human wants the system to perform. This is obviously reminiscent of tierra and a million and one other alife system. The difference being is that I want the whole system to exhibit intelligence. Any form of variation is allowed, from random to getting in programs from the outside. It should be able to change the whole from the OS level up based on the variation. What is your meaning of `intelligence'? I now see it as merely the efficiency of optimization process that drives the environment towards higher utility, according to whatever criterion (reinforcement, in your case). In this view, how does I'll do the same, but with intelligence differ from I'll do the same, but better? Terran's artificial chemistry as whole could not be said to have a goal. Or to put it another way applying the intentional stance to it probably wouldn't help you predict what it did next. Applying the intentional stance to what my system does should help you predict what it does. What
Re: [agi] Can We Start Somewhere was Approximations of Knowledge
2008/6/27 Steve Richfield [EMAIL PROTECTED]: Russell and William, OK, I think that I am finally beginning to get it. No one here is really planning to do wonderful things that people can't reasonably do, though Russell has pointed out some improvements which I will comment on separately. I still don't think you do. The general as far as I am concerned means it can reconfigure itself to do other things it couldn't previously. Just like a human learns differentiation. So when you ask for shopping list of things for it to do, you will get our first steps and things we know that can be done (because they have been done by humans) for testing etc... Consider a human computer team. Now the human can code/configure the machine to help her/him do pretty much anything. I just want to shift that coding/configuring work to the machine. That is hard convey in concrete examples, although I tried. I am interested in things that people can NOT reasonably do. Note that many computer programs have been written to way outperform people in specific tasks, and my own Dr. Eliza would seem to far exceed human capability in handling large amounts of qualitative knowledge that work within its paradigm limits. Hence, it would seem that I may have stumbled into the wrong group (opinions invited). Probably so ;) The solution to a specific tasks is not within the remit of the study of generality. You would be like someone going up to Turing and asking him what specific tasks ACE was going to solve. If he said cryptography, you would go on about the Bombe cracking engima. Unfortunately, no one here appears to be interested in understanding this landscape of solving future hyper-complex problems, but instead apparently everyone wishes to leave this work to some future AGI, that cannot possibly be constructed in the short time frame that I have in mind. Of course, future AGIs are doomed to fail at such efforts, just as people have failed for the last million years or so. If Humans and AGIs are doomed to fail at the task perhaps it is impossible? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Can We Start Somewhere was Approximations of Knowledge
I'm going to ignore the oversimplifications of a variety of peoples positions. But no one in AGI knows how to design or instruct a machine to work without algorithms - or, to be more precise, *complete* algorithms. It's unthinkable - it seems like asking someone not to breathe... until, like every problem,. you start thinking about it. Have you managed to create/design a toy system that can do this very basically. If so, do share. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Can We Start Somewhere was Approximations of Knowledge
2008/6/26 Steve Richfield [EMAIL PROTECTED]: Jiri previously noted that perhaps AGIs would best be used to manage the affairs of humans so that we can do as we please without bothering with the complex details of life. Of course, people and some (communist) governments now already perform this function, so while this might be a potential application, it doesn't count for this posting, as I am looking for things that people either can not do at all, or can not do adequately well. snip Thanks in advance for your concrete examples. Personally I concentrate on things humans could do, but that they don't have the time to do. Mostly I want to do Intelligence Augmentation through augmented reality. Highlight on a heads up display - food that corresponds to a certain health guidelines/ethical standards by object recognition and searching on-line information - books that might be interesting (again by searching information) or other people the user has known has read. None of these should have to be explictly programmed/configured by the user, the system should pick them up by interacting with the user and other machines. They should also only be done in contexts when the user is looking at the items involved (in a book store/library), and not just all the time. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Equivalent of the bulletin for atomic scientists or CRN for AI?
2008/6/23 Bob Mottram [EMAIL PROTECTED]: 2008/6/22 William Pearson [EMAIL PROTECTED]: 2008/6/22 Vladimir Nesov [EMAIL PROTECTED]: Well since intelligence explosions haven't happened previously in our light cone, it can't be a simple physical pattern Probably the last intelligence explosion - a relatively rapid increase in the degree of adaptability capabile of being exhibited by an organism - was the appearance of the first Homo sapiens. The number and variety of tools created by Homo sapiens compared to earlier hominids indicate that this was one of the great leaps forward in history (probably greatly facilitated by a more elaborate language ability). I am using intelligence explosion to mean what would Eliezer mean by it. See http://www.overcomingbias.com/2008/06/optimization-an.html#more I.e. something never seen on this planet. I am sceptical of whether such a process is theoretically possible. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Equivalent of the bulletin for atomic scientists or CRN for AI?
2008/6/23 Vladimir Nesov [EMAIL PROTECTED]: On Mon, Jun 23, 2008 at 12:50 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/6/22 Vladimir Nesov [EMAIL PROTECTED]: Two questions: 1) Do you know enough to estimate which scenario is more likely? Well since intelligence explosions haven't happened previously in our light cone, it can't be a simple physical pattern, so I think non-exploding intelligences have the evidence for being simpler on their side. This message that I'm currently writing hasn't happened previously in out light code. By your argument, it is evidence for it being more difficult to write, than to recreate life on Earth and human intellect, which is clearly false, for all practical purposes. You should state that argument more carefully, in order for it to make sense. If your message was an intelligent entity then you would have a point. I'm looking at classes of technologies and their natural or current human created analogues. Let me give you an example. You have two people claiming to be able to give you an improved TSP solver. One person claims to be able to do all examples in polynomial time the other simply has a better algorithm which can do certain types of graphs in polynomial time, but resorts to exponential time for random graphs. Which would you consider more likely if neither of them have detailed proofs and why? So we might find them more easily. I also think I have solid reasoning to think intelligence exploding is unlikely, which requires paper length rather than post length. So it I think I do, but should I trust my own rationality? But not too much, especially when the argument is not technical (which is clearly the case for questions such as this one). The question is one of theoretical computer science and should be able to be decided as well as the resolution to the halting problem. I'm leaning towards something like Russell Wallace's resolution, but there maybe some complications when you have a program that learns from the environment. I would like to see it done in formally at some point. If argument is sound, you should be able to convince seed AI crowd too Since the concept is their idea they have to be the ones to define it. They won't accept any arguments against it otherwise. They haven't as yet formally defined it, or if they have I haven't seen it. I agree, but it works only if you know that the answer is correct, and (which you didn't address and which is critical for these issues) you won't build a doomsday machine as a result of your efforts, even if this particular path turns out to be more feasible. I don't think a doomsday machine is possible. But considering I would be doing my best to make the system incapable of modifying it's own source code *in the fashion that eliezer wants/is afraid of* anyway, I am not too worried. See http://www.sl4.org/archive/0606/15131.html Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
[agi] Equivalent of the bulletin for atomic scientists or CRN for AI?
While SIAI fills that niche somewhat, it concentrates on the Intelligence explosion scenario. Is there a sufficient group of researchers/thinkers with a shared vision of the future of AI coherent enough to form an organisation? This organisation would discus, explore and disseminate what can be done to make the introduction as painless as possible. The base beliefs shared between the group would be something like - The entities will not have goals/motivations inherent to their form. That is robots aren't likely to band together to fight humans, or try to take over the world for their own means. These would have to be programmed into them, as evolution has programmed group loyalty and selfishness into humans. - The entities will not be capable of fully wrap around recursive self-improvement. They will improve in fits and starts in a wider economy/ecology like most developments in the world * - The goals and motivations of the entities that we will likely see in the real world will be shaped over the long term by the forces in the world, e.g. evolutionary, economic and physics. Basically an organisation trying to prepare for a world where AIs aren't sufficiently advanced technology or magic genies, but still dangerous and a potentially destabilising world change. Could a coherent message be articulated by the subset of the people that agree with these points. Or are we all still too fractured? Will Pearson * I will attempt to give an inside view of why I take this view, at a later date. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Equivalent of the bulletin for atomic scientists or CRN for AI?
2008/6/22 Vladimir Nesov [EMAIL PROTECTED]: Two questions: 1) Do you know enough to estimate which scenario is more likely? Well since intelligence explosions haven't happened previously in our light cone, it can't be a simple physical pattern, so I think non-exploding intelligences have the evidence for being simpler on their side. So we might find them more easily. I also think I have solid reasoning to think intelligence exploding is unlikely, which requires paper length rather than post length. So it I think I do, but should I trust my own rationality? Getting a bunch of people together to argue for both paths seems like a good bet at the moment. 2) What does this difference change for research at this stage? It changes the focus of research from looking for simple principles of intelligence (that can be improved easily on the fly), to one that expects intelligence creation to be a societal process over decades. It also makes secrecy no longer be the default position. If you take the intelligence explosion scenario seriously you won't write anything in public forums that might help other people make AI. As bad/ignorant people might get hold of it and cause the first explosion. Otherwise it sounds like you are just calling to start a cult that believes in this particular unsupported thing, for no good reason. ;-) Hope that gives you some reasons. Let me know if I have misunderstood your questions. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Breaking Solomonoff induction (really)
2008/6/21 Wei Dai [EMAIL PROTECTED]: A different way to break Solomonoff Induction takes advantage of the fact that it restricts Bayesian reasoning to computable models. I wrote about this in is induction unformalizable? [2] on the everything mailing list. Abram Demski also made similar points in recent posts on this mailing list. I think this is a lot stronger objection when you actually implement an implementable variant of Solomonoff Induction (it has started to make me chuckle that a model of induction makes assumptions about the universe that would have to be broken to have it implemented). When you restrict the the memory space of a system a lot more functions become uncomputable with respects to that system. It is not a safe assumption that the world is computable in this restricted notion of computable, i.e. computable with respect to a finite system. Also solomonoff induction ignores any potential physical affects of the computation, as does all probability theory. See section 5 of this attempted paper by me of an formalised example of where things could go wrong. http://codesoup.sourceforge.net/easa.pdf It is not quite an anthropic problem, but it is closely related. I'll tentatively label the observer-world interaction problem. That is the exact nature of the world you see is altered dependent upon the type of system you happen to be. All these are problem with tacit (a la Dennet) representations of beliefs embedded within the Solomonoff induction formalism. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Nirvana
2008/6/12 J Storrs Hall, PhD [EMAIL PROTECTED]: I'm getting several replies to this that indicate that people don't understand what a utility function is. If you are an AI (or a person) there will be occasions where you have to make choices. In fact, pretty much everything you do involves making choices. You can choose to reply to this or to go have a beer. You can choose to spend your time on AGI or take flying lessons. Even in the middle of typing a word, you have to choose which key to hit next. One way of formalizing the process of making choices is to take all the actions you could possibly do at a given point, predict as best you can the state the world will be in after taking such actions, and assign a value to each of them. Then simply do the one with the best resulting value. It gets a bit more complex when you consider sequences of actions and delayed values, but that's a technicality. Basically you have a function U(x) that rank-orders ALL possible states of the world (but you only have to evaluate the ones you can get to at any one time). We do mean slightly different things then. By U(x) I am just talking about a function that generates the set of scalar rewards for actions performed for a reinforcement learning algorithm. Not that evaluates every potential action from where the current system is (since I consider computation an action in order to take energy efficiency into consideration, this would be a massive space). Economists may crudely approximate it, but it's there whether they study it or not, as gravity is to physicists. ANY way of making decisions can either be reduced to a utility function, or it's irrational -- i.e. you would prefer A to B, B to C, and C to A. The math for this stuff is older than I am. If you talk about building a machine that makes choices -- ANY kind of choices -- without understanding it, you're talking about building moon rockets without understanding the laws of gravity, or building heat engines without understanding the laws of thermodynamics. The kinds of choices I am interested in designing for at the moment are should program X or program Y get control of this bit of memory or IRQ for the next time period. X and Y can also make choices and you would need to nail them down as well in order to get the entire U(x) as you talk about it. As the function I am interested in is only concerned about programmatic changes call it PCU(x). Can you give me a reason why the utility function can't be separated out this way? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Nirvana
2008/6/12 J Storrs Hall, PhD [EMAIL PROTECTED]: On Thursday 12 June 2008 02:48:19 am, William Pearson wrote: The kinds of choices I am interested in designing for at the moment are should program X or program Y get control of this bit of memory or IRQ for the next time period. X and Y can also make choices and you would need to nail them down as well in order to get the entire U(x) as you talk about it. As the function I am interested in is only concerned about programmatic changes call it PCU(x). Can you give me a reason why the utility function can't be separated out this way? This is roughly equivalent to a function where the highest-level arbitrator gets to set the most significant digit, the programs X,Y the next most, and so forth. As long as the possibility space is partitioned at each stage, the whole business is rational -- doesn't contradict itself. Modulo special cases, agreed. Allowing the program to play around with the less significant digits, i.e. to make finer distinctions, is probably pretty safe (and the way many AIers envisioning doing it). It's also reminiscent of the way Maslow's hierarchy works. But it doesn't work for full fledged AGI. It is the best design I have at the moment, whether it can make what you want is another matter. I'll continue to try to think of better ones. It should get me a useful system if nothing else, and hopefully more people interested in the full AGI problem, if it proves inadequate. What path are you going to continue down? Suppose you are a young man who's always been taught not to get yourself killed, and not to kill people (as top priorities). You are confronted with your country being invaded and faced with the decision to join the defense with a high liklihood of both. With the system I am thinking of it can get stuck in positions that aren't optimal as the the program control utility function only chooses from the extant programs in the system. It is possible for the system to be dominated by a monopoly or cartel of programs, such that the program chooser doesn't have a choice. This would only happen if there was a long period of stasis and a very powerful/useful set of programs. Such as possibly patriotism or the protection of other sentients in this case, being very useful during peace time. This does seem like you would consider it a bug, and it might be. It is not one I can currently see a guard against. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Nirvana
2008/6/11 J Storrs Hall, PhD [EMAIL PROTECTED]: Vladimir, You seem to be assuming that there is some objective utility for which the AI's internal utility function is merely the indicator, and that if the indicator is changed it is thus objectively wrong and irrational. There are two answers to this. First is to assume that there is such an objective utility, e.g. the utility of the AI's creator. I implicitly assumed such a point of view when I described this as the real problem. But consider: Any AI who believes this must realize that there may be errors and approximations in its own utility function as judged by the real utility, and must thus have as a first priority fixing and upgrading its own utility function. Thus it turns into a moral philosopher and it never does anything useful -- exactly the kind of Nirvana attractor I'm talking about. On the other hand, it might take its utility function for granted, i.e. assume (or agree to act as if) there were no objective utility. It's pretty much going to have to act this way just to get on with life, as indeed most people (except moral philosophers) do. But this leaves it vulnerable to modifications to its own U(x), as in my message. You could always say that you'll build in U(x) and make it fixed, which not only solves my problem but friendliness -- but leaves the AI unable to learn utility. I.e. the most important part of the AI mind is forced to remain brittle GOFAI construct. Solution unsatisfactory. I'm not quite sure what you find unsatisfactory. I think humans have a fixed U(x), but it is not a hard goal for the system but an implicit tendency for the internal programs to not self-modify away from (an agoric economy of programs is not oblidged to find better ways of getting credit, but a good set of programs is hard to dislodge by a bad set). I also think that part of humanity's U(x) relies on social interaction which can be a very complex function. Which can lead to very complex behaviour. Imagine if we were trying to raise children like we teach computers, we wouldn't reward the socially for playing with balls or saying their first words, but would put them straight into designing electronic circuits. Hence why I think that having one or more humans act as part of the U(x) of a system is necessary for interesting behaviour. If there is only one human acting as the input to the U(x) then I think the system and human should be considered part of a larger intentional system, as it will be trying to optimise one goal. Unless the human decides to try and teach it to think for itself, with its own goals. Which would be odd for an intentional system. I claim that there's plenty of historical evidence that people fall into this kind of attractor, as the word nirvana indicates (and you'll find similar attractors at the core of many religions). I don't know many people that have actively wasted away due to self-modification of their goals. Hunger strikes is the closest, but not many people fall into it. Our U(x) is quite limited, and easily satisified in the current economy (food, sexual stimulation, warmth, positive social indicators). This leaves the rest of our software to range all over the place as long as these are satifisfied. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
[Humour of some sort] Re: Are rocks conscious? (was RE: [agi] Did this message get completely lost?)
2008/6/4 Bob Mottram [EMAIL PROTECTED]: 2008/6/4 J Storrs Hall, PhD [EMAIL PROTECTED]: What is the rock thinking? T h i s i s w a a a y o f f t o p i c . . . Rocks are obviously superintelligences. By behaving like inert matter and letting us build monuments and gravel pathways out of them they're just lulling us into a false sense of security. Nope they are just seed AIs which started off with the goal of being a rock, and recursively self-improved themselves to fulfil their goal. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Merging - or: Multiplicity
2008/5/27 Mike Tintner [EMAIL PROTECTED]: Will:And you are part of the problem insisting that an AGI should be tested by its ability to learn on its own and not get instruction/help from other agents be they human or other artificial intelligences. I insist[ed] that an AGI should be tested on its ability to solve some *problems* on its own - cross-domain problems - just as we do. Of course, it should learn from others, and get help on other problems, as we do too. But you don't test for that, and as the loebner prize shows you only tend to get what you test for. But if it can't solve many general problems on its own - which seemed OK by you (after setting up your initially appealing submersible problem - solutio interrupta!) - then it's only a narrow AI. I am happy for the baby machine (which is what we will be dealing with to start with) not to be able to solve general problems on its own. Later on I would be disappointed. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Design Phase Announce - VRRM project
2008/5/27 Steve Richfield [EMAIL PROTECTED]: William, This sounds like you should be announcing the analysis phase! Detailed comments follow... Design/research/analysis, call it what you will. On 5/26/08, William Pearson [EMAIL PROTECTED] wrote: VRRM - Virtual Reinforcement Resource Managing Machine Overview This is a virtual machine designed to allow non-catastrophic unconstrained experimentation of programs in a system as close to the hardware as possible. There have been some interesting real machines in the past, e.g. the Burroughs 5000 and 6000 series computers that seldom crashed. When they did, it was presumed to be either an OS or a hardware problem. At Remote Time-Sharing we extended these in a virtual machine, to make a commercial time-sharing system that NEVER EVER crashed after initial debugging. This while servicing secondary schools in the Seattle Area with many hackers, including a very young Bill Gates and Paul Allen. Systems now crash NOT because of the lack of some whiz-bang technology, but because architectural development has been in a state of arrested development for the last ~35 years. It is not just crashes that I worry about but memory corruption and other forms of subversion. This should allow the system to change as much as is possible and needed for the application under consideration. Currently the project expects to go to the operating system level (including experimentation on schedulers and device drivers). A separate sub-system supplies information on how well the experiment is going. The information is made affective by making it a form of credit periodically used to bid for computational system resources and to pass around between programs. This sounds like a problem for real-time applications. In what sense? Expected deployment scenarios - Research and possible small scale applications on the following - Autonomous Self-managing robotics - A Smart operating system that customises itself to the users preferences without extensive knowledge on the users part Language - C Whoops, there are SERIOUS limitations to what can be made reliable in C. C is purely the language for the VRRM, what the programs will be implemented inside the VM is completely up to the people that implement them. Progress Currently I am hacking/designing my own, but I am open to going to a standard machine emulator if that seems easy at any point. I expect to heavily re-factor. I am focussing on the architectural registers, memory space and memory protection first and will get on to the actual instruction set last. This effort would most usefully be merged with the 10K architectures that I have discussed on this forum. Merging disparate concerns might actually result in a design that someone actually constructs. Possibly after I have completed the VRRM and tested it to see if it works how I think it works. But silicon implementation is not on the agenda at the moment. I'm also in parallel trying to design a high level language for this architecture so the internals initial programs can be cross-compiled for it more easily. Does this require a new language, or just some cleverly-named subroutines? A different set of system calls in the least. Some indication of how important the memory is in dynamic memory creation is needed for example. Current Feature plans - Differentiation between transient and long term storage to avoid unwanted disk thrashing Based on the obsolete concept of virtual memory rather than limitless RAM. We don't have limitless RAM, and I won't be implementing virtual memory. snip because I don't have time - Specialised Capability registers as well as floating point and integers Have you seen my/our proposed improvements to IEEE-754 floating point, that itself incorporates a capability register?! Perhaps we should look at a common design? Do you mean capability in the same sense as me? http://en.wikipedia.org/wiki/Capability-based_security Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Merging - or: Multiplicity
2008/5/27 Mike Tintner [EMAIL PROTECTED]: Actually, that's an absurdity. The whole story of evolution tells us that the problems of living in this world for any species of creature/intelligence at any level can only be solved by a SOCIETY of individuals. This whole dimension seems to be entirely missing from AGI. And you are part of the problem insisting that an AGI should be tested by its ability to learn on its own and not get instruction/help from other agents be they human or other artificial intelligences. The social aspect of mimicry has been picked up Ben Goertzel at least in the initial stages of development of his AGI, he may think it will evolve beyond that eventually. I don't think it will, as every mind is capable of getting stuck in a rut (they are attractor states), getting out of that rut is easier with other intelligences to show the way out (themselves getting stuck in different ruts). Societies can get stuck in their own ruts but generally have bigger spaces to explore, so might find their way out in a long time. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
[agi] Design Phase Announce - VRRM project
VRRM - Virtual Reinforcement Resource Managing Machine Overview This is a virtual machine designed to allow non-catastrophic unconstrained experimentation of programs in a system as close to the hardware as possible. This should allow the system to change as much as is possible and needed for the application under consideration. Currently the project expects to go to the operating system level (including experimentation on schedulers and device drivers). A separate sub-system supplies information on how well the experiment is going. The information is made affective by making it a form of credit periodically used to bid for computational system resources and to pass around between programs. Expected deployment scenarios - Research and possible small scale applications on the following - Autonomous Self-managing robotics - A Smart operating system that customises itself to the users preferences without extensive knowledge on the users part Language - C Progress Currently I am hacking/designing my own, but I am open to going to a standard machine emulator if that seems easy at any point. I expect to heavily re-factor. I am focussing on the architectural registers, memory space and memory protection first and will get on to the actual instruction set last. I'm also in parallel trying to design a high level language for this architecture so the internals initial programs can be cross-compiled for it more easily. Current Feature plans - Differentiation between transient and long term storage to avoid unwanted disk thrashing - Unified memory space - Capability based security between programs - Specialised Capability registers as well as floating point and integers - Keyboard, mouse and display virtual devices as well as extensible models for people to build their own Comments and criticisms welcomed. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Different problem types was Re: [agi] AGI and Wiki, was Understanding a sick puppy
2008/5/16 Steve Richfield [EMAIL PROTECTED]: Does anyone else here share my dream of a worldwide AI with all of the knowledge of the human race to support it - built with EXISTING Wikipedia and Dr. Eliza software and a little glue to hold it all together? I'm taking this as a jumping off point to try and describe and expand upon something I have been mulling over whilst reading your messages. I think you and Matt are interested in solving the oracle problem. That is going to one entity for answers to general questions. I am interested in solving more personal problems. That is there are problems that are unique to the individual at each time. The search problem is a good example. To present the optimal search for an individual you must have as much data about the individual as possible. For example if the search engine knew I had been talking to you it would return different results when I searched for Dr. Eliza (assuming google knows anything about your system). As I would not be comfortable with this level of information being known about me, a centralized search oracle will not work (I will have to stop using gmail when AI gets too advanced). The problems essence is finding pertinent information and presenting it at the right time to the user. I shall call it the whisperer class of problems for the moment. I am also strongly interested in Augmented Reality, where knowing when to interrupt you with emails and other communications is an important thing for the system to do. Both types of system are important, I don't think I can do a decent whisperer system with current technologies, including Dr Eliza. Not to denigrate your approach, but to acknowledge that there are more types of problems out there to be solved. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
Re: [agi] Defining understanding (was Re: Newcomb's Paradox)
Matt mahoney: I am not sure what you mean by AGI. I consider a measure of intelligence to be the degree to which goals are satisfied in a range of environments. It does not matter what the goals are. They may seem irrational to you. The goal of a smart bomb is to blow itself up at a given target. I would consider bombs that hit their targets more often to be more intelligent. I consider understanding to mean intelligence in this context. You can't say that a robot that does nothing is unintelligent unless you specify its goals. We may consider intelligence as a measure and AGI as a threshold. AGI is not required for understanding. You can measure the degree to which various search engines understand your query, spam filters understand your email, language translators understand your document, vision systems understand images, intrusion detection systems understand network traffic, etc. Each system was designed with a goal and can be evaluated according to how well that goal is met. AIXI allows us to evaluate intelligence independent of goals. An agent understands its input if it can predict it. This can be measured precisely. I think you are thinking of solomonoff induction, AIXI won't answer your questions unless it has the goal of getting reward from you for answering the question. It will do what it predicts will get it reward, not try and output the end of all strings given to it. I propose prediction as a general test of understanding. For example, do you understand the sequence 0101010101010101 ? If I asked you to predict the next bit and you did so correctly, then I would say you understand it. What would happen if I said, I don't have time for silly games, please stop emailing me. Would you consider that I understood it? If I want to test your understanding of X, I can describe X, give you part of the description, and test if you can predict the rest. If I want to test if you understand a picture, I can cover part of it and ask you to predict what might be there. This only works if my goal includes revealing my understanding to you. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] Self-maintaining Architecture first for AI
2008/5/11 Russell Wallace [EMAIL PROTECTED]: On Sat, May 10, 2008 at 10:10 PM, William Pearson [EMAIL PROTECTED] wrote: It depends on the system you are designing on. I think you can easily create as many types of sand box as you want in programming language E (1) for example. If the principle of least authority (2) is embedded in the system, then you shouldn't have any problems. Sure, I'm talking about much lower-level concepts though. For example, on a system with 8 gigabytes of memory, a candidate program has computed a 5 gigabyte string. For its next operation, it appends that string to itself, thereby crashing the VM due to running out of memory. How _exactly_ do you prevent this from happening (while meeting all the other requirements for an AI platform)? It's a trickier problem than it sounds like it ought to be. I'm starting to mod qemu (it is not a straightforward process) to add capabilities. The VM will have a set amount of memory and if a location outside this memory is referenced, it will throw a page fault inside the VM, not crash it directly. The system will be able to deal with it how it wants to, something smarter than, Oh no I have done a bad memory reference, I must stop all my work and lose everything!!! Hopefully. In the greater scheme of things the model that a computer has unlimited virtual memory has to go as well. Else you might get important things on the hard disk and have much thrashing and ephemera in main memory. You could still make high level abstractions but the virtual memory one is not the one to display to the low level programs. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] Self-maintaining Architecture first for AI
2008/5/10 Richard Loosemore [EMAIL PROTECTED]: This is still quite ambiguous on a number of levels, so would it be possible for you to give us a road map of where the argument is going? At the moment I am not sure what the theme is. That is because I am still ambiguous as to what the later levels are. A fair amount depends upon the application. For example If you are trying to build an AI for jet plane you need to be a lot more careful in how the system explores. I have in my mind a number of things I can't do with current computers and would like to experiment with. 1) A system that has programs that can generate a new learning program dependent upon the inputs it received. The learning program would have inputs either from the outside, the outputs of other learning programs or the value of other learning programs. It would use a variable number of resources. The outputs could be to learning systems, or it could be the generation of a new learning program. Self-maintenance is required to rate and whittle down the learning programs. The learning program could be anything from a SVM to a genetic programming system. 2) A system similar to automatic programming that takes descriptions in a formal language given from the outside and potentially malicious sources and generates a program from them. The language would be sufficient to specify new generative elements in and so extensible in that fashion. A system that cannot maintain itself trying to do this would quickly get swamped by viruses and the like. I'd probably create a hybrid of the two, but I am fully aware that I don't have enough knowledge to discount other approaches. Once I have got the system working to my satisfaction (both experimentally and by showing that being good is evolutionarily stable and I have minimised tragedy of the commons type failures), I'll go on to study more about higher level problems. I have the (slight) hope that other people will pick up my system and take it places I can't currently imagine. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
[agi] Self-maintaining Architecture first for AI
After getting completely on the wrong foot last time I posted something, and not having had time to read the papers I should have. I have decided to try and start afresh and outline where I am coming from. I'll get around to do a proper paper later. There are two possible modes for designing a computer system. I shall characterise them as the active and the passive. The active approach attempts to solve the problem directly where as the passive approach gives a framework under which the problem and related other ones can be more easily solved. The passive approach is generally less efficient but much more reconfigurable. The passive approach is used when there is large number of related possible problems, with a large variety of solutions. Examples of the passive approach are mainly architectures, programming languages and operating systems, with a variety of different goals. They are not always completely passive, for example automatic garbage collection impacts the system somewhat. One illuminating example is the variety of security systems that have been built along this structure. Security in this sense means that the computer system is composed of domains, where not all of them are equally trusted or allowed resources. Now it is possible to set up a passive system designed with security in mind insecurely, by allowing all domains to access every file on the hard disk. Passive systems do not guarantee the solution they are aiming to aid, the most they can do is allow as many possible things to be represented and permit the prevention of certain things. A passive security system allows the prevention of a domain lowering the security of a part of another domain. The set of problems that I intend to help solve is the set of self-maintainanceing computer systems. Self-maintainance is basically reconfiguring the computer to be suited to the environment it finds itself in. The reason why I think it needs to be solved first before AI is attempted is 1) humans self-maintenance, 2) otherwise the very complex computer systems we build for AI will have to be maintained by ourselves which may become increasingly difficult as they approach human level. It is worth noting that I am using AI in the pure sense of being able to solve problems. It is entirely possible to get very high complexity problem solvers (including potentially passing the turing test) that cannot self-maintaince. There a large variety of possible AIs (different bodies/environments/computational resources/goals) as can be seen from the variety of humans and (proto?) intelligences of animals, so a passive approach is not unreasonable. In the case of self-maintaining system, what is that we wish the architecture to prevent? About the only thing we can prevent is a useless program being able to degrade of the system from the current level of operation by taking control of resources. However we also want to enable useful programs to be able to control more resources. To do this we must protect the resources and make sure the correct programs can somehow get the correct resources, the internal programs should do the rest. So it is a resource management problem. Any active force for better levels of operation has to come from the internal programs of the architecture, and once the higher level of operation has been reached the architecture should act as a ratchet to prevent it from slipping down again. Protecting resources amounts to the security problem which we have a fair amount of literature on and the only passive form of resource allocation we know of is a economic system. ... to be continued I might go into further detail about what I mean by resource but that will have to wait for a further post. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] How general can be and should be AGI?
2008/4/27 Dr. Matthias Heger [EMAIL PROTECTED]: Ben Goertzel [mailto:[EMAIL PROTECTED] wrote 26. April 2008 19:54 Yes, truly general AI is only possible in the case of infinite processing power, which is likely not physically realizable. How much generality can be achieved with how much Processing power, is not yet known -- math hasn't advanced that far yet. My point is not only that 'general intelligence without any limits' would need infinite resources of time and memory. This is trivial of course. What I wanted to say is that any intelligence has to be narrow in a sense if it wants be powerful and useful. There must always be strong assumptions of the world deep in any algorithm of useful intelligence. I am probably the one on this list the closest to the position you think AGI means. I would agree. Any algorithms needs to be very specific to be useful. However the *architecture*, of an AGI needs to be general (by this I mean capable of instantiating any TM equivalent function, from input and current state to output and current state). So I think the the lowest level of the system space should be massive as you argue against. However I would not make it a search space, as such, with a fixed method searching it. On its own, it should be passive, however it is be able to have active programs within it. As these are programs on there own they can search the space of possible programs. These programs could search sub spaces of the entire space, or get information from the outside about which subspaces to search. However, there is no limit to which subspaces they do actually search. What makes my approach different to a bog standard computer system, is that it would guide the searching of the programs within it, by acting as reinforcement based ratchet. Those programs with the most reinforcement, that act sensibly, will be able to protect and expand the influence they have over the system. With the right internal programs and environment, this will look as if the system has a goal for what it is trying to become. See this post for more details. http://www.mail-archive.com/agi@v2.listbox.com/msg02892.html Every recursive procedure has to have a non-reducible base and it is clear, that the overall performance and abilities depend crucially on that basic non-reducible procedure. If this procedure is too general, the performance slows exponentially with the space with which this basic procedure works. There are recursive procedures that abandon the base, see for example booting a machine. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] How general can be and should be AGI?
2008/4/26 Dr. Matthias Heger [EMAIL PROTECTED]: How general should be AGI? My answer, as *potentially* general as possible. In a similar fashion that a UTM is as potentially as general as possible, but with more purpose. There are plenty of problems you can define that don't need the halting problem to be impossible to solve, e.g. remember a number with more digits that the potential states of the universe. Some other comments. Have you looked at the literature on neuro plasticity? This wired article is a good introduction. http://www.wired.com/wired/archive/15.04/esp_pr.html Although there are more academic papers out there, a google can find them. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT ARE THE MISSING CONCEPTUAL PIECES IN AGI? --- recent input and responses
On 21/04/2008, Ed Porter [EMAIL PROTECTED] wrote: So when people are given a sentence such as the one you quoted about verbs, pronouns, and nouns, presuming they have some knowledge of most of the words in the sentence, they will understand the concept that verbs are doing words. This is because of the groupings of words that tend to occur in certain syntactical linguistic contexts, the ones that would be most associated with the types of experiences the mind would associates with doing would be largely word senses that are verbs and that the mind's experience and learned patterns most often proceeds by nouns or pronouns. So all this stuff falls out of the magic of spreading activation in a Novamente-like hierarchical experiential memories (with the help of a considerable control structure such as that envisioned for Novamente). Declarative information learned by NL gets projected into the same type of activations in the hierarchical memory How does this happen? What happens when you try and project, This sentence is false. into the activations of the hierarchical memory? And consider that the whole of the english understanding is likely to be in the hierarchical memory. That is the projection must be learnt. as would actual experiences that teaches the same thing, but at least as episodes, and in some patterns generalized from episodes, such declarative information would remain linked to the experience of having been learned from reading or hearing from other humans. So in summary, a Novamete-like system should be able to handle this alleged problem, and at the moment it does not appear to provide an major unanswered conceptual problem. My conversation with Ben about similar subject (words acting on the knowledge of words) didn't get anywhere. The conversation starting here - http://www.mail-archive.com/agi@v2.listbox.com/msg09485.html And I consider him the authority on Novamente-like systems, for now at least. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT ARE THE MISSING CONCEPTUAL PIECES IN AGI?
On 19/04/2008, Ed Porter [EMAIL PROTECTED] wrote: WHAT ARE THE MISSING CONCEPTUAL PIECES IN AGI? I'm not quite sure how to describe it, but this brief sketch will have to do until I get some more time. These may be in some new AI material, but I haven't had the chance to read up much recently. Linguistic information and other non-inductive information integrated into learning/modelling strategies, including the learning of linguistic rules. Consider an AI learning chess, it is told in plain english that Knights move two hops in one direction and one hop 90 degrees to that. Now our AI has learnt english so how do we hook this knowledge into our modelling system, so that it can predict when it might lose or take a piece because of the position of a knight? Consider also the sentence, There are words such as verbs, that are doing words, you need to put a pronoun or noun before the verb. People are given this sort of information when learning languages, it seems to help them. How and why does it help them? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
Re: [agi] The resource allocation problem
On 05/04/2008, Vladimir Nesov [EMAIL PROTECTED] wrote: On Sat, Apr 5, 2008 at 12:24 AM, William Pearson [EMAIL PROTECTED] wrote: On 01/04/2008, Vladimir Nesov [EMAIL PROTECTED] wrote: This question supposes a specific kind of architecture, where these things are in some sense separate from each other. I am agnostic to how much things are separate. At any particular time a machine can be doing less or more of each of these things. For example in humans it is quite common to talk of concentration. E.g. I'm sorry I wasn't concentrating on what you said, could you repeat it. Stop thinking about the girl, concentrate on the problem at hand. Do you think this is meaningful? It is in some sense, but you need to distinguish levels of description. Implementation of system doesn't have a thinking-about-the-girl component, Who ever said it did? All I have said is there needs to be the mechanisms for an economy, not exactly what the economic agents are. I don't know what they should be. It is body/environment specific, most likely. but when system obtains certain behaviors, you can say that this process that is going now is a thinking-about-the-girl process. If, along with learning this process, you form a mechanism for moving attention elsewhere, you can evoke that mechanism by, for example, sending a phrase Stop thinking about the girl to sensory input. But these specific mechanisms are learned, what you need as a system designer is provide ways of their formation in general case. You also need a way to decide that something should get more attention than something else. Being told to attend to something is not always enough. Also, your list contained 'reasoning', 'seeing past experiences and how they apply to the current one', 'searching for new ways of doing things' and 'applying each heuristic'. Only in some architectures will these things be explicit parts of system design. I don't have them as explicit parts of system design, I have nothing that people would call a cognitive design at the moment. I am not so interested in thinking at the moment as building a more *useful* system (although under some circumstances a thinking system will be a useful one). From my perspective, it's analogous to adding special machine instructions for handling 'Internet browsing' in general-purpose processor, where browser is just one of thousands of applications that can run on it, and it would be inadequately complex for processor anyway. I'd agree, I'm just adding a very loose economy. Any actor is allowed to exist in an economy, I was just giving some examples of potential ways to separate things. If they don't fit in your system ignore them and add what does fit. You need to ration resources, but these are anonymous modelling resources that don't have inherent 'bicycle-properties' or 'language-processing-properties'. So does the whatever allows your system to differentiate between bicycle and non-bicycle somehow manage to not take up resources when not being used? Some of them happen to correlate with things we want them to, by virtue of being placed in contact with sensory input that can communicate structure of those things. Resources are used to build inference structures within the system that allow it to model hidden processes, which in turn should allow it to achieve its goals. I'm still not seeing why it should model the right hidden processes. Stick your system in the real world, which processes (from other people, the weather, fluid dynamics, itself) should it try and model? Why do some people have a lot more elaborate models of these things than other people? If there are high-level resource allocation rules to be discovered, these rules will look at goals and formed inference structures and determine that certain changes are good for overall performance. What happens if two rules conflict? Which rule wins? What happens if rules can only be discovered experimentally? Discussion of such rules needs at least some notions about makeup of inference process and its relation to goals. I'm not creating rules to determine how resources are distributed. That would not be a free market economy. I agree the creation of the rules will come about when the cognitive system is being designed, but would be local to each agent. Even worse, goals can be implicit in inference system itself and be learned starting from a blank slate, There is no useful system that is a blank slate. All learning systems have bias as you well know, and so have implicit information about the world. I would view an economy as having an implicit goal. The closest thing to an explicit goal for an agent in my economy is, to survive, but it is in no way hard binding. To survive credit is needed to purchase resources (including memory to stay in and processing power to earn more credit), for which you need to please
Re: [agi] The resource allocation problem
On 01/04/2008, Vladimir Nesov [EMAIL PROTECTED] wrote: On Tue, Apr 1, 2008 at 6:30 PM, William Pearson [EMAIL PROTECTED] wrote: The resource allocation problem and why it needs to be solved first How much memory and processing power should you apply to the following things?: Visual Processing Reasoning Sound Processing Seeing past experiences and how they apply to the current one Searching for new ways of doing things Applying each heuristic This question supposes a specific kind of architecture, where these things are in some sense separate from each other. I am agnostic to how much things are separate. At any particular time a machine can be doing less or more of each of these things. For example in humans it is quite common to talk of concentration. E.g. I'm sorry I wasn't concentrating on what you said, could you repeat it. Stop thinking about the girl, concentrate on the problem at hand. Do you think this is meaningful? If they are but aspects of the same process, with modalities integrated parts of reasoning, resources can't be rationed on such a high level. Rather underlying low-level elements should be globally restricted and differentiate to support different high-level processes (so that certain portion of them gets mainly devoted to visual processing, high-level reasoning, language, etc.). It boils down to the same thing. If more low level elements are devoted (neuron-equivalents?) to a task in general it is giving it more potential memory, processing power and bandwidth. How is that decided? In some connectionist systems, I would associate it with the stability-plasticity problem described here in section 6. http://www.cns.bu.edu/Profiles/Grossberg/Gro1987CogSci.pdf The nutshell is if you have new learning, should it overwrite the old? If so, which information, if not, please can I have your infinite memory system. Assuming you are implementing this on a normal computer you can easily see that it all boils down to these resources. In the brain not all elements can work at peak effectiveness at the same time (consider the perils of driving while on the mobile *). So even if you had devoted the elements, then further decisions need to be made at run time. Different regions of elements become the resources to be rationed. For example Short-term memory becomes a resource you have to decide how to use. Also oxygen would seem to be a resource in short supply, within the brain. The question remains the same, how should a system choose what to do or what to be. Will * http://www.bmj.com/cgi/content/abstract/bmj.38537.397512.55v1 --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
[agi] The resource allocation problem
The resource allocation problem and why it needs to be solved first How much memory and processing power should you apply to the following things?: Visual Processing Reasoning Sound Processing Seeing past experiences and how they apply to the current one Searching for new ways of doing things Applying each heuristic Is there one right way of deciding these things when you have limited resources? At time A might you want more reasoning done (while in a debate) and at time B more visual processing (while driving). There is also the long term memory problem, should you remember your first kiss or the first star trek episode you saw. Which is more important? An intelligent system needs to solve this problem for itself, as only it will know what is important for the problems it faces. That is it is a local problem. It also requires resources itself. If resources are tight then very approximate methods of determining how many resources to spend on each activity. Due to this, the resource management should not be algorithmic, but free to adapt to the amount of resources at hand. I'm intent on a economic solution to the problem, where each activity is an economic actor. This approach needs to be at the lowest level because each activity has to be programmed with the knowledge of how to act in an economic setting as well as to perform its job. How much should it pay for the other activities of the the programs around it? I'll attempt to write a paper on this, with proper references (Baum, Mark Miller et Al.) But I would be interested in feedback at this stage, Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] Instead of an AGI Textbook
On 26/03/2008, Ben Goertzel [EMAIL PROTECTED] wrote: Hi all, A lot of students email me asking me what to read to get up to speed on AGI. So I started a wiki page called Instead of an AGI Textbook, http://www.agiri.org/wiki/Instead_of_an_AGI_Textbook#Computational_Linguistics I've decided to go my own way and have started a new annotated text book, trying to link in all the topics I think relevant to my current state of work. http://www.agiri.org/wiki/AACA_Textbook I'll try putting in content in for each of those links. But coding for the architecture is probably more pointful at this point. Once I have it up and running on QEmu, I'll try and devote more time to education. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] Intelligence: a pattern discovery algorithm of scalable complexity.
On 30/03/2008, Kingma, D.P. [EMAIL PROTECTED] wrote: Although I symphathize with some of Hawkin's general ideas about unsupervised learning, his current HTM framework is unimpressive in comparison with state-of-the-art techniques such as Hinton's RBM's, LeCun's convolutional nets and the promising low-entropy coding variants. But it should be quite clear that such methods could eventually be very handy for AGI. For example, many of you would agree that a reliable, computationally affordable solution to Vision is a crucial factor for AGI: much of the world's information, even on the internet, is encoded in audiovisual information. Extracting (sub)symbolic semantics from these sources would open a world of learning data to symbolic systems. An audiovisual perception layer generates semantic interpretation on the (sub)symbolic level. How could a symbolic engine ever reason about the real world without access to such information? So a deafblind person couldn't reason about the real world? Put ear muffs and a blind fold on, see what you can figure out about the world around you. Less certainly, but then you could figure out more about the world if you had magnetic sense like pidgeons. Intelligence is not about the modalities of the data you get, it is about the what you do with the data you do get. All of the data on the web is encoded in electronic form, it is only because of our comfort with incoming photons and phonons that it is translated to video and sound. This fascination with A/V is useful, but does not help us figure out the core issues that are holding us up whilst trying to create AGI. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] Intelligence: a pattern discovery algorithm of scalable complexity.
On 30/03/2008, Kingma, D.P. [EMAIL PROTECTED] wrote: Intelligence is not *only* about the modalities of the data you get, but modalities are certainly important. A deafblind person can still learn a lot about the world with taste, smell, and touch, but the senses one has access to defines the limits to the world model one can build. As long as you have one high bandwidth modality you should be able to add on technological gizmos to convert information to that modality, and thus be able to model the phenomenon from that part of the world. Humans manage to convert modalities E.g. http://www.engadget.com/2006/04/25/the-brain-port-neural-tongue-interface-of-the-future/ Using touch on the tongue. We don't do it very well, but that is mainly because we don't have to do it it very often. AIs that are designed to have new modalities added to them, using their major modality of their memory space+interrupts (or other computational modality), may be even more flexible than humans and able to adapt to to a new modality as quickly as a current computer is able to add a new device. If I put on ear muffs and a blind fold right now, I can still reason quite well using touch, since I have access to a world model build using e.g. vision. If you were deafblind and paralysed since your birth, would you have any possibility of spatial reasoning? No, maybe except for some extremely crude genetically coded heuristics. Sure if you don't get any spatial information you won't be able to model spatially. But getting the information is different from having a dedicated modality. My point was that audiovisual is not the only way to get spatial information. It may not even be the best way for what we happen to want to do. So not to get too hung up on any specific modality when discussing intelligence. Sure, you could argue that an intelligence purely based on text, disconnected from the physical world, could be intelligent, but it would have a very hard time reasoning about interaction of entities in the physicial world. It would be unable to understand humans in many aspects: I wouldn't call that generally intelligent. I'm not so much interested in this case, but what about the case where you have a robot with Sonar, Radar and other sensors. But not the normal 2 camera +2 microphone thing people imply when they say audiovisual. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] The Effect of Application of an Idea
On 25/03/2008, Vladimir Nesov [EMAIL PROTECTED] wrote Simple systems can be computationally universal, so it's not an issue in itself. On the other hand, no learning algorithm is universal, there are always distributions that given algorithms will learn miserably. The problem is to find a learning algorithm/representation that has the right kind of bias to implement human-like performance. First a riddle: What can be all learning algorithms, but is none? I'd disagree. Okay simple systems can be computationally universal, but what does that really mean. Computational universality means to be able to represent any computable function, the range and domain of this function are assumed to be from the natural numbers to itself. Most AI formulations when they say that are computationally universal are only talking about function of F: I → O where I is the input and O is the output. These include the formulations of neural networks/GA etc that I have seen. However there are lots of interesting programs in computers that do not map the input to the output. Humans also do not just map the input to the output, we also think, ruminate, model and remember. This does not affect the range of functions from the input to the output, but it does change how quickly they can be moved between. What I am interested in is in systems where the ranges and domains of the functions are entities inside the system. That is the F: I → S, F: S → O, and F: S→ S are important and should be potentially computationally universal. Where S is the internal memory of the system. This allows the system to be all possible learning algorithms (although only one at any time), but also it is no algorithm (else F: I x S → S, would be fixed). General purpose desktop computers are these kinds of systems. If they weren't how else could we implement any type of learning system on them? Thus the answer to my riddle. The question I have been trying to answer precisely is how to govern these sorts of systems so they roughly do what you want, without you having to give precise instructions. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: I know, I KNOW :-) WAS Re: [agi] The Effect of Application of an Idea
On 26/03/2008, Mark Waser [EMAIL PROTECTED] wrote: First a riddle: What can be all learning algorithms, but is none? A human being! Well my answer was a common PC, which I hope is more illuminating because we know it well. But human being works, as does any future AI design, as far as I am concerned. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] The Effect of Application of an Idea
On 24/03/2008, Jim Bromer [EMAIL PROTECTED] wrote: To try to understand what I am talking about, start by imagining a simulation of some physical operation, like a part of a complex factory in a Sim City kind of game. In this kind of high-level model no one would ever imagine all of the objects should interact in one stereotypical way, different objects would interact with other objects in different kinds of ways. And no one would imagine that the machines that operated on other objects in the simulation were not also objects in their own right. For instance the machines used in production might require the use of other machines to fix or enhance them. And the machines might produce or operate on objects that were themselves machines. When you think about a simulation of some complicated physical systems it becomes very obvious that different kinds of objects can have different effects on other objects. And yet, when it comes to AI, people go on an on about systems that totally disregard this seemingly obvious divergence of effect that is so typical of nature. Instead most theories see insight as if it could be funneled through some narrow rational system or other less rational field operations where the objects of the operations are only seen as the ineffective object of the pre-defined operations of the program. How would this differ from the sorts of computational systems I have been muttering about? Where you have an architecture where an active bit of code or program is equivalent to an object in the above paragraph. Also have a look at Eurisko by Doug Lenat. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
Re: [agi] Flies Neural Networks
On 16/03/2008, Ed Porter [EMAIL PROTECTED] wrote: I am not an expert on neural nets, but from my limited understanding it is far from clear exactly what the new insight into neural nets referred to in this article is, other than that timing neuron firings is important in the brain, which is something multiple people have been saying for years. So I fail to understand what is new here, other than a reiteration of the view that most of the traditional neural net models used in machine learning are gross simplifications of the neural nets in the brain, particularly with regard to their simplification of the temporal complexity of brain networks. Disclaimer: I am not a neuro scientist either. The spiking neuron model, as I understand it, has the neuron when it is activated give not one signal but a series of signals. This was called a spike train. It was thought that the exact timing of this series of signals was unimportant, just the number and so the neurons just acted as integrators getting an average value. So if a train was like this A) _|_|_| was thought to be the same as this B) __||_| Having the same number of spikes on average. So people simulating real neural networks (not back prop etc) i.e. computational neuroscientists the sorts of people working on blue brain*, used this as the model for there systems. Now it turns out the spike train is more like a message in a packet rather than a number, so A and B are different signals and may be treated differently. It might also have implications for the computational capacity of the brain. I think the max bandwidth calculations would have gone up. * I don't know what model of neuron they are using, but it is that sort of neural network researcher I am talking about. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
[agi] AGI 08 blogging?
Anyone blogging what they are finding interesting in AGI 08? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Thought experiment on informationally limited systems
On 28/02/2008, YKY (Yan King Yin) [EMAIL PROTECTED] wrote: On 2/28/08, William Pearson [EMAIL PROTECTED] wrote: Note I want something different than computational universality. E.g. Von Neumann architectures are generally programmable, Harvard architectures aren't. As they can't be reprogrammed at run time. It seems that you want to build the AGI from the programming level. This is in contrast to John MacCarthy's declarative paradigm. Your approach offers more flexibility (perhaps maximum flexibility), but may not make AGI easier to build. Learning, in your case, is a matter of algorithmic learning. It may be harder / less efficient than logic-based learning. Algorithmic learning is hard. But just because the system is based upon programs as its lowest level representation, does not mean that all learning is going to be algorithmic learning. It is possible to have programs that learn in any fashion within the system. If it makes sense in the system, you could have a logic based learning program. Just that it will be in competition with other learners to see which is the most useful for the system. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Thought experiment on informationally limited systems
On 28/02/2008, Mike Tintner [EMAIL PROTECTED] wrote: You must first define its existing skills, then define the new challenge with some degree of precision - then explain the principles by which it will extend its skills. It's those principles of extension/generalization that are the be-all and end-all, (and NOT btw, as you suggest, any helpful info that the robot will receive - that,sir, is cheating - it has to work these things out for itself - although perhaps it could *ask* for info). Why is that cheating? Would you never give instructions to a child about what to do? Taking instuctions is something that all intelligences need to be able to do, but it should be attempted to be minimised. I'm not saying it should take instructions unquestioningly either, ideally it should figure out whether the instructions you give are any use for it. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Thought experiment on informationally limited systems
On 02/03/2008, Mike Tintner [EMAIL PROTECTED] wrote: Jeez, Will, the point of Artificial General Intelligence is that it can start adapting to an unfamiliar situation and domain BY ITSELF. And your FIRST and only response to the problem you set was to say: I'll get someone to tell it what to do. Nothing we ever do is by ourselves, entirely, we have a wealth of examples to draw from that we have acquired from family/friends and teachers. The situation I described was like throwing a baby into a completely unfamiliar problem, without the wealth of experience we have built up over the years, so some hand holding is to be expected. Also I'm not planning to have a full AI made any time soon, I'm merely laying the ground work, for many other people to work upon. I may get animal level adaptivity/intelligence myself, it depends how quickly I can build the first layer and the tools I need for the next. This is also why I concentrate on the most flexible system possible, I do not wish to constrain the system to do any more than needs be done to achieve my current goal. This goal is to add a way of selecting between the programs within a computer system dependent upon what the system needs to do. It is more fundamental than your cross-over idea, in that it is a lower level phenomenon, but not in the sense it is more important for acting intelligently. IOW you simply avoided the problem and thought only of cheating. What a solution, or merest idea for a solution, must do is tell me how that intelligence will start adapting by itself - will generalize from its existing skills to cross over domains. I'm not building the solution, merely a framework which I think will enable people to build the solution. I think this needs to be done first, in essence I am trying to deal with the develop and acquire skills problem. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Solomonoff Induction Question
On 29/02/2008, Abram Demski [EMAIL PROTECTED] wrote: I'm an undergrad who's been lurking here for about a year. It seems to me that many people on this list take Solomonoff Induction to be the ideal learning technique (for unrestricted computational resources). I'm wondering what justification there is for the restriction to turing-machine models of the universe that Solomonoff Induction uses. Restricting an AI to computable models will obviously make it more realistically manageable. However, Solomonoff induction needs infinite computational resources, so this clearly isn't a justification. There is a gotcha here, at least when you are trying to go to a computable solution (that doesn't require infinite memory). When you go to a FSM (which all our computers are ), then there opens up a whole range of things that are uncomputable for the FSM in question. Including a whole raft of FSM more complex than it. Keeping the same general shape of the system (trying to account for all the detail) means we are likely to overfit, due to trying to model systems that are are too complex for us to be able to model, whilst trying to model for the noise in our systems. This would make the most probable TM more complex than it needs to be, without actually improving its predictive power. Not quite what you worried about, but might add weight to your call to have uncomputablility included in general models of intelligence. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
Re: [agi] Solomonoff Induction Question
On 01/03/2008, Jey Kottalam [EMAIL PROTECTED] wrote: On Sat, Mar 1, 2008 at 3:10 AM, William Pearson [EMAIL PROTECTED] wrote: Keeping the same general shape of the system (trying to account for all the detail) means we are likely to overfit, due to trying to model systems that are are too complex for us to be able to model, whilst trying to model for the noise in our systems. Could you explain this further? I followed what you're saying up to this paragraph. How and why does the overfitting happen? Lets say you have a sine wave on an electrical cable with some noise, you are trying to predict its next value. The noise isn't actually unpredictable, its main components are from radiation from the ionosphere and sun spot activity (or magnetic field variations from the earths crust or whatever). The Turing machine to represent such noise precisely (that is model the sun/ionosphere in enough detail) is more complex than you inductor can represent. However as it is trying to find a program that *exactly* predicts the sequence, sin x will be discarded for some other program that achieves greater accuracy on the training set. Not that it would be bad over fitting, just that it would likely be sin x + a function for pseudo random noise. Real infinite resources Solomonoff induction avoids this by eventually being able to predict what we generally think of as noise. Hope this clears up what I mean. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=95818715-a78a9b Powered by Listbox: http://www.listbox.com
