Re: [agi] If your AGI can't learn to play chess it is no AGI
Matthias, You've presented a straw man argument to criticize embodiment; As a counter-example, in the OCP AGI-development plan, embodiment is not primarily used to provide domains (via artificial environments) in which an AGI might work out abstract problems, directly or comparatively (not to discount the potential utility of this approach in many scenarios), but rather to provide an environment for the grounding of symbols (which include concepts important for doing mathematics), similar to the way in which humans (from infants through to adults) learn through play and also through guided education. 'Abstraction' is so named because it involves generalizing from the specifics of one or more domains (d1, d2), and is useful when it can be applied (with *any* degree of success) to other domains (d3, ...). Virtual embodied interactive learning utilizes virtual objects and their properties as a way of generating these specifics for artificial minds to use to build abstractions, to grok the abstractions of others, and ultimately to build a deep understanding of our reality (yes, 'deep' in this sense is used in a very human-mind-centric way). Of course, few people claim that machine learning with the help of virtually embodied environments is the ONLY way to approach building an AI capable of doing and mathematics (and communicating with humans about mathematics), but it is an approach that has *many* good things going for it, including proving tractable via measurable incremental improvements (even though it is admittedly still at a *very* early stage). -dave On Wed, Oct 22, 2008 at 4:20 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: It seems to me that many people think that embodiment is very important for AGI. For instance some people seem to believe that you can't be a good mathematician if you haven't made some embodied experience. But this would have a rather strange consequence: If you give your AGI a difficult mathematical problem to solve, then it would answer: Sorry, I still cannot solve your problem, but let me walk with my body through the virtual world. Hopefully, I will then understand your mathematical question end even more hopefully I will be able to solve it after some further embodied experience. AGI is the ability to solve different problems in different domains. But such an AGI would need to make experiences in domain d1 in order to solve problems of domain d2. Does this really make sense, if every information necessary to solve problems of d2 is in d2? I think an AGI which has to make experiences in d1 in order to solve a problem of domain d2 which contains everything to solve this problem is no AGI. How should such an AGI know what experiences in d1 are necessary to solve the problem of d2? In my opinion a real AGI must be able to solve a problem of a domain d without leaving this domain if in this domain there is everything to solve this problem. From this we can define a simple benchmark which is not sufficient for AGI but which is **necessary** for a system to be an AGI system: Within the domain of chess there is everything to know about chess. So if it comes up to be a good chess player learning chess from playing chess must be sufficient. Thus, an AGI which is not able to enhance its abilities in chess from playing chess alone is no AGI. Therefore, my first steps in the roadmap towards AGI would be the following: 1. Make a concept for your architecture of your AGI 2. Implement the software for your AGI 3. Try if your AGI is able to become a good chess player from learning in the domain of chess alone. 4. If your AGI can't even learn to play good chess then it is no AGI and it would be a waste of time to make experiences with your system in more complex domains. -Matthias -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscription 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] If your AGI can't learn to play chess it is no AGI
On Wed, Oct 22, 2008 at 3:20 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: It seems to me that many people think that embodiment is very important for AGI. I'm not one of these people, but I at least learn what their arguments. You seem to have made up an argument which you've then knocked down (poorly) and claimed success. Which, BTW, is a very human thing to do and is not something an AGI could learn without being embodied and surrounded by other people who do it ;) Trent --- 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
AW: [agi] If your AGI can't learn to play chess it is no AGI
I see no argument in your text against my main argumentation, that an AGI should be able to learn chess from playing chess alone. This I call straw man replies. My main point against embodiment is just the huge effort for embodiment. You could work for years with this approach and a certain AGI concept until you recognize that it doesn't work. If you apply your AGI concept in a small and even not necessarily AGI-complete domain you would come much faster to a benchmark whether your concept is even worth to make difficult studies with embodiment. Chess is a very good domain for this benchmark because it is very easy to program and it is very difficult to outperform human intelligence in this domain. - Matthias Von: David Hart [mailto:[EMAIL PROTECTED] Gesendet: Mittwoch, 22. Oktober 2008 09:43 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI Matthias, You've presented a straw man argument to criticize embodiment; As a counter-example, in the OCP AGI-development plan, embodiment is not primarily used to provide domains (via artificial environments) in which an AGI might work out abstract problems, directly or comparatively (not to discount the potential utility of this approach in many scenarios), but rather to provide an environment for the grounding of symbols (which include concepts important for doing mathematics), similar to the way in which humans (from infants through to adults) learn through play and also through guided education. 'Abstraction' is so named because it involves generalizing from the specifics of one or more domains (d1, d2), and is useful when it can be applied (with *any* degree of success) to other domains (d3, ...). Virtual embodied interactive learning utilizes virtual objects and their properties as a way of generating these specifics for artificial minds to use to build abstractions, to grok the abstractions of others, and ultimately to build a deep understanding of our reality (yes, 'deep' in this sense is used in a very human-mind-centric way). Of course, few people claim that machine learning with the help of virtually embodied environments is the ONLY way to approach building an AI capable of doing and mathematics (and communicating with humans about mathematics), but it is an approach that has *many* good things going for it, including proving tractable via measurable incremental improvements (even though it is admittedly still at a *very* early stage). -dave On Wed, Oct 22, 2008 at 4:20 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: It seems to me that many people think that embodiment is very important for AGI. For instance some people seem to believe that you can't be a good mathematician if you haven't made some embodied experience. But this would have a rather strange consequence: If you give your AGI a difficult mathematical problem to solve, then it would answer: Sorry, I still cannot solve your problem, but let me walk with my body through the virtual world. Hopefully, I will then understand your mathematical question end even more hopefully I will be able to solve it after some further embodied experience. AGI is the ability to solve different problems in different domains. But such an AGI would need to make experiences in domain d1 in order to solve problems of domain d2. Does this really make sense, if every information necessary to solve problems of d2 is in d2? I think an AGI which has to make experiences in d1 in order to solve a problem of domain d2 which contains everything to solve this problem is no AGI. How should such an AGI know what experiences in d1 are necessary to solve the problem of d2? In my opinion a real AGI must be able to solve a problem of a domain d without leaving this domain if in this domain there is everything to solve this problem. From this we can define a simple benchmark which is not sufficient for AGI but which is *necessary* for a system to be an AGI system: Within the domain of chess there is everything to know about chess. So if it comes up to be a good chess player learning chess from playing chess must be sufficient. Thus, an AGI which is not able to enhance its abilities in chess from playing chess alone is no AGI. Therefore, my first steps in the roadmap towards AGI would be the following: 1. Make a concept for your architecture of your AGI 2. Implement the software for your AGI 3. Try if your AGI is able to become a good chess player from learning in the domain of chess alone. 4. If your AGI can't even learn to play good chess then it is no AGI and it would be a waste of time to make experiences with your system in more complex domains. -Matthias _ agi | https://www.listbox.com/member/archive/303/=now Archives https://www.listbox.com/member/archive/rss/303/ Fehler! Es wurde kein Dateiname angegeben.| https://www.listbox.com/member/?; Modify Your
AW: [agi] If your AGI can't learn to play chess it is no AGI
The restriction is by far not arbitrary. If your AGI is in a space ship or on a distant planet and has to solve the problems in this domain then it has no chance to leave this domain. If this domain contains every information which is necessary to solve the problem then an AGI *must* be able to solve this problem without leaving this domain. Otherwise it would have an essential lack of intelligence and it would not be a real AGI. By the way: Generalization is a mythical thing, because you can never make conclusions from past visited state-action pairs to still unvisited state-action pairs. The reason why this often works are just regularities in the environment. But of course you can not presume that these regularities hold for arbitrary domains. The only thing you can do is to use your past experiences and *hope* they will apply in still unknown domains. - Matthias Von: David Hart [mailto:[EMAIL PROTECTED] Gesendet: Mittwoch, 22. Oktober 2008 11:27 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI I see no reason to impose on AGI the arbitrary restriction that it need posses the ability to learn to perform in a given domain by learning from only within that domain. An AGI should be able to, by definition, adapt itself to function across different and varied domains, using its multi-domain knowledge and experience to improve its performance in any single domain. Choosing a performance metric from only a single domain as a benchmark for an AGI is antithetical to this definition, because, e.g., software that can perform well at chess without being adaptable to other domains is not AGI, but merely narrow AI, and such simplistic single-domain benchmarks can be easily tricked by collections of well orchestrated narrow AI programs. Rather, good benchmarks should be composite benchmarks with component sub-benchmarks spanning multiple and varied domains. A human analogue of the multi-domain AGI concept is nicely paraphrased by Robert A. Heinlein: A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -dave On Wed, Oct 22, 2008 at 7:23 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I see no argument in your text against my main argumentation, that an AGI should be able to learn chess from playing chess alone. This I call straw man replies. My main point against embodiment is just the huge effort for embodiment. You could work for years with this approach and a certain AGI concept until you recognize that it doesn't work. If you apply your AGI concept in a small and even not necessarily AGI-complete domain you would come much faster to a benchmark whether your concept is even worth to make difficult studies with embodiment. Chess is a very good domain for this benchmark because it is very easy to program and it is very difficult to outperform human intelligence in this domain. - Matthias Von: David Hart [mailto:[EMAIL PROTECTED] Gesendet: Mittwoch, 22. Oktober 2008 09:43 An: agi@v2.listbox.com Betreff: Re: [agi] If your AGI can't learn to play chess it is no AGI Matthias, You've presented a straw man argument to criticize embodiment; As a counter-example, in the OCP AGI-development plan, embodiment is not primarily used to provide domains (via artificial environments) in which an AGI might work out abstract problems, directly or comparatively (not to discount the potential utility of this approach in many scenarios), but rather to provide an environment for the grounding of symbols (which include concepts important for doing mathematics), similar to the way in which humans (from infants through to adults) learn through play and also through guided education. 'Abstraction' is so named because it involves generalizing from the specifics of one or more domains (d1, d2), and is useful when it can be applied (with *any* degree of success) to other domains (d3, ...). Virtual embodied interactive learning utilizes virtual objects and their properties as a way of generating these specifics for artificial minds to use to build abstractions, to grok the abstractions of others, and ultimately to build a deep understanding of our reality (yes, 'deep' in this sense is used in a very human-mind-centric way). Of course, few people claim that machine learning with the help of virtually embodied environments is the ONLY way to approach building an AI capable of doing and mathematics (and communicating with humans about mathematics), but it is an approach that has *many* good things going for it, including proving tractable via measurable incremental improvements (even though it is
AW: [agi] If your AGI can't learn to play chess it is no AGI
I do not claim that AGI might not have bias which is equivalent to genes of your example. The point is that AGI is the union set of all AI sets. If I have a certain domain d and a problem p and I know that p can be solved using nothing else than d, then AGI must be able to solve problem p in d otherwise it is not AGI. - Matthias Bob Mottram wrote In the case of humans embodied experience also includes the experience accumulated by our genes over many generations of evolutionary time. This means that even if you personally have not had much embodied experience during your lifetime evolution has shaped your brain wiring ready for that sort of cognition to take place (for instance the ability to perform mental rotations). --- 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] If your AGI can't learn to play chess it is no AGI
On Wed, Oct 22, 2008 at 6:23 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I see no argument in your text against my main argumentation, that an AGI should be able to learn chess from playing chess alone. This I call straw man replies. No-one can learn chess from playing chess alone. Chess is necessarily a social activity. As such, your suggestion isn't even sensible, let alone reasonable. Trent --- 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] If your AGI can't learn to play chess it is no AGI
On Wed, Oct 22, 2008 at 2:10 PM, Trent Waddington [EMAIL PROTECTED] wrote: No-one can learn chess from playing chess alone. Chess is necessarily a social activity. As such, your suggestion isn't even sensible, let alone reasonable. Current AIs learn chess without engaging in social activities ;-). And chess might be a good drosophila for AI, if it's treated as such ( http://www-formal.stanford.edu/jmc/chess.html ). This was uncalled for. -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: [agi] If your AGI can't learn to play chess it is no AGI
If you give the system the rules of chess then it has all which is necessary to know to become a good chess player. It may play against itself or against a common chess program or against humans. - Matthias Trent Waddington [mailto:[EMAIL PROTECTED] wrote No-one can learn chess from playing chess alone. Chess is necessarily a social activity. As such, your suggestion isn't even sensible, let alone reasonable. Trent --- 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: [agi] If your AGI can't learn to play chess it is no AGI
I do not regard chess as important as a drosophila for AI. It would just be a first milestone where we can make a fast proof of concept for an AGI approach. The faster we can sort out bad AGI approaches the sooner we will obtain a successful one. Chess has the advantage to be an easy programmable domain. The domain of chess is not AGI-complete but crucial problems for AGI can be found in chess as well. AGI can be trained automatically against strong chess programs because those engines offer an open API. The performance can be evaluated by elo-ranking, i.e. a common evaluation algorithm for chess players But I do not emphasize performance evaluation too much. The milestone would be passed successfully if the AGI would use a current PC and would be able to beat average human chess players after it has played many thousand chess games against chess programs. It would be a big step towards AGI if someone could build a chess playing program by a learning software which is pattern based and is not inherently build for chess. I think such a program would gain much attention in the community of AI which is also necessary to accelerate the research of AGI. Of course successful experiments with embodiment would probably gain more attention. But the development cycle from concept to experiment would take much longer time with embodiment than with an easy to program and automatically testable chess domain. We should suspect that we still have to go many times through this cycle and therefore it is essential that the cycle should need as few efforts and time as possible. - Matthias Vladimir Nesov [mailto:[EMAIL PROTECTED] wrote Current AIs learn chess without engaging in social activities ;-). And chess might be a good drosophila for AI, if it's treated as such ( http://www-formal.stanford.edu/jmc/chess.html ). This was uncalled for. --- 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
AW: AW: [agi] If your AGI can't learn to play chess it is no AGI
I agree that chess is far from sufficient for AGI. But I have mentioned this already at the beginning of this thread. The important role of chess for AGI could be to rule out bad AGI approaches as fast as possible. Before you go to more complex domains you should consider chess as a first important milestone which helps you not to go a long way towards a dead end with the wrong approach for AGI. If chess is so easy because it is completely described, complete information about state available, fully deterministic etc. then the more important it is that your AGI can learn such an easy task before you try something more difficult. -Matthias Derek Zahn [mailto:[EMAIL PROTECTED] wrote I would agree with this and also with your thesis that a true AGI must be able to learn chess in this way. However, although this ability is necessary it is far from sufficient for AGI, and thinking about AGI from this very narrow perspective seems to me to be a poor way to attack the problem. Very few of the things an AGI must be able to do (as the Heinlein quote points out) are similar to chess -- completely described, complete information about state available, fully deterministic. If you aim at chess you might hit chess but there's no reason that you will achieve anything higher. Still, using chess as a test case may not be useless; a system that produces a convincing story about concept formation in the chess domain (that is, that invents concepts for pinning, pawn chains, speculative sacrifices in exchange for piece mobility, zugzwang, and so on without an identifiable bias toward these things) would at least be interesting to those interested in AGI. Mathematics, though, is interesting in other ways. I don't believe that much of mathematics involves the logical transformations performed in proof steps. A system that invents new fields of mathematics, new terms, new mathematical ideas -- that is truly interesting. Inference control is boring, but inventing mathematical induction, complex numbers, or ring theory -- THAT is AGI-worthy. _ agi | https://www.listbox.com/member/archive/303/=now Archives https://www.listbox.com/member/archive/rss/303/ | https://www.listbox.com/member/?; 7 Modify Your Subscription 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] constructivist issues
You may not like Therefore, we cannot understand the math needed to define our own intelligence., but I'm rather convinced that it's correct. Do you mean to say that there are parts that we can't understand or that the totality is too large to fit and that it can't be cleanly and completely decomposed into pieces (i.e. it's a complex system ;-). Personally, I believe that the foundational pieces necessary to construct/boot-strap an intelligence are eminently understandable (if not even fairly simple) but that the resulting intelligence that a) organically grows from it's interaction with an environment that it can only extract partial, dirty, and ambiguous data and b) does not have the time, computational capability, or data to make itself even remotely consistent past a certain level IS large and complex enough that you will never truly understand it (which is where I have sympathy with Richard Loosemore's arguments -- but don't buy that the interaction of the pieces is necessarily so complex that we can't make broad predictions that are accurate enough to be able to engineer intelligence). If you say parts we can't understand, how do you reconcile that with your statements of yesterday about what general intelligences can learn? --- 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
[agi] A huge amount of math now in standard first-order predicate logic format!
I had not noticed this before, though it was posted earlier this year. Finally Josef Urban translated Mizar into a standard first-order logic format: http://www.cs.miami.edu/~tptp/MizarTPTP/http://www.cs.miami.edu/%7Etptp/MizarTPTP/ Note that there are hyperlinks pointing to the TPTP-ized proofs of each theorem. This is math with **no steps left out of the proofs** ... everything included ... This should be a great resource for AI systems that want to learn about math by reading definitions/theorems/proofs without needing to grok English language or diagrams... Translating this TPTP format into something easily loadable into OpenCog, for example, would not be a big trick Doing useful inference on the data, on the other hand, is another story ;-) To try this in OpenCog, we gotta wait for Joel to finish porting the backward-chainer from NM to OpenCog ... and then, dealing with all this data would be a mighty test of adaptive inference control ;-O ben g -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- 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: AW: [agi] Re: Defining AGI
However, the point I took issue with was your claim that a stupid person could be taught to effectively do science ... or (your later modification) evaluation of scientific results. At the time I originally took exception to your claim, I had not read the earlier portion of the thread, and I still haven't; so I still do not know why you made the claim in the first place. In brief -- You've agreed that even a stupid person is a general intelligence. By do science, I (originally and still) meant the amalgamation that is probably best expressed as a combination of critical thinking and/or the scientific method. My point was a combination of both a) to be a general intelligence, you really must have a domain model and the rudiments of critical thinking/scientific methodology in order to be able to competently/effectively update it and b) if you're a general intelligence, even if you don't need it, you should be able to be taught the rudiments of critical thinking/scientific methodology. Are those points that you would agree with? (A serious question -- and, in particular, if you don't agree, I'd be very interested in why since I'm trying to arrive at a reasonable set of distinctions that define a general intelligence). In typical list fashion, rather than asking what I meant (or, in your case, even having the courtesy to read what came before -- so that you might have *some* chance of understanding what I was trying to get at -- in case my immediate/proximate phrasing was as awkward as I'll freely admit that it was ;-), it effectively turned into an argument past each other when your immediate concept/interpretation of *science = advanced statistical interpretation* hit the blindingly obvious shoals of it's not easy teaching stupid people complicated things (I mean -- seriously, dude --do you *really* think that I'm going to be that far off base? And, if not, why disrupt the conversation so badly by coming in in such a fashion?).. (And I have to say -- As list owner, it would be helpful if you would set a good example of reading threads and trying to understand what people meant rather than immediately coming in and flinging insults and accusations of ignorance e.g. This is obviously spoken by someone who has never . . . . ). So . . . . can you agree with the claim as phrased above? (i.e. What were we disagreeing on again? ;-) Oh, and the original point was part of a discussion about the necessary and sufficient pre-requisites for general intelligence so it made sense to (awkwardly :-) say that a domain model and the rudiments of critical thinking/scientific methodology are a (major but not complete) part of that. - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Tuesday, October 21, 2008 8:51 PM Subject: Re: AW: AW: [agi] Re: Defining AGI Mark W wrote: What were we disagreeing on again? This conversation has drifted into interesting issues in the philosophy of science, most of which you and I seem to substantially agree on. However, the point I took issue with was your claim that a stupid person could be taught to effectively do science ... or (your later modification) evaluation of scientific results. At the time I originally took exception to your claim, I had not read the earlier portion of the thread, and I still haven't; so I still do not know why you made the claim in the first place. ben -- 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: AW: [agi] If your AGI can't learn to play chess it is no AGI
I don't agree at all. The ability to cope with narrow, closed, deterministic environments in an isolated way is VERY DIFFERENT from the ability to cope with a more open-ended, indeterminate environment like the one humans live in Not everything that is a necessary capability of a completed human-level, roughly human-like AGI, is a sensible first step toward a human-level, roughly human-like AGI I'm not saying that making a system that's able to learn chess is a **bad** idea. I am saying that I suspect it's not the best path to AGI. I'm slightly more attracted to the General Gameplaying (GGP) Competition than to a narrow-focus on chess http://games.stanford.edu/ but not so much to that either... I look at it this way. I have a basic understanding of how a roughly human-like AGI mind (with virtual embodiment and language facility) might progress from the preschool level up through the university level, by analogy to human cognitive development. On the other hand, I do not have a very good understanding at all of how a radically non-human-like AGI mind would progress from learn to play chess level to the university level, or to the level of GGP, or robust mathematical theorem-proving, etc. If you have a good understanding of this I'd love to hear it. -- Ben G On Wed, Oct 22, 2008 at 9:47 AM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: I agree that chess is far from sufficient for AGI. But I have mentioned this already at the beginning of this thread. The important role of chess for AGI could be to rule out bad AGI approaches as fast as possible. Before you go to more complex domains you should consider chess as a first important milestone which helps you not to go a long way towards a dead end with the wrong approach for AGI. If chess is so easy because it is completely described, complete information about state available, fully deterministic etc. then the more important it is that your AGI can learn such an easy task before you try something more difficult. -Matthias Derek Zahn [mailto:[EMAIL PROTECTED] wrote I would agree with this and also with your thesis that a true AGI must be able to learn chess in this way. However, although this ability is necessary it is far from sufficient for AGI, and thinking about AGI from this very narrow perspective seems to me to be a poor way to attack the problem. Very few of the things an AGI must be able to do (as the Heinlein quote points out) are similar to chess -- completely described, complete information about state available, fully deterministic. If you aim at chess you might hit chess but there's no reason that you will achieve anything higher. Still, using chess as a test case may not be useless; a system that produces a convincing story about concept formation in the chess domain (that is, that invents concepts for pinning, pawn chains, speculative sacrifices in exchange for piece mobility, zugzwang, and so on without an identifiable bias toward these things) would at least be interesting to those interested in AGI. Mathematics, though, is interesting in other ways. I don't believe that much of mathematics involves the logical transformations performed in proof steps. A system that invents new fields of mathematics, new terms, new mathematical ideas -- that is truly interesting. Inference control is boring, but inventing mathematical induction, complex numbers, or ring theory -- THAT is AGI-worthy. -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/| Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- 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: AW: [agi] If your AGI can't learn to play chess it is no AGI
Matthias Heger: If chess is so easy because it is completely described, complete information about state available, fully deterministic etc. then the more important it is that your AGI can learn such an easy task before you try something more difficult. Chess is not easy. Becoming good at chess is something that most humans never accomplish and none accomplish without years of training in background material. The question is whether chess is representative of the domains we want AGIs to master. I think a case could be made either way. I don't want to be discouraging -- any concrete demonstration of AGI ideas is of great interest, even in formal toy domains. --- 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: AW: [agi] Re: Defining AGI
In brief -- You've agreed that even a stupid person is a general intelligence. By do science, I (originally and still) meant the amalgamation that is probably best expressed as a combination of critical thinking and/or the scientific method. My point was a combination of both a) to be a general intelligence, you really must have a domain model and the rudiments of critical thinking/scientific methodology in order to be able to competently/effectively update it and b) if you're a general intelligence, even if you don't need it, you should be able to be taught the rudiments of critical thinking/scientific methodology. Are those points that you would agree with? The rudiments, yes. But the rudiments are not enough to perform effectively by accepted standards ... e.g. they are not enough to avoid getting fired from your job as a scientist... unless it's a government job ;-) ben --- 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] constructivist issues
It doesn't, because **I see no evidence that humans can understand the semantics of formal system in X in any sense that a digital computer program cannot** I just argued that humans can't understand the totality of any formal system X due to Godel's Incompleteness Theorem but the rest of this is worth addressing . . . . Whatever this mysterious understanding is that you believe you possess, **it cannot be communicated to me in language or mathematics**. Because any series of symbols you give me, could equally well be produced by some being without this mysterious understanding. Excellent! Except for the fact that the probability of the being *continuing* to emit those symbols without this mysterious understanding rapidly approaches zero. So I'm going to argue that understanding *can* effectively be communicated/determined. Arguing otherwise is effectively arguing for vanishingly small probabilities in infinities (and why I hate most arguments involving AIXI as proving *anything* except absolute limits c.f. Matt Mahoney and compression = intelligence). I'm going to continue arguing that understanding exactly equates to having a competent domain model and being able to communicate about it (i.e. that there is no mystery about understanding -- other than not understanding it ;-). Can you describe any possible finite set of finite-precision observations that could provide evidence in favor of the hypothesis that you possess this posited understanding, and against the hypothesis that you are something equivalent to a digital computer? I think you cannot. But I would argue that this is because a digital computer can have understanding (and must and will in order to be an AGI). So, your belief in this posited understanding has nothing to do with science, it's basically a kind of religious faith, it seems to me... '-) If you're assuming that humans have it and computers can't, then I have to strenuously agree. There is no data (that I am aware of) to support this conclusion so it's pure faith, not science. --- 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] constructivist issues
I don't want to diss the personal value of logically inconsistent thoughts. But I doubt their scientific and engineering value. I doesn't seem to make sense that something would have personal value and then not have scientific or engineering value. I can sort of understand science if you're interpreting science looking for the final correct/optimal value but engineering generally goes for either good enough or the best of the currently known available options and anything that really/truly has personal value would seem to have engineering value. --- 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
AW: [agi] A huge amount of math now in standard first-order predicate logic format!
Very useful link. Thanks. -Matthias Von: Ben Goertzel [mailto:[EMAIL PROTECTED] Gesendet: Mittwoch, 22. Oktober 2008 15:40 An: agi@v2.listbox.com Betreff: [agi] A huge amount of math now in standard first-order predicate logic format! I had not noticed this before, though it was posted earlier this year. Finally Josef Urban translated Mizar into a standard first-order logic format: http://www.cs.miami.edu/~tptp/MizarTPTP/ http://www.cs.miami.edu/%7Etptp/MizarTPTP/ Note that there are hyperlinks pointing to the TPTP-ized proofs of each theorem. This is math with **no steps left out of the proofs** ... everything included ... This should be a great resource for AI systems that want to learn about math by reading definitions/theorems/proofs without needing to grok English language or diagrams... Translating this TPTP format into something easily loadable into OpenCog, for example, would not be a big trick Doing useful inference on the data, on the other hand, is another story ;-) To try this in OpenCog, we gotta wait for Joel to finish porting the backward-chainer from NM to OpenCog ... and then, dealing with all this data would be a mighty test of adaptive inference control ;-O ben g -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein _ agi | https://www.listbox.com/member/archive/303/=now Archives https://www.listbox.com/member/archive/rss/303/ | https://www.listbox.com/member/?; 7 Modify Your Subscription 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: AW: [agi] Re: Defining AGI
I'm also confused. This has been a strange thread. People of average and around-average intelligence are trained as lab technicians or database architects every day. Many of them are doing real science. Perhaps a person with down's syndrome would do poorly in one of these largely practical positions. Perhaps. The consensus seems to be that there is no way to make a fool do a scientist's job. But he can do parts of it. A scientist with a dozen fools at hand could be a great deal more effective than a rival with none, whereas a dozen fools on their own might not be expected to do anything at all. So it is complicated. Or maybe another way to rephrase it is combine it with another thread . . . . Any individual piece of science is understandable/teachable to (or my original point -- verifiable or able to be validated by) any general intelligence but the totally of science combined with the world is far too large to . . . . (which is effectively Ben's point) --- 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] constructivist issues
(1) We humans understand the semantics of formal system X. No. This is the root of your problem. For example, replace formal system X with XML. Saying that We humans understand the semantics of XML certainly doesn't work and why I would argue that natural language understanding is AGI-complete (i.e. by the time all the RDF, OWL, and other ontology work is completed -- you'll have an AGI). Any formal system can always be extended *within it's defined syntax* to have any meaning. That is the essence of Godel's Incompleteness Theorem. It's also sort of the basis for my argument with Dr. Matthias Heger. Semantics are never finished except when your model of the world is finished (including all possibilities and infinitudes) so language understanding can't be simple and complete. Personally, rather than starting with NLP, I think that we're going to need to start with a formal language that is a disambiguated subset of English and figure out how to use our world model/knowledge to translate English to this disambiguated subset -- and then we can build from there. (or maybe this makes Heger's argument for him . . . . ;-) --- 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] constructivist issues
On Wed, Oct 22, 2008 at 10:51 AM, Mark Waser [EMAIL PROTECTED] wrote: I don't want to diss the personal value of logically inconsistent thoughts. But I doubt their scientific and engineering value. I doesn't seem to make sense that something would have personal value and then not have scientific or engineering value. Come by the house, we'll drop some acid together and you'll be convinced ;-) --- 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] constructivist issues
Well, if you are a computable system, and if by think you mean represent accurately and internally then you can only think that odd thought via being logically inconsistent... ;-) True -- but why are we assuming *internally*? Drop that assumption as Charles clearly did and there is no problem. It's like infrastructure . . . . I don't have to know all the details of something to use it under normal circumstances though I frequently need to know the details is I'm doing something odd with it or looking for extreme performance and I definitely need to know the details if I'm diagnosing/fixing/debugging it -- but I can always learn them as I go . . . . - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Tuesday, October 21, 2008 11:26 PM Subject: Re: [agi] constructivist issues Well, if you are a computable system, and if by think you mean represent accurately and internally then you can only think that odd thought via being logically inconsistent... ;-) On Tue, Oct 21, 2008 at 11:23 PM, charles griffiths [EMAIL PROTECTED] wrote: I disagree, and believe that I can think X: This is a thought (T) that is way too complex for me to ever have. Obviously, I can't think T and then think X, but I might represent T as a combination of myself plus a notebook or some other external media. Even if I only observe part of T at once, I might appreciate that it is one thought and believe (perhaps in error) that I could never think it. I might even observe T in action, if T is the result of billions of measurements, comparisons and calculations in a computer program. Isn't it just like thinking This is an image that is way too detailed for me to ever see? Charles Griffiths --- On Tue, 10/21/08, Ben Goertzel [EMAIL PROTECTED] wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] constructivist issues To: agi@v2.listbox.com Date: Tuesday, October 21, 2008, 7:56 PM I am a Peircean pragmatist ... I have no objection to using infinities in mathematics ... they can certainly be quite useful. I'd rather use differential calculus to do calculations, than do everything using finite differences. It's just that, from a science perspective, these mathematical infinities have to be considered finite formal constructs ... they don't existP except in this way ... I'm not going to claim the pragmatist perspective is the only subjectively meaningful one. But so far as I can tell it's the only useful one for science and engineering... To take a totally different angle, consider the thought X = This is a thought that is way too complex for me to ever have Can I actually think X? Well, I can understand the *idea* of X. I can manipulate it symbolically and formally. I can reason about it and empathize with it by analogy to A thought that is way too complex for my three-year-old past-self to have ever had , and so forth. But it seems I can't ever really think X, except by being logically inconsistent within that same thought ... this is the Godel limitation applied to my own mind... I don't want to diss the personal value of logically inconsistent thoughts. But I doubt their scientific and engineering value. -- Ben G On Tue, Oct 21, 2008 at 10:43 PM, Abram Demski [EMAIL PROTECTED] wrote: Ben, How accurate would it be to describe you as a finitist or ultrafinitist? I ask because your view about restricting quantifiers seems to reject even the infinities normally allowed by constructivists. --Abram --- 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 -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] Nothing will ever be attempted if all possible objections must be first overcome - Dr Samuel Johnson agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] Nothing will ever be attempted if all possible objections must be first overcome - Dr
[agi] Fun with first-order inference in OpenCog ...
http://brainwave.opencog.org/ -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- 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] constructivist issues
I disagree, and believe that I can think X: This is a thought (T) that is way too complex for me to ever have. Obviously, I can't think T and then think X, but I might represent T as a combination of myself plus a notebook or some other external media. Even if I only observe part of T at once, I might appreciate that it is one thought and believe (perhaps in error) that I could never think it. I might even observe T in action, if T is the result of billions of measurements, comparisons and calculations in a computer program. Isn't it just like thinking This is an image that is way too detailed for me to ever see? Excellent! This is precisely how I feel about intelligence . . . . (and why we *can* understand it even if we can't hold the totality of it -- or fully predict it -- sort of like the weather :-) --- 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] constructivist issues
You have not convinced me that you can do anything a computer can't do. And, using language or math, you never will -- because any finite set of symbols you can utter, could also be uttered by some computational system. -- Ben G Can we pin this somewhere? (Maybe on Penrose? ;-) --- 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] constructivist issues
The problem is to gradually improve overall causal model of environment (and its application for control), including language and dynamics of the world. Better model allows more detailed experience, and so through having a better inbuilt model of an aspect of environment, such as language, it's possible to communicate richer description of other aspects of environment. But it's not obvious that bandwidth of experience is the bottleneck here. No, but nor is it obvious that this *isn't* one of the major bottlenecks... It's probably just limitations of the cognitive algorithm that simply can't efficiently improve its model, and so feeding it more experience through tricks like this is like trying to get a hundredfold speedup in the O(log(log(n))) algorithm by feeding it more hardware. Hard to say... Remember, we humans have a load of evolved inductive bias for understanding human language ... AGI's don't ... so using Lojban to talk to an AGI could be a way to partly make up for this deficit in inductive bias... It should be possible to get a proof-of-concept level results about efficiency without resorting to Cycs and Lojbans, and after that they'll turn out to be irrelevant. Cyc and Lojban are not comparable, one is a knowledge-base, the other is a language Cyc-L and Lojban are more closely comparable, though still very different because Lojban allows for more ambiguity (as well as Cyc-L level precision, depending on speaker's choice) ... and of course Lojban is intended for interactive conversation rather than knowledge entry ben g --- 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
AW: AW: [agi] If your AGI can't learn to play chess it is no AGI
Ben wrote: The ability to cope with narrow, closed, deterministic environments in an isolated way is VERY DIFFERENT from the ability to cope with a more open-ended, indeterminate environment like the one humans live in These narrow, closed, deterministic domains are *subsets* of what AGI is intended to do and what humans can do. Chess can be learned by young children. Not everything that is a necessary capability of a completed human-level, roughly human-like AGI, is a sensible first step toward a human-level, roughly human-like AGI This is surely true. But let's say someone wants to develop a car. Doesn't it makes sense first to develop and test its essential parts before I put everything together and go to the road? I think chess is a good testing area because in the domain of chess there are too many situations to consider them all. This is a very typical and very important problem of human environments as well. On the other hand there are patterns in chess which can be learned and which makes life less complex. This is the second analogy to human environments. Therefore the domain of chess is not so different. It contains an important subset of typical problems for human-level AI. And if you want to solve the complex problem to build AGI then you cannot avoid the task of solving every single of its sub problems. If your system sees no patterns in chess, then I would doubt whether it is really suitable for AGI. I'm not saying that making a system that's able to learn chess is a **bad** idea. I am saying that I suspect it's not the best path to AGI. Ok. I'm slightly more attracted to the General Gameplaying (GGP) Competition than to a narrow-focus on chess http://games.stanford.edu/ http://games.stanford.edu/ but not so much to that either... I look at it this way. I have a basic understanding of how a roughly human-like AGI mind (with virtual embodiment and language facility) might progress from the preschool level up through the university level, by analogy to human cognitive development. On the other hand, I do not have a very good understanding at all of how a radically non-human-like AGI mind would progress from learn to play chess level to the university level, or to the level of GGP, or robust mathematical theorem-proving, etc. If you have a good understanding of this I'd love to hear it. Ok. I do not say that your approach is wrong. In fact I think it is very interesting and ambitious. But as you think that my approach is not the best one I think that your approach is not the best one. Probably, the discussion could be endless. And probably you already have invested too much effort in your approach that you really can consider to change it. I hope you are right because I would be very happy to see the first AGI soon, regardless who will build it and regardless which concept is used. -Matthias --- 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] constructivist issues
IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic English) actually *favored* the small tremendously over-used/ambiguous words (because you got so much more bang for the buck with them). Try only using big unambiguous words and see if you still have the same opinion. If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. My first reaction is . . . . Take a Lojban dictionary and see if you can come up with an unambiguous English word or very short phrase for each Lojban word. If you can do it, my approach will work and will have the advantage that the output can be read by anyone (i.e. it's the equivalent of me having done it in Lojban and then added a Lojban - English translation on the end) though the input is still *very* problematical (thus the need for a semantically-driven English-subset translator). If you can't do it, then my approach won't work. Can you do it? Why or why not? If you can, do you still believe that my approach won't work? Oh, wait . . . . a Lojban-to-English dictionary *does* attempt to come up with an unambiguous English word or very short phrase for each Lojban word. :-) Actually, h . . . . a Lojban dictionary would probably help me focus my efforts a bit better and highlight things that I may have missed . . . . do you have a preferred dictionary or resource? (Google has too many for me to do a decent perusal quickly) - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 11:11 AM Subject: Re: [agi] constructivist issues Personally, rather than starting with NLP, I think that we're going to need to start with a formal language that is a disambiguated subset of English IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. ben g -- 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] constructivist issues
(joke) What? You don't love me any more? /thread - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 11:11 AM Subject: Re: [agi] constructivist issues (joke) On Wed, Oct 22, 2008 at 11:11 AM, Ben Goertzel [EMAIL PROTECTED] wrote: On Wed, Oct 22, 2008 at 10:51 AM, Mark Waser [EMAIL PROTECTED] wrote: I don't want to diss the personal value of logically inconsistent thoughts. But I doubt their scientific and engineering value. I doesn't seem to make sense that something would have personal value and then not have scientific or engineering value. Come by the house, we'll drop some acid together and you'll be convinced ;-) -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein -- 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] constructivist issues
Come by the house, we'll drop some acid together and you'll be convinced ;-) Been there, done that. Just because some logically inconsistent thoughts have no value doesn't mean that all logically inconsistent thoughts have no value. Not to mention the fact that hallucinogens, if not the subsequently warped thoughts, do have the serious value of raising your mental Boltzmann temperature. - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 11:11 AM Subject: Re: [agi] constructivist issues On Wed, Oct 22, 2008 at 10:51 AM, Mark Waser [EMAIL PROTECTED] wrote: I don't want to diss the personal value of logically inconsistent thoughts. But I doubt their scientific and engineering value. I doesn't seem to make sense that something would have personal value and then not have scientific or engineering value. Come by the house, we'll drop some acid together and you'll be convinced ;-) -- 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: AW: [agi] Language learning (was Re: Defining AGI)
--- On Tue, 10/21/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: Sorry, but this was no proof that a natural language understanding system is necessarily able to solve the equation x*3 = y for arbitrary y. 1) You have not shown that a language understanding system must necessarily(!) have made statistical experiences on the equation x*3 =y. A language model is a probability distribution P over text of human origin. If you can compute P(x) for given text string x, then you can pass the Turing test because for any question Q and answer A you can compute P(A|Q) = P(QA)/P(Q) using the same distribution that a human would use to answer the question. This includes any math questions that the average human could answer. 2) you give only a few examples. For a proof of the claim, you have to prove it for every(!) y. You originally allowed *any* y. To quote your earlier email: For instance, I doubt that anyone can prove that any system which understands natural language is necessarily able to solve the simple equation x *3 = y for a given y. Anyway I did the experiment for y = 12. You can try the experiment for other values of y if you wish. Let me know what happens. 3) you apply rules such as 5 * 7 = 35 - 35 / 7 = 5 but you have not shown that 3a) that a language understanding system necessarily(!) has this rules 3b) that a language understanding system necessarily(!) can apply such rules It must have the rules and apply them to pass the Turing test. -- Matt Mahoney, [EMAIL PROTECTED] --- 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] constructivist issues
On Wed, Oct 22, 2008 at 7:47 PM, Ben Goertzel [EMAIL PROTECTED] wrote: The problem is to gradually improve overall causal model of environment (and its application for control), including language and dynamics of the world. Better model allows more detailed experience, and so through having a better inbuilt model of an aspect of environment, such as language, it's possible to communicate richer description of other aspects of environment. But it's not obvious that bandwidth of experience is the bottleneck here. No, but nor is it obvious that this *isn't* one of the major bottlenecks... My intuition is that it's very easy to steadily increase bandwidth of experience, the more you know the more you understand. If you start from simple sensors/actuators (or even chess or Go), progress is gradual and open-ended. It's probably just limitations of the cognitive algorithm that simply can't efficiently improve its model, and so feeding it more experience through tricks like this is like trying to get a hundredfold speedup in the O(log(log(n))) algorithm by feeding it more hardware. Hard to say... Remember, we humans have a load of evolved inductive bias for understanding human language ... AGI's don't ... so using Lojban to talk to an AGI could be a way to partly make up for this deficit in inductive bias... Any language at all is a way of increasing experiential bandwidth about environment. If bandwidth isn't essential, bootstrapping this process through a language is equally irrelevant. At some point, however inefficiently, language can be learned if system allows open-ended learning. This is a question of not doing premature optimization of a program that is not even designed yet, not talking about being implemented and profiled. It should be possible to get a proof-of-concept level results about efficiency without resorting to Cycs and Lojbans, and after that they'll turn out to be irrelevant. Cyc and Lojban are not comparable, one is a knowledge-base, the other is a language Cyc-L and Lojban are more closely comparable, though still very different because Lojban allows for more ambiguity (as well as Cyc-L level precision, depending on speaker's choice) ... and of course Lojban is intended for interactive conversation rather than knowledge entry (as tools towards improving bandwidth of experience, they do the same thing) -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] constructivist issues
Too many responses for me to comment on everything! So, sorry to those I don't address... Ben, When I claim a mathematical entity exists, I'm saying loosely that meaningful statements can be made using it. So, I think meaning is more basic. I mentioned already what my current definition of meaning is: a statement is meaningful if it is associated with a computable rule of deduction that it can use to operate on other (meaningful) statements. This is in contrast to positivist-style definitions of meaning, that would instead require a computable test of truth and/or falsehood. So, a statement is meaningful if it has procedural deductive meaning. We *understand* a statement if we are capable of carrying out the corresponding deductive procedure. A statement is *true* if carrying out that deductive procedure only produces more true statements. We *believe* a statement if we not only understand it, but proceed to apply its deductive procedure. There is of course some basic level of meaningful statements, such as sensory observations, so that this is a working recursive definition of meaning and truth. By this definition of meaning, any statement in the arithmetical hierarchy is meaningful (because each statement can be represented by computable consequences on other statements in the arithmetical hierarchy). I think some hyperarithmetical truths are captured as well. I am more doubtful about it capturing anything beyond the first level of the analytic hierarchy, and general set-theoretic discourse seems far beyond its reach. Regardless, the definition of meaning makes a very large number of uncomputable truths nonetheless meaningful. Russel, I think both Ben and I would approximately agree with everything you said, but that doesn't change our disagreeing with each other :). Mark, Good call... I shouldn't be talking like I think it is terrifically unlikely that some more-intelligent alien species would find humans mathematically crude. What I meant was, it seems like humans are logically complete in some sense. In practice we are greatly limited by memory and processing speed and so on; but I *don't* think we're limited by lacking some important logical construct. It would be like us discovering some alien species whose mathematicians were able to understand each individual case of mathematical induction, but were unable to comprehend the argument for accepting it as a general principle, because they lack the abstraction. Something like that is what I find implausible. --Abram --- 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] constructivist issues
This is the standard Lojban dictionary http://jbovlaste.lojban.org/ I am not so worried about word meanings, they can always be handled via reference to WordNet via usages like run_1, run_2, etc. ... or as you say by using rarer, less ambiguous words Prepositions are more worrisome, however, I suppose they can be handled in a similar way, e.g. by defining an ontology of preposition meanings like with_1, with_2, with_3, etc. In fact we had someone spend a couple months integrating existing resources into a preposition-meaning ontology like this a while back ... the so-called PrepositionWordNet ... or as it eventually came to be called the LARDict or LogicalArgumentRelationshipDictionary ... I think it would be feasible to tweak RelEx to recognize these sorts of subscripts, and in this way to recognize a highly controlled English that would be unproblematic to map semantically... We would then say e.g. I ate dinner with_2 my fork I live in_2 Maryland I have lived_6 for_3 41 years (where I suppress all _1's, so that e.g. ate means ate_1) Because, RelEx already happily parses the syntax of all simple sentences, so the only real hassle to deal with is disambiguation. We could use similar hacking for reference resolution, temporal sequencing, etc. The terrorists_v1 robbed_v2 my house. After that_v2, the jerks_v1 urinated in_3 my yard. I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. -- Ben G On Wed, Oct 22, 2008 at 12:00 PM, Mark Waser [EMAIL PROTECTED] wrote: IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic English) actually *favored* the small tremendously over-used/ambiguous words (because you got so much more bang for the buck with them). Try only using big unambiguous words and see if you still have the same opinion. If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. My first reaction is . . . . Take a Lojban dictionary and see if you can come up with an unambiguous English word or very short phrase for each Lojban word. If you can do it, my approach will work and will have the advantage that the output can be read by anyone (i.e. it's the equivalent of me having done it in Lojban and then added a Lojban - English translation on the end) though the input is still *very* problematical (thus the need for a semantically-driven English-subset translator). If you can't do it, then my approach won't work. Can you do it? Why or why not? If you can, do you still believe that my approach won't work? Oh, wait . . . . a Lojban-to-English dictionary *does* attempt to come up with an unambiguous English word or very short phrase for each Lojban word. :-) Actually, h . . . . a Lojban dictionary would probably help me focus my efforts a bit better and highlight things that I may have missed . . . . do you have a preferred dictionary or resource? (Google has too many for me to do a decent perusal quickly) - Original Message - *From:* Ben Goertzel [EMAIL PROTECTED] *To:* agi@v2.listbox.com *Sent:* Wednesday, October 22, 2008 11:11 AM *Subject:* Re: [agi] constructivist issues Personally, rather than starting with NLP, I think that we're going to need to start with a formal language that is a disambiguated subset of English IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. ben g -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- *agi* | Archives https://www.listbox.com/member/archive/303/=now https://www.listbox.com/member/archive/rss/303/ | Modifyhttps://www.listbox.com/member/?;Your Subscription http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able
Re: [agi] constructivist issues
So, a statement is meaningful if it has procedural deductive meaning. We *understand* a statement if we are capable of carrying out the corresponding deductive procedure. A statement is *true* if carrying out that deductive procedure only produces more true statements. We *believe* a statement if we not only understand it, but proceed to apply its deductive procedure. OK, then according to your definition, Godel's Theorem says that if humans are computable there are some things that we cannot understand ... just as, for any computer program, there are some things it can't understand. It just happens that according to your definition, a computer system can understand some fabulously uncomputable entities. But there's no contradiction there. Just like a human can, a digital theorem prover can understand some uncomputable entities in the sense you specify... ben g --- 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] If your AGI can't learn to play chess it is no AGI
Not everything that is a necessary capability of a completed human-level, roughly human-like AGI, is a sensible first step toward a human-level, roughly human-like AGI This is surely true. But let's say someone wants to develop a car. Doesn't it makes sense first to develop and test its essential parts before I put everything together and go to the road? Yes, and we are of course doing that I think chess is a good testing area I strongly disagree... If your system sees no patterns in chess, then I would doubt whether it is really suitable for AGI. I strongly suspect that OpenCog ... once more of the NM tools are ported to it (e.g. the completion of the backward chainer port) ... could learn to play chess legally but not very well. To get it to play really well would probably require either a lot of specialized hacking with inference control, or a broader AGI approach going beyond the chess domain... or a lot more advancement of the learning mechanisms (along lines already specified in the OCP design) To me, teaching OpenCog to play chess poorly would prove almost nothing. And getting it to play chess well via tailoring the inference control mechanisms would prove little that's relevant to AGI, though it would be cool. Ok. I do not say that your approach is wrong. In fact I think it is very interesting and ambitious. But as you think that my approach is not the best one I think that your approach is not the best one. Probably, the discussion could be endless. And probably you already have invested too much effort in your approach that you really can consider to change it. I hope you are right because I would be very happy to see the first AGI soon, regardless who will build it and regardless which concept is used. I would change my approach if I thought there were a better one. But you haven't convinced me, just as I haven't convinced you ;-) Anyway, to take your approach I would not need to change my AGI design at all: OCP could be pursued in the domain of learning to play chess. I just don't think that's the best choice. BTW, if I were going to pursue a board game I'd choose Go not chess ... at least it hasn't been solved by narrow-AI very well yet ... so a really good OpenCog-based Go program would have more sex appeal ... there has not been a Deep Blue of Go My son is a good Go player so maybe I'll talk him into trying this one day ;-) ben g --- 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] constructivist issues
Mark, The way you invoke Godel's Theorem is strange to me... perhaps you have explained your argument more fully elsewhere, but as it stands I do not see your reasoning. --Abram On Wed, Oct 22, 2008 at 12:20 PM, Mark Waser [EMAIL PROTECTED] wrote: It looks like all this disambiguation by moving to a more formal language is about sweeping the problem under the rug, removing the need for uncertain reasoning from surface levels of syntax and semantics, to remember about it 10 years later, retouch the most annoying holes with simple statistical techniques, and continue as before. That's an excellent criticism but not the intent. Godel's Incompleteness Theorem means that you will be forever building . . . . All that disambiguation does is provides a solid, commonly-agreed upon foundation to build from. English and all natural languages are *HARD*. They are not optimal for simple understanding particularly given the realms we are currently in and ambiguity makes things even worse. Languages have so many ambiguities because of the way that they (and concepts) develop. You see something new, you grab the nearest analogy and word/label and then modify it to fit. That's why you then later need the much longer words and very specific scientific terms and names. Simple language is what you need to build the more specific complex language. Having an unambiguous constructed language is simply a template or mold that you can use as scaffolding while you develop NLU. Children start out very unambiguous and concrete and so should we. (And I don't believe in statistical techniques unless you have the resources of Google or AIXI) --- 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] constructivist issues
What I meant was, it seems like humans are logically complete in some sense. In practice we are greatly limited by memory and processing speed and so on; but I *don't* think we're limited by lacking some important logical construct. It would be like us discovering some alien species whose mathematicians were able to understand each individual case of mathematical induction, but were unable to comprehend the argument for accepting it as a general principle, because they lack the abstraction. Something like that is what I find implausible. I like the phrase logically complete. The way that I like to think about it is that we have the necessary seed of whatever intelligence/competence is that can be logically extended to cover all circumstances. We may not have the personal time or resources to do so but given infinite time and resources there is no block on the path from what we have to getting there. Note, however, that it is my understanding that a number of people on this list do not agree with this statement (feel free to chime in with you reasons why folks). - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 12:20 PM Subject: Re: [agi] constructivist issues Too many responses for me to comment on everything! So, sorry to those I don't address... Ben, When I claim a mathematical entity exists, I'm saying loosely that meaningful statements can be made using it. So, I think meaning is more basic. I mentioned already what my current definition of meaning is: a statement is meaningful if it is associated with a computable rule of deduction that it can use to operate on other (meaningful) statements. This is in contrast to positivist-style definitions of meaning, that would instead require a computable test of truth and/or falsehood. So, a statement is meaningful if it has procedural deductive meaning. We *understand* a statement if we are capable of carrying out the corresponding deductive procedure. A statement is *true* if carrying out that deductive procedure only produces more true statements. We *believe* a statement if we not only understand it, but proceed to apply its deductive procedure. There is of course some basic level of meaningful statements, such as sensory observations, so that this is a working recursive definition of meaning and truth. By this definition of meaning, any statement in the arithmetical hierarchy is meaningful (because each statement can be represented by computable consequences on other statements in the arithmetical hierarchy). I think some hyperarithmetical truths are captured as well. I am more doubtful about it capturing anything beyond the first level of the analytic hierarchy, and general set-theoretic discourse seems far beyond its reach. Regardless, the definition of meaning makes a very large number of uncomputable truths nonetheless meaningful. Russel, I think both Ben and I would approximately agree with everything you said, but that doesn't change our disagreeing with each other :). Mark, Good call... I shouldn't be talking like I think it is terrifically unlikely that some more-intelligent alien species would find humans mathematically crude. What I meant was, it seems like humans are logically complete in some sense. In practice we are greatly limited by memory and processing speed and so on; but I *don't* think we're limited by lacking some important logical construct. It would be like us discovering some alien species whose mathematicians were able to understand each individual case of mathematical induction, but were unable to comprehend the argument for accepting it as a general principle, because they lack the abstraction. Something like that is what I find implausible. --Abram --- 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: AW: [agi] Language learning (was Re: Defining AGI)
You make the implicit assumption that a natural language understanding system will pass the turing test. Can you prove this? Furthermore, it is just an assumption that the ability to have and to apply the rules are really necessary to pass the turing test. For these two reasons, you still haven't shown 3a and 3b. By the way: The turing test must convince 30% of the people. Today there is a system which can already convince 25% http://www.sciencedaily.com/releases/2008/10/081013112148.htm -Matthias 3) you apply rules such as 5 * 7 = 35 - 35 / 7 = 5 but you have not shown that 3a) that a language understanding system necessarily(!) has this rules 3b) that a language understanding system necessarily(!) can apply such rules It must have the rules and apply them to pass the Turing test. -- Matt Mahoney, [EMAIL PROTECTED] --- 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: [OpenCog] Re: [agi] constructivist issues
I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. Awesome. Like I said, it's a piece of something that I'm trying currently. If I get positive results, I'm certainly not going to hide the fact. ;-) (or, it could turn into a learning experience like my attempts with Simplified English and Basic English :-) - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Cc: [EMAIL PROTECTED] Sent: Wednesday, October 22, 2008 12:27 PM Subject: [OpenCog] Re: [agi] constructivist issues This is the standard Lojban dictionary http://jbovlaste.lojban.org/ I am not so worried about word meanings, they can always be handled via reference to WordNet via usages like run_1, run_2, etc. ... or as you say by using rarer, less ambiguous words Prepositions are more worrisome, however, I suppose they can be handled in a similar way, e.g. by defining an ontology of preposition meanings like with_1, with_2, with_3, etc. In fact we had someone spend a couple months integrating existing resources into a preposition-meaning ontology like this a while back ... the so-called PrepositionWordNet ... or as it eventually came to be called the LARDict or LogicalArgumentRelationshipDictionary ... I think it would be feasible to tweak RelEx to recognize these sorts of subscripts, and in this way to recognize a highly controlled English that would be unproblematic to map semantically... We would then say e.g. I ate dinner with_2 my fork I live in_2 Maryland I have lived_6 for_3 41 years (where I suppress all _1's, so that e.g. ate means ate_1) Because, RelEx already happily parses the syntax of all simple sentences, so the only real hassle to deal with is disambiguation. We could use similar hacking for reference resolution, temporal sequencing, etc. The terrorists_v1 robbed_v2 my house. After that_v2, the jerks_v1 urinated in_3 my yard. I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. -- Ben G On Wed, Oct 22, 2008 at 12:00 PM, Mark Waser [EMAIL PROTECTED] wrote: IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic English) actually *favored* the small tremendously over-used/ambiguous words (because you got so much more bang for the buck with them). Try only using big unambiguous words and see if you still have the same opinion. If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. My first reaction is . . . . Take a Lojban dictionary and see if you can come up with an unambiguous English word or very short phrase for each Lojban word. If you can do it, my approach will work and will have the advantage that the output can be read by anyone (i.e. it's the equivalent of me having done it in Lojban and then added a Lojban - English translation on the end) though the input is still *very* problematical (thus the need for a semantically-driven English-subset translator). If you can't do it, then my approach won't work. Can you do it? Why or why not? If you can, do you still believe that my approach won't work? Oh, wait . . . . a Lojban-to-English dictionary *does* attempt to come up with an unambiguous English word or very short phrase for each Lojban word. :-) Actually, h . . . . a Lojban dictionary would probably help me focus my efforts a bit better and highlight things that I may have missed . . . . do you have a preferred dictionary or resource? (Google has too many for me to do a decent perusal quickly) - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 11:11 AM Subject: Re: [agi] constructivist issues Personally, rather than starting with NLP, I think that we're going to need to start with a formal language that is a disambiguated subset of English IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g
Re: [agi] constructivist issues
All theorems in the same formal system are equivalent anyways ;-) On Wed, Oct 22, 2008 at 1:03 PM, Abram Demski [EMAIL PROTECTED] wrote: Ben, What, then, do you make of my definition? Do you think deductive consequence is insufficient for meaningfulness? I am not sure exactly where you draw the line as to what is really meaningful (as in finite collections of finite statements about finite-precision measurements) and what is only indirectly meaningful by its usefulness (as in differential calculus). Perhaps any universal statements are only meaningful by usefulness? Also, it seems like when you say Godel's Incompleteness, you mean Tarski's Undefinability? (Can't let the theorems be misused!) About the theorem prover; yes, absolutely, so long as the mathematical entity is understandable by the definition I gave. Unfortunately, I still have some work to do, because as far as I can tell that definition does not explain how uncountable sets are meaningful... (maybe it does and I am just missing something...) --Abram On Wed, Oct 22, 2008 at 12:30 PM, Ben Goertzel [EMAIL PROTECTED] wrote: So, a statement is meaningful if it has procedural deductive meaning. We *understand* a statement if we are capable of carrying out the corresponding deductive procedure. A statement is *true* if carrying out that deductive procedure only produces more true statements. We *believe* a statement if we not only understand it, but proceed to apply its deductive procedure. OK, then according to your definition, Godel's Theorem says that if humans are computable there are some things that we cannot understand ... just as, for any computer program, there are some things it can't understand. It just happens that according to your definition, a computer system can understand some fabulously uncomputable entities. But there's no contradiction there. Just like a human can, a digital theorem prover can understand some uncomputable entities in the sense you specify... ben g 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/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- 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] constructivist issues
Also, I don't prefer to define meaning the way you do ... so clarifying issues with your definition is your problem, not mine!! On Wed, Oct 22, 2008 at 1:03 PM, Abram Demski [EMAIL PROTECTED] wrote: Ben, What, then, do you make of my definition? Do you think deductive consequence is insufficient for meaningfulness? I am not sure exactly where you draw the line as to what is really meaningful (as in finite collections of finite statements about finite-precision measurements) and what is only indirectly meaningful by its usefulness (as in differential calculus). Perhaps any universal statements are only meaningful by usefulness? Also, it seems like when you say Godel's Incompleteness, you mean Tarski's Undefinability? (Can't let the theorems be misused!) About the theorem prover; yes, absolutely, so long as the mathematical entity is understandable by the definition I gave. Unfortunately, I still have some work to do, because as far as I can tell that definition does not explain how uncountable sets are meaningful... (maybe it does and I am just missing something...) --Abram On Wed, Oct 22, 2008 at 12:30 PM, Ben Goertzel [EMAIL PROTECTED] wrote: So, a statement is meaningful if it has procedural deductive meaning. We *understand* a statement if we are capable of carrying out the corresponding deductive procedure. A statement is *true* if carrying out that deductive procedure only produces more true statements. We *believe* a statement if we not only understand it, but proceed to apply its deductive procedure. OK, then according to your definition, Godel's Theorem says that if humans are computable there are some things that we cannot understand ... just as, for any computer program, there are some things it can't understand. It just happens that according to your definition, a computer system can understand some fabulously uncomputable entities. But there's no contradiction there. Just like a human can, a digital theorem prover can understand some uncomputable entities in the sense you specify... ben g 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/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -- Robert Heinlein --- 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] constructivist issues
Douglas Hofstadter's newest book I Am A Strange Loop (currently available from Amazon for $7.99 - http://www.amazon.com/Am-Strange-Loop-Douglas-Hofstadter/dp/B001FA23HM) has an excellent chapter showing Godel in syntax and semantics. I highly recommend it. The upshot is that while it is easily possible to define a complete formal system of syntax, that formal system can always be used to convey something (some semantics) that is (are) outside/beyond the system -- OR, to paraphrase -- meaning is always incomplete because it can always be added to even inside a formal system of syntax. This is why I contend that language translation ends up being AGI-complete (although bounded subsets clearly don't need to be -- the question is whether you get a usable/useful subset more easily with or without first creating a seed AGI). - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 12:38 PM Subject: Re: [agi] constructivist issues Mark, The way you invoke Godel's Theorem is strange to me... perhaps you have explained your argument more fully elsewhere, but as it stands I do not see your reasoning. --Abram On Wed, Oct 22, 2008 at 12:20 PM, Mark Waser [EMAIL PROTECTED] wrote: It looks like all this disambiguation by moving to a more formal language is about sweeping the problem under the rug, removing the need for uncertain reasoning from surface levels of syntax and semantics, to remember about it 10 years later, retouch the most annoying holes with simple statistical techniques, and continue as before. That's an excellent criticism but not the intent. Godel's Incompleteness Theorem means that you will be forever building . . . . All that disambiguation does is provides a solid, commonly-agreed upon foundation to build from. English and all natural languages are *HARD*. They are not optimal for simple understanding particularly given the realms we are currently in and ambiguity makes things even worse. Languages have so many ambiguities because of the way that they (and concepts) develop. You see something new, you grab the nearest analogy and word/label and then modify it to fit. That's why you then later need the much longer words and very specific scientific terms and names. Simple language is what you need to build the more specific complex language. Having an unambiguous constructed language is simply a template or mold that you can use as scaffolding while you develop NLU. Children start out very unambiguous and concrete and so should we. (And I don't believe in statistical techniques unless you have the resources of Google or AIXI) --- 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/?; 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
RE: [agi] Who is smart enough to answer this question?
Vlad, Thanks for your below reply to my prior email of Tue 10/21/2008 7:08 PM I agreed with most of your reply. There are only two major issues upon which I wanted further confirmation, clarification, or comment. 1. WHY C(N,S) IS DIVIDED BY T(N,S,O) TO FORM A LOWER BOUNDS FOR A(N,S,O) You have stated that C(N,S) / T(N,S,O) is a lower bounds for the value A(N,S,O) where A is the number of sets of length S (which I will also refer to as assemblies) that can be formed from N nodes, where none of the A assemblies overlap the population of any other of such assemblies by an overlap of O or larger. I want to see if my understanding of this is correct. I understand what C(N,S) is. Thanks to your explanations I think I understand what T(N,S,O) is. And as is explained below under heading 2, I think I understand why T(N,S,O) is likely to over count the number of assemblies that impermissibly overlap the population of any allowed assembly that is counted as part of A. But until a few minutes ago I didn't understand the mathematical bases for A = C(N,S) / T(N,S,O) I understood why T(N,S,O) could be subtracted from C(N,S), but not why should be its divisor. Now I think I do. PLEASE CONFIRM IF MY NEW UNDERSTANDING IS CORRECT. For purposes of simplicity, until I state otherwise let us assume there is no multiple counting of excluded assemblies in the calculation of T(N,S,O), and, thus, that T(N,S,O) is exact. As I understand your argument you are saying every time you increase the count of allowable assemblies by 1, you increase the count of unallowable assemblies --- i.e., those having an impermissible overlap --- by T. Thus if you had A allowable assemblies, you would have A x T unallowable assemblies and, thus, A + A x T = C(N,S) This says all the allowable and unallowable sets of length S would equal the total number of different possible sets of length S that can be made from N elements. If one solves this for A one gets A(1 + T) = C(N,S) A = C(N,S) / (1 + T) And since T is normally much, much larger than 1, we can forget the 1 to give your formula A = C(N,S) / T Now lets take into account the fact that it appears T(N,S,O) is normally larger than the number of actual sets excluded by the addition of each allowable set. That makes A smaller than it should be in the above equation and thus, changes the equation to A = C(N,S) / T Which is equivalent to saying C(N,S) / T is a lower bounds for A, just as you have been saying. IS THIS EXPLANATION CORRECT? 2. THE SOURCE OF THE OVER COUNTING ASSOCIATED WITH T(S,N,O) I think I understand why there would be over counting in the formula T(N,S,O) used in your formula A = C(N,S) / T(N,S,O). It appears to results from the fact that --- when you calculate T for a given allowable assembly of length S using the formula: T(N,S,O) = SUM FROM X =O TO S OF C(S,X)*C(N-S,S-X) to estimate the number T of possible assemblies that have impermissible overlap with the given allowable assembly --- it would appear that some assemblies counted as excluded by an iteration of T with a smaller value of X would also be counted as excluded in other iterations having a larger value of X. This is because all the overlapping sub-combinations C(S,X) that would occur for a smaller value of X, would also occur as part of one or more of the sub-combinations C(S,X) that would occur for a larger value of X. Thus, T would create a number that is larger than the actual number of assemblies that would have impermissible overlap with a given allowable assembly. IS THIS CORRECT? IS THERE ANY OTHER SOURCE OF OVERCOUNTING? I though their might also be over counting because it appears --- as is stated under heading 1 --- that in the formula A = C(N,S) / T(N,S,O), T is implicitly calculated for each allowable assembly in A, and I though their might be overlap between the count T of excluded assembles made for different allowable assemblies. But it now appears to me that since none of the allowable assemblies share any overlaps X of length O to S with any other allowable set, it would appear that there would be no overlap between any of the T(N,S,O) assemblies counted as being unallowable for any first allowed assembly and those calculated as unallowable for any second allowed assembly. That is, none of the C(S,X) sub-combinations, with X O, that could be made from any first allowable assembly of length S, would be shared with any other allowable assemblies, meaning that none of the assemblies excluded by the calculation of T for one allowable assembly, could be included in the calculation of T for another allowable assembly. IS THIS CORRECT? = Finally, I have to ask if you came up with the equation A = C(N,S) / T yourself, or if you got it from some other source (and if so which source). I AM VERY THANKFUL IF YOU FOUND IT FROM ANOTHER
Re: [OpenCog] Re: [agi] constructivist issues
Well, I am confident my approach with subscripts to handle disambiguation and reference resolution would work, in conjunction with the existing link-parser/RelEx framework... If anyone wants to implement it, it seems like just some hacking with the open-source Java RelEx code... Like what I called a semantically-driven English-subset translator?. Oh, I'm pretty confidant that it will work as well . . . . after the LaBrea tar pit of implementations . . . . (exactly how little semantic-related coding do you think will be necessary? ;-) - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Cc: [EMAIL PROTECTED] Sent: Wednesday, October 22, 2008 1:06 PM Subject: Re: [OpenCog] Re: [agi] constructivist issues Well, I am confident my approach with subscripts to handle disambiguation and reference resolution would work, in conjunction with the existing link-parser/RelEx framework... If anyone wants to implement it, it seems like just some hacking with the open-source Java RelEx code... ben g On Wed, Oct 22, 2008 at 12:59 PM, Mark Waser [EMAIL PROTECTED] wrote: I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. Awesome. Like I said, it's a piece of something that I'm trying currently. If I get positive results, I'm certainly not going to hide the fact. ;-) (or, it could turn into a learning experience like my attempts with Simplified English and Basic English :-) - Original Message - From: Ben Goertzel To: agi@v2.listbox.com Cc: [EMAIL PROTECTED] Sent: Wednesday, October 22, 2008 12:27 PM Subject: [OpenCog] Re: [agi] constructivist issues This is the standard Lojban dictionary http://jbovlaste.lojban.org/ I am not so worried about word meanings, they can always be handled via reference to WordNet via usages like run_1, run_2, etc. ... or as you say by using rarer, less ambiguous words Prepositions are more worrisome, however, I suppose they can be handled in a similar way, e.g. by defining an ontology of preposition meanings like with_1, with_2, with_3, etc. In fact we had someone spend a couple months integrating existing resources into a preposition-meaning ontology like this a while back ... the so-called PrepositionWordNet ... or as it eventually came to be called the LARDict or LogicalArgumentRelationshipDictionary ... I think it would be feasible to tweak RelEx to recognize these sorts of subscripts, and in this way to recognize a highly controlled English that would be unproblematic to map semantically... We would then say e.g. I ate dinner with_2 my fork I live in_2 Maryland I have lived_6 for_3 41 years (where I suppress all _1's, so that e.g. ate means ate_1) Because, RelEx already happily parses the syntax of all simple sentences, so the only real hassle to deal with is disambiguation. We could use similar hacking for reference resolution, temporal sequencing, etc. The terrorists_v1 robbed_v2 my house. After that_v2, the jerks_v1 urinated in_3 my yard. I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. -- Ben G On Wed, Oct 22, 2008 at 12:00 PM, Mark Waser [EMAIL PROTECTED] wrote: IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic English) actually *favored* the small tremendously over-used/ambiguous words (because you got so much more bang for the buck with them). Try only using big unambiguous words and see if you still have the same opinion. If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. My first reaction is . . . . Take a Lojban dictionary and see if you can come up with an unambiguous English word or very short phrase for each Lojban word. If you can do it, my approach will work and will have the advantage that the output can be read by anyone (i.e. it's the equivalent
AW: AW: [agi] If your AGI can't learn to play chess it is no AGI
It depends what to play chess poorly mean. No one would expect that a general AGI architecture can outperform special chess programs with the same computational resources. I think you could convince a lot of people if you demonstrate that your approach which is obviously completely different from brute force chess can learn chess to a moderate level of a let's say average 10 year old human chess player. At least when you are in your open cog roadmap between phase artificial child and artificial adult then your system should necessarily be able to learn chess without any special hacking of hidden chess knowledge. BTW, Computer GO is already not so bad: http://www.engadget.com/2008/08/15/supercomputer-huygens-beats-go-professio nal-no-one-is-safe/ http://www.engadget.com/2008/08/15/supercomputer-huygens-beats-go-profession al-no-one-is-safe/ - Matthias Ben wrote: I strongly suspect that OpenCog ... once more of the NM tools are ported to it (e.g. the completion of the backward chainer port) ... could learn to play chess legally but not very well. To get it to play really well would probably require either a lot of specialized hacking with inference control, or a broader AGI approach going beyond the chess domain... or a lot more advancement of the learning mechanisms (along lines already specified in the OCP design) To me, teaching OpenCog to play chess poorly would prove almost nothing. And getting it to play chess well via tailoring the inference control mechanisms would prove little that's relevant to AGI, though it would be cool. Ok. I do not say that your approach is wrong. In fact I think it is very interesting and ambitious. But as you think that my approach is not the best one I think that your approach is not the best one. Probably, the discussion could be endless. And probably you already have invested too much effort in your approach that you really can consider to change it. I hope you are right because I would be very happy to see the first AGI soon, regardless who will build it and regardless which concept is used. I would change my approach if I thought there were a better one. But you haven't convinced me, just as I haven't convinced you ;-) Anyway, to take your approach I would not need to change my AGI design at all: OCP could be pursued in the domain of learning to play chess. I just don't think that's the best choice. BTW, if I were going to pursue a board game I'd choose Go not chess ... at least it hasn't been solved by narrow-AI very well yet ... so a really good OpenCog-based Go program would have more sex appeal ... there has not been a Deep Blue of Go My son is a good Go player so maybe I'll talk him into trying this one day ;-) ben g _ agi | https://www.listbox.com/member/archive/303/=now Archives https://www.listbox.com/member/archive/rss/303/ | https://www.listbox.com/member/?; 7 Modify Your Subscription 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=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: [agi] If your AGI can't learn to play chess it is no AGI
A couple of distinctions that I think would be really helpful for this discussion . . . . There is a profound difference between learning to play chess legally and learning to play chess well. There is an equally profound difference between discovering how to play chess well and being taught to play chess well. Personally, I think that a minimal AGI should be able to be taught to play chess reasonably well (i.e. about how well an average human would play after being taught the rules and playing a reasonable number of games with hints/pointers/tutoring provided) at about the same rate as a human when given the same assistance as that human. Given that grandmasters don't learn solely from chess-only examples without help or without analogies and strategies from other domains, I don't see why an AGI should be forced to operate under those constraints. Being taught is much faster and easier than discovering on your own. Translating an analogy or transferring a strategy from another domain is much faster than discovering something new or developing something from scratch. Why are we crippling our AGI in the name of simplicity? (And Go is obviously the same) --- 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] If your AGI can't learn to play chess it is no AGI
* * Mathematics, though, is interesting in other ways. I don't believe that much of mathematics involves the logical transformations performed in proof steps. A system that invents new fields of mathematics, new terms, new mathematical ideas -- that is truly interesting. Inference control is boring, but inventing mathematical induction, complex numbers, or ring theory -- THAT is AGI-worthy. Is this different from generic concept formulation and explanation (just in a slightly different domain)? No system can make those kinds of inventions without sophisticated inference control. Concept creation of course is required also, though. -- Ben --- 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: AW: [agi] Language learning (was Re: Defining AGI)
--- On Wed, 10/22/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: You make the implicit assumption that a natural language understanding system will pass the turing test. Can you prove this? If you accept that a language model is a probability distribution over text, then I have already proved something stronger. A language model exactly duplicates the distribution of answers that a human would give. The output is indistinguishable by any test. In fact a judge would have some uncertainty about other people's language models. A judge could be expected to attribute some errors in the model to normal human variation. Furthermore, it is just an assumption that the ability to have and to apply the rules are really necessary to pass the turing test. For these two reasons, you still haven't shown 3a and 3b. I suppose you are right. Instead of encoding mathematical rules as a grammar, with enough training data you can just code all possible instances that are likely to be encountered. For example, instead of a grammar rule to encode the commutative law of addition, 5 + 3 = a + b = b + a = 3 + 5 a model with a much larger training data set could just encode instances with no generalization: 12 + 7 = 7 + 12 92 + 0.5 = 0.5 + 92 etc. I believe this is how Google gets away with brute force n-gram statistics instead of more sophisticated grammars. It's language model is probably 10^5 times larger than a human model (10^14 bits vs 10^9 bits). Shannon observed in 1949 that random strings generated by n-gram models of English (where n is the number of either letters or words) look like natural language up to length 2n. For a typical human sized model (1 GB text), n is about 3 words. To model strings longer than 6 words we would need more sophisticated grammar rules. Google can model 5-grams (see http://googleresearch.blogspot.com/2006/08/all-our-n-gram-are-belong-to-you.html ), so it is able to generate and recognize (thus appear to understand) sentences up to about 10 words. By the way: The turing test must convince 30% of the people. Today there is a system which can already convince 25% http://www.sciencedaily.com/releases/2008/10/081013112148.htm It would be interesting to see a version of the Turing test where the human confederate, machine, and judge all have access to a computer with an internet connection. I wonder if this intelligence augmentation would make the test easier or harder to pass? -Matthias 3) you apply rules such as 5 * 7 = 35 - 35 / 7 = 5 but you have not shown that 3a) that a language understanding system necessarily(!) has this rules 3b) that a language understanding system necessarily(!) can apply such rules It must have the rules and apply them to pass the Turing test. -- Matt Mahoney, [EMAIL PROTECTED] -- Matt Mahoney, [EMAIL PROTECTED] --- 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] constructivist issues
Mark, I own and have read the book-- but my first introduction to Godel's Theorem was Douglas Hofstadter's earlier work, Godel Escher Bach. Since I had already been guided through the details of the proof (and grappled with the consequences), to be honest chapter 10 you refer to was a little boring :). But, I still do not agree with the way you are using the incompleteness theorem. It is important to distinguish between two different types of incompleteness. 1. Normal Incompleteness-- a logical theory fails to completely specify something. 2. Godelian Incompleteness-- a logical theory fails to completely specify something, even though we want it to. Logicians always mean type 2 incompleteness when they use the term. To formalize the difference between the two, the measuring stick of semantics is used. If a logic's provably-true statements don't match up to its semantically-true statements, it is incomplete. However, it seems like all you need is type 1 completeness for what you are saying. Nobody claims that there is a complete, well-defined semantics for natural language against which we could measure the provably-true (whatever THAT would mean). So, Godel's theorem is way overkill here in my opinion. --Abram On Wed, Oct 22, 2008 at 7:48 PM, Mark Waser [EMAIL PROTECTED] wrote: Most of what I was thinking of and referring to is in Chapter 10. Gödel's Quintessential Strange Loop (pages 125-145 in my version) but I would suggest that you really need to read the shorter Chapter 9. Pattern and Provability (pages 113-122) first. I actually had them conflated into a single chapter in my memory. I think that you'll enjoy them tremendously. - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 4:19 PM Subject: Re: [agi] constructivist issues Mark, Chapter number please? --Abram On Wed, Oct 22, 2008 at 1:16 PM, Mark Waser [EMAIL PROTECTED] wrote: Douglas Hofstadter's newest book I Am A Strange Loop (currently available from Amazon for $7.99 - http://www.amazon.com/Am-Strange-Loop-Douglas-Hofstadter/dp/B001FA23HM) has an excellent chapter showing Godel in syntax and semantics. I highly recommend it. The upshot is that while it is easily possible to define a complete formal system of syntax, that formal system can always be used to convey something (some semantics) that is (are) outside/beyond the system -- OR, to paraphrase -- meaning is always incomplete because it can always be added to even inside a formal system of syntax. This is why I contend that language translation ends up being AGI-complete (although bounded subsets clearly don't need to be -- the question is whether you get a usable/useful subset more easily with or without first creating a seed AGI). - Original Message - From: Abram Demski [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Wednesday, October 22, 2008 12:38 PM Subject: Re: [agi] constructivist issues Mark, The way you invoke Godel's Theorem is strange to me... perhaps you have explained your argument more fully elsewhere, but as it stands I do not see your reasoning. --Abram On Wed, Oct 22, 2008 at 12:20 PM, Mark Waser [EMAIL PROTECTED] wrote: It looks like all this disambiguation by moving to a more formal language is about sweeping the problem under the rug, removing the need for uncertain reasoning from surface levels of syntax and semantics, to remember about it 10 years later, retouch the most annoying holes with simple statistical techniques, and continue as before. That's an excellent criticism but not the intent. Godel's Incompleteness Theorem means that you will be forever building . . . . All that disambiguation does is provides a solid, commonly-agreed upon foundation to build from. English and all natural languages are *HARD*. They are not optimal for simple understanding particularly given the realms we are currently in and ambiguity makes things even worse. Languages have so many ambiguities because of the way that they (and concepts) develop. You see something new, you grab the nearest analogy and word/label and then modify it to fit. That's why you then later need the much longer words and very specific scientific terms and names. Simple language is what you need to build the more specific complex language. Having an unambiguous constructed language is simply a template or mold that you can use as scaffolding while you develop NLU. Children start out very unambiguous and concrete and so should we. (And I don't believe in statistical techniques unless you have the resources of Google or AIXI) --- 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
Re: Lojban (was Re: [agi] constructivist issues)
[Usual disclaimer: this is not the approach I'm taking, but I don't find it stupid] The idea is that by teaching an AI in a minimally-ambiguous language, one can build up its commonsense understanding such that it can then deal with the ambiguities of natural language better, using this understanding... Just because Cyc failed doesn't mean teaching a system using Lojban would necessarily fail. Lojban is a lot more interesting than Cyc-L because it can tractably be used by people to informally chat with AI's, just as can a natural language... For instance, one could chat in Lojban with an embodied AI system, and it would then get strong symbol groundings for its Lojban ;-) ben g On Wed, Oct 22, 2008 at 9:23 PM, Matt Mahoney [EMAIL PROTECTED] wrote: Why would anyone use a simplified or formalized English (with regular grammar and no ambiguities) as a path to natural language understanding? Formal language processing has nothing to do with natural language processing other than sharing a common lexicon that make them appear superficially similar. - Natural language can be learned from examples. Formal language can not. - Formal language has an exact grammar and semantics. Natural language does not. - Formal language must be parsed before it can be understood. Natural language must be understood before it can be parsed. - Formal language is designed to be processed efficiently on a fast, reliable, sequential computer that neither makes nor tolerates errors, between systems that have identical, fixed language models. Natural language evolved to be processed efficiently by a slow, unreliable, massively parallel computer with enormous memory in a noisy environment between systems that have different but adaptive language models. So how does yet another formal language processing system help us understand natural language? This route has been a dead end for 50 years, in spite of the ability to always make some initial progress before getting stuck. -- Matt Mahoney, [EMAIL PROTECTED] --- On *Wed, 10/22/08, Ben Goertzel [EMAIL PROTECTED]* wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] constructivist issues To: agi@v2.listbox.com Cc: [EMAIL PROTECTED] Date: Wednesday, October 22, 2008, 12:27 PM This is the standard Lojban dictionary http://jbovlaste.lojban.org/ I am not so worried about word meanings, they can always be handled via reference to WordNet via usages like run_1, run_2, etc. ... or as you say by using rarer, less ambiguous words Prepositions are more worrisome, however, I suppose they can be handled in a similar way, e.g. by defining an ontology of preposition meanings like with_1, with_2, with_3, etc. In fact we had someone spend a couple months integrating existing resources into a preposition-meaning ontology like this a while back ... the so-called PrepositionWordNet ... or as it eventually came to be called the LARDict or LogicalArgumentRelationshipDictionary ... I think it would be feasible to tweak RelEx to recognize these sorts of subscripts, and in this way to recognize a highly controlled English that would be unproblematic to map semantically... We would then say e.g. I ate dinner with_2 my fork I live in_2 Maryland I have lived_6 for_3 41 years (where I suppress all _1's, so that e.g. ate means ate_1) Because, RelEx already happily parses the syntax of all simple sentences, so the only real hassle to deal with is disambiguation. We could use similar hacking for reference resolution, temporal sequencing, etc. The terrorists_v1 robbed_v2 my house. After that_v2, the jerks_v1 urinated in_3 my yard. I think this would be a relatively pain-free way to communicate with an AI that lacks the common sense to carry out disambiguation and reference resolution reliably. Also, the log of communication would provide a nice training DB for it to use in studying disambiguation. -- Ben G On Wed, Oct 22, 2008 at 12:00 PM, Mark Waser [EMAIL PROTECTED] wrote: IMHO that is an almost hopeless approach, ambiguity is too integral to English or any natural language ... e.g preposition ambiguity Actually, I've been making pretty good progress. You just always use big words and never use small words and/or you use a specific phrase as a word. Ambiguous prepositions just disambiguate to one of three/four/five/more possible unambiguous words/phrases. The problem is that most previous subsets (Simplified English, Basic English) actually *favored* the small tremendously over-used/ambiguous words (because you got so much more bang for the buck with them). Try only using big unambiguous words and see if you still have the same opinion. If you want to take this sort of approach, you'd better start with Lojban instead Learning Lojban is a pain but far less pain than you'll have trying to make a disambiguated subset of English. My first reaction is . . . . Take a Lojban dictionary and see
Re: Lojban (was Re: [agi] constructivist issues)
On Thu, Oct 23, 2008 at 11:23 AM, Matt Mahoney [EMAIL PROTECTED] wrote: So how does yet another formal language processing system help us understand natural language? This route has been a dead end for 50 years, in spite of the ability to always make some initial progress before getting stuck. Although I mostly agree with you, I do often think that humans understand formal languages very differently to, say, compilers (if they can be said to understand them at all) and I think it is interesting to study how one might build an AGI system that understands formal languages the way humans do. I have no idea whether it is easier to do this with formal languages than it is to do this with natural languages. Trent --- 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
