Re: Re[2]: [agi] Self-building AGI
On 02/12/2007, Ed Porter [EMAIL PROTECTED] wrote: I currently think there are some human human-level intelligences who know how to build most of an AGI, at least enough to get up and running systems that would solve many aspects of the AGI problem and help us better understand what, if any other aspects of the problem needed to be solved. I think the Novamente team is one example. I think you may be right, although there have been many people in the past who believed they knew how to create an intelligent system, but who made little progress on the problem. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71392535-36c9a6
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
On Sunday 02 December 2007, John G. Rose wrote: Building up parse trees and word sense models, let's say that would be a first step. And then say after a while this was accomplished and running on some peers. What would the next theoretical step be? I am not sure what the next step would be. The first step might be enough for the moment. When you have the network functioning at all, expose an API so that other programmers can come in and try to utilize sentence analysis (and other functions) as if the network is just another lobe of the brain or another component for ai. This would allow others who are possibly more creative than us to take advantage of what looks to be interesting work. - Bryan - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71422338-8cb1da
[agi] AGI first mention on NPR!
Yesterday I heard the phrase Artificial General Intelligence on the radio for the first time ever: http://www.npr.org/templates/story/story.php?storyId=16816185 Weekend Edition Sunday, December 2, 2007 · The idea of what Artificial Intelligence should be has evolved over the past 50 years — from solving puzzles and playing chess to emulating the abilities of a child: walking, recognizing objects. A recent conference brought together those who invent the future. A recent Singularity Summit brought together those who imagine — and invent — the future. Unfortunately, most of the report was filled with sound bites that were, to my mind, ridiculously naive extrapolations and speculations, like: Paul Saffo: The optimistic scenario is they will treat us like pets most of which were calculated to horrify the audience. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71450599-b9df52
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: Once you build up good models for parsing and word sense, then you read large amounts of text and start building up model of the realities described and generalizations from them. Assuming this is a continuation of the discussion of an AGI-at-home P2P system, you are going to be very limited by the lack of bandwidth, particularly for attacking the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space, which is going to be difficult with limited bandwidth. But a large amount of text with appropriate parsing and word sense labeling would still provide a valuable aid for web and text search and for many forms of automatic learning. And the level of understanding that such a P2P system could derive from reading huge amounts of text could be a valuable initial source of one component of world knowledge for use by AGI. I know you always find it teious when I express scepticism, so I will preface my remarks with: take this advice or ignore it, your choice. This description of how to get AGI done reminds me of my childhood project to build a Mars-bound spacecraft modeled after James Blish's Book Welcome to Mars. I Knew that I could build it in time for the next conjunction of Mars, but I hadn't quite gotten the anti-gravity drive sorted out, so instead I collected all the other materials described in the book, so everything would be ready when the AG drive started working... The reason it reminds me of this episode is that you are calmly talking here about the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space . this is your equivalent of the anti-gravity drive. This is the part that needs extremely detailed knowledge of AI and psychology, just to be understand the nature of the problem (never mind to solve it). If you had any idea bout how to solve this part of the problem, everything else would drop into your lap. You wouldn't need a P2P AGI-at-home system, because with this solution in hand you would have people beating down your door to give you a supercomputer. Menawhile, unfortunately, solving all those other issues like making parsers and trying to do word-sense disambiguation would not help one whit to get the real theoretical task done. I am not being negative, I am just relaying the standard understanding of priorities in the AGI field as a whole. Send complaints addressed to AGI Community, not to me, please. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71451441-4352c5
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Once you build up good models for parsing and word sense, then you read large amounts of text and start building up model of the realities described and generalizations from them. Assuming this is a continuation of the discussion of an AGI-at-home P2P system, you are going to be very limited by the lack of bandwidth, particularly for attacking the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space, which is going to be difficult with limited bandwidth. But a large amount of text with appropriate parsing and word sense labeling would still provide a valuable aid for web and text search and for many forms of automatic learning. And the level of understanding that such a P2P system could derive from reading huge amounts of text could be a valuable initial source of one component of world knowledge for use by AGI. Ed Porter -Original Message- From: Bryan Bishop [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 7:33 AM To: agi@v2.listbox.com Subject: Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research] On Sunday 02 December 2007, John G. Rose wrote: Building up parse trees and word sense models, let's say that would be a first step. And then say after a while this was accomplished and running on some peers. What would the next theoretical step be? I am not sure what the next step would be. The first step might be enough for the moment. When you have the network functioning at all, expose an API so that other programmers can come in and try to utilize sentence analysis (and other functions) as if the network is just another lobe of the brain or another component for ai. This would allow others who are possibly more creative than us to take advantage of what looks to be interesting work. - Bryan - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71438525-d92982
Re: [agi] AGI first mention on NPR!
Bob Mottram wrote: Perhaps a good word of warning is that it will be really easy to satirise/lampoon/misrepresent AGI and its proponents until such time as one is actually created. The problem is that these two activities - denigrating AGI, and actually building one - are not two independent things. The first could have a serious effect on the second. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71453422-d780e4
Re: [agi] AGI first mention on NPR!
Perhaps a good word of warning is that it will be really easy to satirise/lampoon/misrepresent AGI and its proponents until such time as one is actually created. On 03/12/2007, Richard Loosemore [EMAIL PROTECTED] wrote: Yesterday I heard the phrase Artificial General Intelligence on the radio for the first time ever: http://www.npr.org/templates/story/story.php?storyId=16816185 Weekend Edition Sunday, December 2, 2007 · The idea of what Artificial Intelligence should be has evolved over the past 50 years — from solving puzzles and playing chess to emulating the abilities of a child: walking, recognizing objects. A recent conference brought together those who invent the future. A recent Singularity Summit brought together those who imagine — and invent — the future. Unfortunately, most of the report was filled with sound bites that were, to my mind, ridiculously naive extrapolations and speculations, like: Paul Saffo: The optimistic scenario is they will treat us like pets most of which were calculated to horrify the audience. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71452571-92cff0
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
From: Richard Loosemore [mailto:[EMAIL PROTECTED] The reason it reminds me of this episode is that you are calmly talking here about the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space . this is your equivalent of the anti-gravity drive. This is the part that needs extremely detailed knowledge of AI and psychology, just to be understand the nature of the problem (never mind to solve it). If you had any idea bout how to solve this part of the problem, everything else would drop into your lap. You wouldn't need a P2P AGI-at-home system, because with this solution in hand you would have people beating down your door to give you a supercomputer. This is naïve. It almost never works this way, where if someone has a solution to a well known unsolved engineering problem that resources just come knocking at the door. Menawhile, unfortunately, solving all those other issues like making parsers and trying to do word-sense disambiguation would not help one whit to get the real theoretical task done. This is impractical. ... I am not being negative, I am just relaying the standard understanding of priorities in the AGI field as a whole. Send complaints addressed to AGI Community, not to me, please. You are being negative! And since when have the priorities of understandings in the AGI field been standardized? Perhaps that is part the limiting factor and self-defeating narrow-mindedness. John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71501965-68a77a
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Menawhile, unfortunately, solving all those other issues like making parsers and trying to do word-sense disambiguation would not help one whit to get the real theoretical task done. I agree. AI has a long history of doing the easy part of the problem first: solving the mathematics or logic of a word problem, and deferring the hard part, which is extracting the right formal statement from the natural language input. This is the opposite order of how children learn. The proper order is: lexical rules first, then semantics, then grammar, and then the problem solving. The whole point of using massive parallel computation is to do the hard part of the problem. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71493437-c427ac
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Matt:: The whole point of using massive parallel computation is to do the hard part of the problem. I get it : you and most other AI-ers are equating hard with very, very complex, right? But you don't seriously think that the human mind successfully deals with language by massive parallel computation, do you? Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, yes, (in order to understand, for example, any of the sentences you or I are writing here), but only a tiny fraction of the operations that computers currently perform? The whole idea of massive parallel computation here, surely has to be wrong. And yet none of you seem able to face this to my mind obvious truth. I only saw this term recently - perhaps it's v. familiar to you (?) - that the human brain works by look-up rather than search. Hard problems can have relatively simple but ingenious solutions. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71504832-b01a2d
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
From: Bryan Bishop [mailto:[EMAIL PROTECTED] I am not sure what the next step would be. The first step might be enough for the moment. When you have the network functioning at all, expose an API so that other programmers can come in and try to utilize sentence analysis (and other functions) as if the network is just another lobe of the brain or another component for ai. This would allow others who are possibly more creative than us to take advantage of what looks to be interesting work. This is true and a way to get utility out of it. And getting the first step accomplished is quite a bit of work as is maintaining it. Having just a few basic baby steps actually materialize in front of you eliminates some of the complexity so that the larger problem may appear just a bit less daunting. Also communal developer feedback is a constructive motivator. John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71510117-536e83
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
From: Ed Porter [mailto:[EMAIL PROTECTED] Once you build up good models for parsing and word sense, then you read large amounts of text and start building up model of the realities described and generalizations from them. Assuming this is a continuation of the discussion of an AGI-at-home P2P system, you are going to be very limited by the lack of bandwidth, particularly for attacking the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space, which is going to be difficult with limited bandwidth. But a large amount of text with appropriate parsing and word sense labeling would still provide a valuable aid for web and text search and for many forms of automatic learning. And the level of understanding that such a P2P system could derive from reading huge amounts of text could be a valuable initial source of one component of world knowledge for use by AGI. I kind of see the small bandwidth between (most) individual nodes as not a limiting factor as sets of nodes act as temporary single group entities. IOW the BW between one set of 50 nodes and another set of 50 nodes is quite large actually and individual nodes' data access would depend on - indexes of indexes to minimize their individual BW requirements. Does this not apply to your model? John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71511001-15807d
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
John G. Rose wrote: From: Richard Loosemore [mailto:[EMAIL PROTECTED] [snip] I am not being negative, I am just relaying the standard understanding of priorities in the AGI field as a whole. Send complaints addressed to AGI Community, not to me, please. You are being negative! And since when have the priorities of understandings in the AGI field been standardized? Perhaps that is part the limiting factor and self-defeating narrow-mindedness. It is easy for a research field to agree that certain problems are really serious and unsolved. A hundred years ago, the results of the Michelson-Morley experiments were a big unsolved problem, and pretty serious for the foundations of physics. I don't think it would have been self-defeating narrow-mindedness for someone to have pointed to that problem and said this is a serious problem. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71517612-4f04ee
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Mike Tintner wrote: Matt:: The whole point of using massive parallel computation is to do the hard part of the problem. I get it : you and most other AI-ers are equating hard with very, very complex, right? But you don't seriously think that the human mind successfully deals with language by massive parallel computation, do you? Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, yes, (in order to understand, for example, any of the sentences you or I are writing here), but only a tiny fraction of the operations that computers currently perform? The whole idea of massive parallel computation here, surely has to be wrong. And yet none of you seem able to face this to my mind obvious truth. I only saw this term recently - perhaps it's v. familiar to you (?) - that the human brain works by look-up rather than search. Hard problems can have relatively simple but ingenious solutions. You need to check the psychology data: it emphatically disagrees with your position here. One thing that can be easily measured is the activation of lexical items related in various ways to a presented word (i.e. show the subject the word Doctor and test to see if the word Nurse gets activated). It turns out that within an extremely short time of the forst word being seen, a very large numbmer of other words have their activations raised significantly. Now, whichever way you interpret these (so called priming) results, one thing is not in doubt: there is massively parallel activation of lexical units going on during language processing. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71515718-ac1ab7
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: Richard, It is false to imply that knowledge of how to draw implications from a series of statements by some sort of search mechanism is equally unknown as that of how to make an anti-gravity drive -- if by anti-gravity drive you mean some totally unknown form of physics, rather than just anything, such as human legs, that can push against gravity. It is unfair because there is a fair amount of knowledge about how to draw implications from sequences of statements. For example view Shastri's www.icsi.berkeley.edu/~shastri/psfiles/cogsci00.ps. Also Ben Goertzel has demonstrated a program that draws implications from statements contained in different medical texts. Ed Porter P.S., I have enclosed an inexact, but, at least to me, useful drawing I made of the type of search involved in understanding the multiple implications contained in the series of statements contained in Shastri's John fell in the Hallway. Tom had cleaned it. He was hurt example. Of course, what is most missing from this drawing are all the other, dead end, implications which do not provide a likely implication. Only one of such dead end is shown (the implication between fall and trip). As a result you don't sense how many dead ends have to be searched to find the implications which best explain the statements. EWP Well, bear in mind that I was not meaning the analogy to be *that* exact, or I would have given up on AGI long ago - I'm sure you know that I don't believe that getting an understanding system working is as impossible as getting an AG drive built. The purpose of my comment was to point to a huge gap in understanding, and the mistaken strategy of dealing with all the peripheral issues before having a clear idea how to solve the central problem. I cannot even begin to do justice, here, to the issues involved in solving the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space You talk as if an extension of some current strategy will solve this ... but it is not at all clear that any current strategy for solving this problem actually does scale up to a full solution to the problem. I don't care how many toy examples you come up with, you have to show a strategy for dealing with some of the core issues, AND reasons to believe that those strategies really will work (other than I find them quite promising). Not only that, but there at least some people (to wit, myself) who believe there are positive reasons to believe that the current strategies *will* not scale up. Richard Loosemore -Original Message- From: Richard Loosemore [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 10:07 AM To: agi@v2.listbox.com Subject: Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research] Ed Porter wrote: Once you build up good models for parsing and word sense, then you read large amounts of text and start building up model of the realities described and generalizations from them. Assuming this is a continuation of the discussion of an AGI-at-home P2P system, you are going to be very limited by the lack of bandwidth, particularly for attacking the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space, which is going to be difficult with limited bandwidth. But a large amount of text with appropriate parsing and word sense labeling would still provide a valuable aid for web and text search and for many forms of automatic learning. And the level of understanding that such a P2P system could derive from reading huge amounts of text could be a valuable initial source of one component of world knowledge for use by AGI. I know you always find it teious when I express scepticism, so I will preface my remarks with: take this advice or ignore it, your choice. This description of how to get AGI done reminds me of my childhood project to build a Mars-bound spacecraft modeled after James Blish's Book Welcome to Mars. I Knew that I could build it in time for the next conjunction of Mars, but I hadn't quite gotten the anti-gravity drive sorted out, so instead I collected all the other materials described in the book, so everything would be ready when the AG drive started working... The reason it reminds me of this episode is that you are calmly talking here about the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space . this is your equivalent of the anti-gravity drive. This is the part that needs extremely detailed knowledge of AI and psychology, just to be
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
From: Richard Loosemore [mailto:[EMAIL PROTECTED] It is easy for a research field to agree that certain problems are really serious and unsolved. A hundred years ago, the results of the Michelson-Morley experiments were a big unsolved problem, and pretty serious for the foundations of physics. I don't think it would have been self-defeating narrow-mindedness for someone to have pointed to that problem and said this is a serious problem. Well the definition of problems and the approaches to solving the problems can be narrow-minded or looked at with a narrow-human-psychological AI perspective. Most of these problems boil down to engineering problems and the theory already exists in some other form; it is a matter of putting things together IMO. But myself not being in the cog sci world for that long, only thinking of AGI in terms of computers, math and AI, I am unaware of the details of some of the particular AGI unsolved mysteries that are talked about. Not to say I haven't thought about them from my own narrow-human-psychological AI perspective :) John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71519373-6b5212
[agi] What are the real unsolved issues in AGI [WAS Re: Hacker intelligence
John G. Rose wrote: From: Richard Loosemore [mailto:[EMAIL PROTECTED] It is easy for a research field to agree that certain problems are really serious and unsolved. A hundred years ago, the results of the Michelson-Morley experiments were a big unsolved problem, and pretty serious for the foundations of physics. I don't think it would have been self-defeating narrow-mindedness for someone to have pointed to that problem and said this is a serious problem. Well the definition of problems and the approaches to solving the problems can be narrow-minded or looked at with a narrow-human-psychological AI perspective. Most of these problems boil down to engineering problems and the theory already exists in some other form; it is a matter of putting things together IMO. I think this is a very important issue in AGI, which is why I felt compelled to say something. As you know, I keep trying to get meaningful debate to happen on the subject of *methodology* in AGI. That is what my claims about the complex systems problem are all about: the very serious possibility that the existing AGI/AI methodology is so seriously broken that virtually everything going on right now will be written up by future historians as a complete waste of effort. In that context - where there is something of an agreement about what the big unsolved problems are, and where I have raised questions about the very foundations of today's AGI methodology - it is truly astonishing to hear people talking about issues being more or less solved, bar the shouting. Richard Loosemore P.S. BTW, it isn't really anything to do with taking a cognitive science perspective. Don't forget that I come from a hybrid background: I am not a cognitive scientist encroaching on hard-science AI and computing, I have done both sides in equal measure. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71525665-a80bc7
[agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
My suggestion, criticized below (criticism can be valuable), was for just one of many possible uses of an open-source P2P AGI-at-home type system. I am totally willing to hear other proposals. Considering how little time I spent coming up with the one being criticized, I have a relatively low ego investment in it and I assume there will be better suggestions from others. I think the hard part of AGI will be difficult to address on a P2P system with low interconnect bandwidth. I do because I believe the hard part of AGI will be learning appropriate dynamic controls for massively parallel systems computing over massive amounts of data, and the creation of automatically self organizing knowledge bases derived from computing from such massive amounts of knowledge in a highly non-localized way. For progress on these fronts at any reasonable speed you need massive bandwidth, which a current P2P system would lack, according to the previous communications on this thread. So a current P2P system on the web is not going to be a good test bed for anything approaching human-level AGI. But interesting things could be learned with P2P AGI-at-Home networks. In the NL example I proposed, the word senses and parsing were all to be learned with generalized AGI learning algorithms (although bootstrapped with some narrow AI tools) I think they could be a good test bed for AGI learning of self organizing gen-comp hierarchies because the training data is plentiful and easy to get, many of the gen-comp hierarchy of patterns that would be formed would be ones that we humans could understand, and the capabilities of the system would be ones we could compare to human level performance in a somewhat intuitive manner. With regard to the statement that The proper order is: lexical rules first, then semantics, then grammar, and then the problem solving. The whole point of using massive parallel computation is to do the hard part of the problem I have the following two comments: (1) As I have said before, the truly hard part of AGI is almost certainly going to be beyond a P2P network of PCs. And (2) with regard to the order of NL learning, I think a child actually learns semantics first (words associated with sets of experience), since most young children I have met start communicating first in single word statements. The word sense experts I proposed in the P2P system would be focusing on this level of knowledge. Unfortunately, they would be largely limited to experience in the form of a textual context, resulting in a quite limited form of experiential grounding. The type of generalized AGI learning algorithm I proposed would address lexical rules and grammar as part of both its study of grammar and word senses. I have only separated out different forms of expertise because each PC can only contain a relatively small amount of information, so there has to be some attempt to separate the P2P's AGI representation into regions with the highest locality of reference. In and ideal world this should be done automatically, but to do this well automatically would tend to require high bandwidth, which the P2P system wouldn't have. So at least initially it probably makes sense to have humans decide what the various fields of expertise are (although such decisions could be based on AGI derived data, such as that obtained from data access patterns on singe PC AGI prototypes, or even on an initial networked system). Also, I think we should take advantage of some of the narrow AI tools we have, such as parsers, WordNet, dictionaries, and word-sense quessers, to bootstrap the system so that we could get more deeply into the more interesting aspect of AGI such as semantic understanding faster. These narrow AI tools could be used in conjunction with AGI learning. For example, the output of a narrow AI parser or word sense labeler could be used to provide initial data used to train up AGI models, which could then replace or run in conjunction with the narrow AGI tools in a set of EM cycles, with the AGI models hopefully providing more consistent labeling at time progresses, and increasingly getting more weight relative to the narrow AI tools. Perhaps one aspect of the AGI-at-home project would be to develop a good generalized architecture for wedding various classes of narrow AI and AGI in such a learning environment. Narrow AI's are often very efficient, but they have very limitations which AGI can often overcome. Perhaps learning how to optimally wed the two could create systems that had the best features of both AGI and narrow AI, greatly increasing the efficiency of AGI. But there are all sorts of other interesting things that could be done with an AGI-at-home P2P system. I am claiming no special expertise as to what is the best use of it. For example, I think it would be interesting to see what sort of AGI's could be built on current PCs with up to 4G or RAM. It would be interesting to see just what
Re: [agi] What are the real unsolved issues in AGI [WAS Re: Hacker intelligence
On 03/12/2007, Richard Loosemore [EMAIL PROTECTED] wrote: it is truly astonishing to hear people talking about issues being more or less solved, bar the shouting. You'll usually find that such people never trouble themselves with implementational details. Intuitive notions about how easy some task ought to be are a notoriously poor guide in this area. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71564707-4a8b1c
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
MIKE TINTNER Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, ED PORTER How do you find the right set of relatively few computations and/or models that are appropriate in a complex context without massive computation? -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 12:12 PM To: agi@v2.listbox.com Subject: Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research] Matt:: The whole point of using massive parallel computation is to do the hard part of the problem. I get it : you and most other AI-ers are equating hard with very, very complex, right? But you don't seriously think that the human mind successfully deals with language by massive parallel computation, do you? Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, yes, (in order to understand, for example, any of the sentences you or I are writing here), but only a tiny fraction of the operations that computers currently perform? The whole idea of massive parallel computation here, surely has to be wrong. And yet none of you seem able to face this to my mind obvious truth. I only saw this term recently - perhaps it's v. familiar to you (?) - that the human brain works by look-up rather than search. Hard problems can have relatively simple but ingenious solutions. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71590357-a986d6
RE: [agi] What are the real unsolved issues in AGI [WAS Re: Hacker intelligence
From: Richard Loosemore [mailto:[EMAIL PROTECTED] I think this is a very important issue in AGI, which is why I felt compelled to say something. As you know, I keep trying to get meaningful debate to happen on the subject of *methodology* in AGI. That is what my claims about the complex systems problem are all about: the very serious possibility that the existing AGI/AI methodology is so seriously broken that virtually everything going on right now will be written up by future historians as a complete waste of effort. I don't think that will happen, sometimes a lot of energy expenditure needs to be made to just move ahead an inch. Also there is some spinning of wheels going on as other technologies mature which is happening quite well BTW. And there has been an awful lot of directly applicable and related theoretical work accomplished and proliferated over the last few decades. In that context - where there is something of an agreement about what the big unsolved problems are, and where I have raised questions about the very foundations of today's AGI methodology - it is truly astonishing to hear people talking about issues being more or less solved, bar the shouting. Excuse my ignorance - top 3 unsolved problems are? - NLP, and what else? And then from what I have gathered on this email list you favor a complex systems emergent approach? But you somehow don't agree with mathematical models. That's an immediate turn-off for implementationalists so it's hard to gain acceptance. Could you give a one liner (or more) description of your theory again if you don't mind, or an URL - my interest is somewhat captivated. John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71597210-e47d1c
RE: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
For some lucky cable folks the BW is getting ready to increase soon: http://arstechnica.com/news.ars/post/20071130-docsis-3-0-possible-100mbps-sp eeds-coming-to-some-comcast-users-in-2008.html I'm yet to fully understand the limitations of a P2P based AGI design or the augmentational ability of a public P2P network on a private P2P network constructed for AGI. I would count out P2P AGI so quickly. John _ From: Ed Porter [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 12:20 PM To: agi@v2.listbox.com Subject: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...] My suggestion, criticized below (criticism can be valuable), was for just one of many possible uses of an open-source P2P AGI-at-home type system. I am totally willing to hear other proposals. Considering how little time I spent coming up with the one being criticized, I have a relatively low ego investment in it and I assume there will be better suggestions from others. I think the hard part of AGI will be difficult to address on a P2P system with low interconnect bandwidth. I do because I believe the hard part of AGI will be learning appropriate dynamic controls for massively parallel systems computing over massive amounts of data, and the creation of automatically self organizing knowledge bases derived from computing from such massive amounts of knowledge in a highly non-localized way. For progress on these fronts at any reasonable speed you need massive bandwidth, which a current P2P system would lack, according to the previous communications on this thread. So a current P2P system on the web is not going to be a good test bed for anything approaching human-level AGI. But interesting things could be learned with P2P AGI-at-Home networks. In the NL example I proposed, the word senses and parsing were all to be learned with generalized AGI learning algorithms (although bootstrapped with some narrow AI tools) I think they could be a good test bed for AGI learning of self organizing gen-comp hierarchies because the training data is plentiful and easy to get, many of the gen-comp hierarchy of patterns that would be formed would be ones that we humans could understand, and the capabilities of the system would be ones we could compare to human level performance in a somewhat intuitive manner. With regard to the statement that The proper order is: lexical rules first, then semantics, then grammar, and then the problem solving. The whole point of using massive parallel computation is to do the hard part of the problem I have the following two comments: (1) As I have said before, the truly hard part of AGI is almost certainly going to be beyond a P2P network of PCs. And (2) with regard to the order of NL learning, I think a child actually learns semantics first (words associated with sets of experience), since most young children I have met start communicating first in single word statements. The word sense experts I proposed in the P2P system would be focusing on this level of knowledge. Unfortunately, they would be largely limited to experience in the form of a textual context, resulting in a quite limited form of experiential grounding. The type of generalized AGI learning algorithm I proposed would address lexical rules and grammar as part of both its study of grammar and word senses. I have only separated out different forms of expertise because each PC can only contain a relatively small amount of information, so there has to be some attempt to separate the P2P's AGI representation into regions with the highest locality of reference. In and ideal world this should be done automatically, but to do this well automatically would tend to require high bandwidth, which the P2P system wouldn't have. So at least initially it probably makes sense to have humans decide what the various fields of expertise are (although such decisions could be based on AGI derived data, such as that obtained from data access patterns on singe PC AGI prototypes, or even on an initial networked system). Also, I think we should take advantage of some of the narrow AI tools we have, such as parsers, WordNet, dictionaries, and word-sense quessers, to bootstrap the system so that we could get more deeply into the more interesting aspect of AGI such as semantic understanding faster. These narrow AI tools could be used in conjunction with AGI learning. For example, the output of a narrow AI parser or word sense labeler could be used to provide initial data used to train up AGI models, which could then replace or run in conjunction with the narrow AGI tools in a set of EM cycles, with the AGI models hopefully providing more
Re: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
On Dec 3, 2007, at 12:52 PM, John G. Rose wrote: For some lucky cable folks the BW is getting ready to increase soon: http://arstechnica.com/news.ars/post/20071130-docsis-3-0-possible-100mbps-sp eeds-coming-to-some-comcast-users-in-2008.html I'm yet to fully understand the limitations of a P2P based AGI design or the augmentational ability of a public P2P network on a private P2P network constructed for AGI. I would count out P2P AGI so quickly. Distributed algorithms tend to be far more sensitivity to latency than bandwidth, except to the extent that low bandwidth induces latency. As a practical matter, the latency floor of P2P is so high that most algorithms would run far faster on a small number of local machines than a large number of geographically distributed machines. There is a reason people interested in high-performance computing tend to spend more on their interconnect than their compute nodes. J. Andrew Rogers - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71602529-7ec12e
RE: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
From: J. Andrew Rogers [mailto:[EMAIL PROTECTED] Distributed algorithms tend to be far more sensitivity to latency than bandwidth, except to the extent that low bandwidth induces latency. As a practical matter, the latency floor of P2P is so high that most algorithms would run far faster on a small number of local machines than a large number of geographically distributed machines. There is a reason people interested in high-performance computing tend to spend more on their interconnect than their compute nodes. The P2P public network is not homogenous. Lower quality nodes far outnumber high quality nodes but high quality nodes do exist. High quality meaning both low latency and high bandwidth (example 3ms ping at 44 mbits). For human equivalent AGI a private P2P network MIGHT be required, inexpensive would be 3ms ping on a gig E clustered segment. More pricier could require an external switched fabric of say around a 5+ gigbit interconnect cluster. Lazy processing on low quality P2P - exactly how invaluable is that? Distributed P2P computing for AGI needs to be self-organizing, detect and adapt to resource conditions. It's not a perfect world. John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71615162-02022c
Re: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
--- Ed Porter [EMAIL PROTECTED] wrote: And (2) with regard to the order of NL learning, I think a child actually learns semantics first Actually Jusczyk showed that babies learn the rules for segmenting continuous speech at 7-10 months. I did some experiments in 1999 following the work of Hutchens and Alder showing that it is possible to learn the rules for segmenting text without spaces using only the simple character n-gram statistics of the input. The word boundaries occur where the mutual information across the boundary is lowest. http://cs.fit.edu/~mmahoney/dissertation/lex1.html Children begin learning the meanings of words around 12 months, and start forming simple sentences around age 2-3. For example, I think it would be interesting to see what sort of AGI's could be built on current PCs with up to 4G or RAM. I did something like that with language models, up to 2 GB. So far, my research suggests you need a LOT more memory. http://cs.fit.edu/~mmahoney/compression/text.html With regard to distributed AI, I believe the protocol should be natural language at the top level (perhaps on top of HTTP), because I think it is essential that live humans can participate. The idea is that each node in the P2P network might be relatively stupid, but would be an expert on some narrow topic, and know how to find other experts on related topics. A node would scan queries for keywords and ignore the messages it doesn't understand (which would be most of them). Overall the network would appear intelligent because *somebody* would know. When a user asks a question or posts information, the message would be broadcast to many nodes, which could choose to ignore them or relay them to other nodes that it believes would find the message more relevant. Eventually the message gets to a number of experts, who then reply to the message. The source and destination nodes would then update their links to each other, replacing the least recently used links. The system would be essentially a file sharing or message posting service with a distributed search engine. It would make no distinctions between queries and updates, because asking a question about a topic indicates knowledge of related topics. Every message you post becomes a permanent part of this gigantic distributed database, tagged with your name (or anonymous ID) and a time stamp. I wrote my thesis on the question of whether such a system would scale to a large, unreliable network. (Short answer: yes). http://cs.fit.edu/~mmahoney/thesis.html Implementation detail: how to make a P2P client useful enough that people will want to install it? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71629115-649a10
Re: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
On Dec 3, 2007 5:07 PM, Matt Mahoney [EMAIL PROTECTED] wrote: When a user asks a question or posts information, the message would be broadcast to many nodes, which could choose to ignore them or relay them to other nodes that it believes would find the message more relevant. Eventually the message gets to a number of experts, who then reply to the message. The source and destination nodes would then update their links to each other, replacing the least recently used links. I wrote my thesis on the question of whether such a system would scale to a large, unreliable network. (Short answer: yes). http://cs.fit.edu/~mmahoney/thesis.html Implementation detail: how to make a P2P client useful enough that people will want to install it? That sounds almost word-for-word like something I was visualizing (though not producing as a thesis) I believe the next step of such a system is to become an abstraction between the user and the network they're using. So if you can hook into your P2P network via a firefox extension, (consider StumbleUpon or Greasemonkey) so it (the agent) can passively monitor your web interaction - then it could be learn to screen emails (for example) or pre-chew either your first 10 google hits or summarize the next 100 for relevance. I have been told that by the time you have an agent doing this well, you'd already have AGI - but i can't believe this kind of data mining is beyond narrow AI (or requires fully general adaptive intelligence) Maybe when I get around to the Science part of my BS degree (after the Arts filler) I will explore to a greater depth for a thesis. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71648663-f0a7ee
Re: [agi] What are the real unsolved issues in AGI [WAS Re: Hacker intelligence
John G. Rose wrote: From: Richard Loosemore [mailto:[EMAIL PROTECTED] I think this is a very important issue in AGI, which is why I felt compelled to say something. As you know, I keep trying to get meaningful debate to happen on the subject of *methodology* in AGI. That is what my claims about the complex systems problem are all about: the very serious possibility that the existing AGI/AI methodology is so seriously broken that virtually everything going on right now will be written up by future historians as a complete waste of effort. I don't think that will happen, sometimes a lot of energy expenditure needs to be made to just move ahead an inch. Also there is some spinning of wheels going on as other technologies mature which is happening quite well BTW. And there has been an awful lot of directly applicable and related theoretical work accomplished and proliferated over the last few decades. In that context - where there is something of an agreement about what the big unsolved problems are, and where I have raised questions about the very foundations of today's AGI methodology - it is truly astonishing to hear people talking about issues being more or less solved, bar the shouting. Excuse my ignorance - top 3 unsolved problems are? - NLP, and what else? And then from what I have gathered on this email list you favor a complex systems emergent approach? But you somehow don't agree with mathematical models. That's an immediate turn-off for implementationalists so it's hard to gain acceptance. Could you give a one liner (or more) description of your theory again if you don't mind, or an URL - my interest is somewhat captivated. Top three? I don't know if anyone ranks them. Try: 1) Grounding Problem (the *real* one, not the cheap substitute that everyone usually thinks of as the symbol grounding problem). 2) The problem of desiging an inference control engine whose behavior is predictable/governable etc. 3) A way to represent things - and in particular, uncertainty - without getting buried up to the eyeballs in (e.g.) temporal logics that nobody believes in. Take this with a pinch of salt: I am sure there are plenty of others. But if you came up with a *principled* solution to these issues, I'd be impressed. One linear description of my theory? I'll think about it. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71653318-ec0059
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
On Dec 3, 2007 12:12 PM, Mike Tintner [EMAIL PROTECTED] wrote: I get it : you and most other AI-ers are equating hard with very, very complex, right? But you don't seriously think that the human mind successfully deals with language by massive parallel computation, do you? Very very complex tends to exceed one's ability to properly model and especially predict. Even if the human mind invokes some special kind of magical cleverness, do you think you (judging from your writing) have some unique ability to isolate that function (noun) without simultaneously using that function (verb) ? I often imagine that I understand the working of my own mind almost perfectly. Those that claim to have grasped the quintessential bit typically end up so far over the edge that they are unable to express it in meaningful or useful terms. Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, yes, (in order to understand, for example, any of the sentences you or I are writing here), but only a tiny fraction of the operations that computers currently perform? I believe you are making that statement because you wish it to be true. I see no basis for anything to be obvious - especially the formalism required to define what the term means. This is due primarily to the complexity associated with recursive self-reflection. The whole idea of massive parallel computation here, surely has to be wrong. And yet none of you seem able to face this to my mind obvious truth. We each continue to persist in our delusions. Yours may be no different in the end. :) I only saw this term recently - perhaps it's v. familiar to you (?) - that the human brain works by look-up rather than search. Hard problems can have relatively simple but ingenious solutions. How is the look-up table built? Usually by experience. When we have enough similar experiences to look up a solution to general adaptive intelligence, we will have likely been close enough to it for so long that (probably) nobody will be surprised. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71652723-808348
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
RL: One thing that can be easily measured is the activation of lexical items related in various ways to a presented word (i.e. show the subject the word Doctor and test to see if the word Nurse gets activated). It turns out that within an extremely short time of the forst word being seen, a very large numbmer of other words have their activations raised significantly. Now, whichever way you interpret these (so called priming) results, one thing is not in doubt: there is massively parallel activation of lexical units going on during language processing. Thanks for reply. How many associations are activated? How do we know neuroscientifically they are associations to the words being processed and not something else entirely? Out of interest, can you give me a ball park estimate of how many associations you personally think are activated, say, in in a few seconds, in processing sentences like: The doctor made a move on the nurse. Relationships between staff in health organizations are fraught with complexities No, I'm not trying to be ridiculously demanding or asking you to be ridiculously exact. As you probably know by now, I see the processing of sentences as involving several levels, especially for the second sentence, but I don't see the number of associations as that many. Let's be generous and guess hundreds for the items in the above sentences. But a computer program, as I understand, will be typically searching through anywhere between thousands, millions and way upwards. On the one hand, we can perhaps agree that one of the brain's glories is that it can very rapidly draw analogies - that I can quickly produce a string of associations like, say, snake, rope, chain, spaghetti strand, - and you may quickly be able to continue that string with further associations, (like string). I believe that power is mainly based on look-up - literally finding matching shapes at speed. But I don't see the brain as checking through huge numbers of such shapes. (It would be enormously demanding on resources, given that these are complex pictures, no?). As evidence , I'd point to what happens if you try to keep producing further analogies. The brain rapidly slows down. It gets harder and harder. And yet you will be able to keep producing further examples from memory virtually for ever - just slower and slower. Relevant images/ concepts are there, but it's not easy to access them. That's why copywriters get well paid to, in effect, keep searching for similar analogies (as cool/refreshing as...). It's hard work. If that many relevant shapes were being unconsciously activated as you seem to be suggesting, it shouldn't be such protracted work. The brain can literally connect any thing to any other thing with, so to speak, 6 degrees of separation - but I don't think it can conect that many things at once. I accept that this is still neuroscientifically an open issue, ( I'd be grateful for pointers to the research you're referring to). But I would have thought it obvious that the brain has massively inferior search capabilities to those of computers - that, surely, is a major reasonwhy we invented computers in the first place - they're a massive extension of our powers. And yet the brain can draw analogies, and basically, with minor exceptions, computers still can't. I think it's clear that computers won't catch up here by quantitatively increasing their powers still further. If you're digging a hole in the wrong place, digging further quicker won't help. (I'm arguing a variant of your own argument against Edward P!). But of course when your education and technology dispose you to dig in just those places, it's extremely hard to change your ways - or even believe, pace Edward, that change is necessary at all. After all, look at the size of those holes.. surely, we'll hit the Promised Land anytime now. P.S. In general, the brain is hugely irrational - it can only maintain a reflective, concentrated train of thought for literally seconds, not minutes before going off at tangents. It continually and necessarily jumps to conclusions. Such irrationality is highly adaptive in a fast-moving world where you can't hang around thinking about things for long. The idea that this same brain is systematically, thoroughly searching through, let's say, thousands or millions of variants on ideas, seems to me seriously at odds with this irrationality. (But I'm interested in all relevant research). - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71651016-b43e51
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
--- Mike Tintner [EMAIL PROTECTED] wrote: On the one hand, we can perhaps agree that one of the brain's glories is that it can very rapidly draw analogies - that I can quickly produce a string of associations like, say, snake, rope, chain, spaghetti strand, - and you may quickly be able to continue that string with further associations, (like string). I believe that power is mainly based on look-up - literally finding matching shapes at speed. But I don't see the brain as checking through huge numbers of such shapes. (It would be enormously demanding on resources, given that these are complex pictures, no?). Semantic models learn associations by proximity in the training text. The degree to which you associate snake and rope depends on how often these words appear near each other. You can create an association matrix A, e.g. A[snake][rope] is the degree of association between these words. Among the most successful of these models is latent semantic analysis (LSA), where A is factored: A = USV by singular value decomposition (SVD), such that U and V are orthonormal and S is diagonal, and then discard all but the largest elements of S. In a typical LSA model, A is 20K by 20K, and S is reduced to about 200. This approximates A to two 20K by 200 matrices, using about 2% as much space. One effect of lossy compression by LSA is to derive associations by the transitive property of semantics. For example, if snake is associated with rope and rope with chain, then the LSA approximation will derive an association of snake with chain even if it was not seen in the training data. SVD has an efficient parallel implementation. It is most easily visualized as a 20K by 200 by 20K 3-layer linear neural network [1]. But this really should not be surprising, because natural language evolved to be processed efficiently on a slow but highly parallel computer. 1. Gorrell, Genevieve (2006), Generalized Hebbian Algorithm for Incremental Singular Value Decomposition in Natural Language Processing, Proceedings of EACL 2006, Trento, Italy. http://www.aclweb.org/anthology-new/E/E06/E06-1013.pdf -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71675396-27fd0e
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
RICHARD LOOSEMORE I cannot even begin to do justice, here, to the issues involved in solving the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space You talk as if an extension of some current strategy will solve this ... but it is not at all clear that any current strategy for solving this problem actually does scale up to a full solution to the problem. I don't care how many toy examples you come up with, you have to show a strategy for dealing with some of the core issues, AND reasons to believe that those strategies really will work (other than I find them quite promising). Not only that, but there at least some people (to wit, myself) who believe there are positive reasons to believe that the current strategies *will* not scale up. ED PORTER I don't know if you read the Shastri paper I linked to or not, but it shows we do know how to do many of the types of implication which are used in NL. What he shows needs some extensions, so it is more generalized, but it and other known inference schemes explain a lot of how text understanding could be done. With regard to the scaling issue, it is a real issue. But there are multiple reasons to believe the scaling problems can be overcome. Not proofs, Richard, so you are entitled to your doubts. But open your mind to the possibilities they present. They include: -1-the likely availability of roughly brain level representational, computational, and interconnect capacities within the several hundred thousand to 1 million dollar range in seven to ten years. -2-the fact that human experience and representation does not explode combinatorially. Instead it is quite finite. It fits insides our heads. Thus, although you are dealing with extremely high dimensional spaces, most of that space is empty. There are know ways to deal with extremely high dimensional spaces while avoiding the exponential explosion made possible by such high dimensionality. Take the well know Growing Neural Gas (GNG) algorithm. It automatically creates a relative compact representation of a possibly infinite dimensional space, by allocated nodes to only those parts of the high dimensional space where there is stuff, or, if resource are more limited, where the most stuff is. Or take indexing, it takes one only to places in the hyperspace where something actually occurred or was thought about. One can have probabilitistically selected hierarchical indexing (something like John Rose suggested) which make indexing much more efficient. -3-experiential computers focus most learning, most models, and most search on things that actually have happened in the past or on things that in many ways are similar to what has happened in the past. This tends to greatly reduce representational and search spaces. When such a system synthesizes or perceives new patterns that have never happened before the system will normally have to explore large search spaces, but because of the capacity of brain level hardware it will have considerable capability to do so. The type of hardware that will be available for human-level agi in the next decade will probably have sustainable cross sectional bandwidths of 10G to 1T messages/sec with 64Byte payloads/msg. With branching tree activations and the fact that many messages will be regional, the total amount of messaging could well be 100G to 100T such msg/sec. Lets assume our hardware has 10T msg/sec and that we want to read 10 words a second. That would allow 1T msg/word. With a dumb spreading activation rule that would allow you to: active the 30K most probably implications; and for each of them the 3K most probable implications; and for each of them the 300 most probable implications; and for each of them the 30 most probable implications. As dumb as this method of inferencing would be, it actually would make a high percent of the appropriate multi-step inferences, particularly when you consider that the probability of activation at the successive stages would be guided by probabilities from other activations in the current context. Of course there are much more intelligent ways to guide activation that this. Also it is important to understand that at every level in many of the searches or explorations in such a system there will be guidance and limitations provided by similar models from past experience, greatly reducing the amount of or the number of explorations that are required to produce reasonable results. -4-Michael Collins a few years ago had was many AI researches considered to be the best grammatical parser, which used the kernel trick to effectively match parse trees in, I think it was, 500K dimensions. By use of the Kernel trick the actual computation usually was performed in a small subset of these dimensions and the parser was
[agi] Priming of associates [WAS Re: Hacker intelligence level]
Mike Tintner wrote: RL: One thing that can be easily measured is the activation of lexical items related in various ways to a presented word (i.e. show the subject the word Doctor and test to see if the word Nurse gets activated). It turns out that within an extremely short time of the forst word being seen, a very large numbmer of other words have their activations raised significantly. Now, whichever way you interpret these (so called priming) results, one thing is not in doubt: there is massively parallel activation of lexical units going on during language processing. Thanks for reply. How many associations are activated? How do we know neuroscientifically they are associations to the words being processed and not something else entirely? Out of interest, can you give me a ball park estimate of how many associations you personally think are activated, say, in in a few seconds, in processing sentences like: The doctor made a move on the nurse. Relationships between staff in health organizations are fraught with complexities No, I'm not trying to be ridiculously demanding or asking you to be ridiculously exact. As you probably know by now, I see the processing of sentences as involving several levels, especially for the second sentence, but I don't see the number of associations as that many. Let's be generous and guess hundreds for the items in the above sentences. But a computer program, as I understand, will be typically searching through anywhere between thousands, millions and way upwards. I am not sure how many, but my understanding of the literature is that very large numbers show priming, and that it is proportional to association strength or semantic relatedness, measured some other way. The speed is also interesting: the effect can occur within about 140 ms of the word being shown. At the brain's clock speed, that would be maybe 300 clocks. Not much time for anything except parallel processing in that short a time. On the one hand, we can perhaps agree that one of the brain's glories is that it can very rapidly draw analogies - that I can quickly produce a string of associations like, say, snake, rope, chain, spaghetti strand, - and you may quickly be able to continue that string with further associations, (like string). I believe that power is mainly based on look-up - literally finding matching shapes at speed. But I don't see the brain as checking through huge numbers of such shapes. (It would be enormously demanding on resources, given that these are complex pictures, no?). It would be a problem if it were checking pictures. The standard model is that there are links already established between concepts, as a result of experience, and all it is doing is propagating activation along links, in parallel. It does depend on whether these are analogies or just associations. (related, of course). As evidence , I'd point to what happens if you try to keep producing further analogies. The brain rapidly slows down. It gets harder and harder. And yet you will be able to keep producing further examples from memory virtually for ever - just slower and slower. Relevant images/ concepts are there, but it's not easy to access them. That's why copywriters get well paid to, in effect, keep searching for similar analogies (as cool/refreshing as...). It's hard work. If that many relevant shapes were being unconsciously activated as you seem to be suggesting, it shouldn't be such protracted work. Generating analogies of that sort would not be the same effect. I make no claims for that specific thing, only for the activation of semantically related or associated concepts. The brain can literally connect any thing to any other thing with, so to speak, 6 degrees of separation - but I don't think it can conect that many things at once. That is just low-level (neuron connectivity). That doesn't speak to higher level systems. I accept that this is still neuroscientifically an open issue, ( I'd be grateful for pointers to the research you're referring to). But I would have thought it obvious that the brain has massively inferior search capabilities to those of computers - that, surely, is a major reasonwhy we invented computers in the first place - they're a massive extension of our powers. Too many imponderable here, but in general, no: the brain may still have the edge for some types of parallel search. And yet the brain can draw analogies, and basically, with minor exceptions, computers still can't. Now you skip to different issue: we don't know the *mechanism* involved in analogy finding. That is why compters cannot do it. It is not that computers lack the processing power or connectivity. I think it's clear that computers won't catch up here by quantitatively increasing their powers still further. If you're digging a hole in the wrong place, digging further quicker won't help. (I'm arguing a variant of your own
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
MIKE TINTNER Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, ED PORTER How do you find the right set of relatively few computations and/or models that are appropriate in a complex context without massive computation? Ed, Contrary to my PM, maybe I should answer this in more precise detail.My hypothesis is as follows: the brain does most of its thinking, and particularly adaptive thinking, by look-up not by blind search. How can you or I deal with : Get that box out of this house now.. How is it say, that I will be able to think of a series of ideas like get ten men to carry it, get a fork-lift truck to move it, use large levers, get hold of some heavy ropes ... etc etc. straight off the top of my head in well under a minute? All of those ideas are derived from visual/sensory images/ schemas of large objects being moved. The brain does not, I suggest, consult digital/ verbal lists or networks of verbal ideas about moving boxes out of houses or any similar set of verbal concepts, (except v. occasionally). How then does the brain rapidly pull relevant large-object-moving shapes out of memory? (There are obviously more operations involved here than just shape search, but that's what I want to concentrate on). Now this is where I confess again to being a general techno-idiot (although I suspect that in this particular area most of you may be, too). My confused idea is that if you have a stack of shapes, there are ways to pull out/ spot the relevant ones quickly without sorting through the stack one by one. I think Hawkins suggests something like this in ON INtelligence. Maybe you can have thoughts about this. (Alternatively, the again confused idea occurs that certain neuronal areas, when stimulated with a certain shape, may be able to remember similar shapes that have been there before - v. loosely as certain metals when heated, can remember/ resume old forms) Whatever, I am increasingly confident that the brain does work v. extensively by matching shapes physically, (rather than by first converting them into digital/symbolic form). And I recommend here Sandra Blakeslee's latest book on body maps - the opening Ramachandran quote - When a reporter asked the famous biologist JBS Haldane what his biological studies had taught about God, Haldane replied:The creator if he exists must have an inordinate fondness for beetles since there are more species of beetle than any other group of living creqtures. By the same token, a neurologist might conclude that God is a cartographer. He must have an inordinate fondness for maps, for everywhere you look in the brain maps abound. If I'm headed even loosely in the right direction here, only analog computation will be able to handle the kind of rapid shape matching and searches I'm talking about, as opposed to the inordinately long, blind symbolic searches of digital computation. And you're going to need a whole new kind of computer. But none of you guys are prepared to even contemplate that. P.S. One important feature of shape searches by contrast with digital, symbolic searches is that you don't make mistakes. IOW when we think about a problem like getting the box out of a house, all our ideas, I suggest, will be to some extent relevant. They may not totally solve the problem, but they will fit some of the requirements, precisely because they have been derived by shape comparison. When a computer blindly searches lists of symbols by contrast, most of them of course are totally irrelevant. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71680486-77dd12
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: RICHARD LOOSEMORE I cannot even begin to do justice, here, to the issues involved in solving the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space You talk as if an extension of some current strategy will solve this ... but it is not at all clear that any current strategy for solving this problem actually does scale up to a full solution to the problem. I don't care how many toy examples you come up with, you have to show a strategy for dealing with some of the core issues, AND reasons to believe that those strategies really will work (other than I find them quite promising). Not only that, but there at least some people (to wit, myself) who believe there are positive reasons to believe that the current strategies *will* not scale up. ED PORTER I don't know if you read the Shastri paper I linked to or not, but it shows we do know how to do many of the types of implication which are used in NL. What he shows needs some extensions, so it is more generalized, but it and other known inference schemes explain a lot of how text understanding could be done. With regard to the scaling issue, it is a real issue. But there are multiple reasons to believe the scaling problems can be overcome. Not proofs, Richard, so you are entitled to your doubts. But open your mind to the possibilities they present. They include: -1-the likely availability of roughly brain level representational, computational, and interconnect capacities within the several hundred thousand to 1 million dollar range in seven to ten years. -2-the fact that human experience and representation does not explode combinatorially. Instead it is quite finite. It fits insides our heads. Thus, although you are dealing with extremely high dimensional spaces, most of that space is empty. There are know ways to deal with extremely high dimensional spaces while avoiding the exponential explosion made possible by such high dimensionality. Take the well know Growing Neural Gas (GNG) algorithm. It automatically creates a relative compact representation of a possibly infinite dimensional space, by allocated nodes to only those parts of the high dimensional space where there is stuff, or, if resource are more limited, where the most stuff is. Or take indexing, it takes one only to places in the hyperspace where something actually occurred or was thought about. One can have probabilitistically selected hierarchical indexing (something like John Rose suggested) which make indexing much more efficient. I'm sorry, but this is not addressing the actual issues involved. You are implicitly assuming a certain framework for solving the problem of representing knowledge ... and then all your discussion is about whether or not it is feasible to implement that framework (to overcome various issues to do with searches that have to be done within that framework). But I am not challenging the implementation issues, I am challenging the viability of the framework itself. My mind is completely open. But right now I raised one issue, and this is not answered. I am talking about issues that could prevent that framework from ever working no matter how much computing power is available. You must be able to see this: you are familiar with the fact that it is possible to frame a solution to certain problems in such a way that the proposed solution is KNOWN to not converge on an answer? An answer can be perfectly findable IF you use a different representation, but there are some ways of representing the problem that lead to a type of solution that is completely incomputable. This is an analogy: I suggest to you that the framework you have in mind when you discuss the solution of the AGI problem is like those broken representations. -3-experiential computers focus most learning, most models, and most search on things that actually have happened in the past or on things that in many ways are similar to what has happened in the past. This tends to greatly reduce representational and search spaces. When such a system synthesizes or perceives new patterns that have never happened before the system will normally have to explore large search spaces, but because of the capacity of brain level hardware it will have considerable capability to do so. The type of hardware that will be available for human-level agi in the next decade will probably have sustainable cross sectional bandwidths of 10G to 1T messages/sec with 64Byte payloads/msg. With branching tree activations and the fact that many messages will be regional, the total amount of messaging could well be 100G to 100T such msg/sec. Lets assume our hardware has 10T msg/sec and that we want to read 10 words a second. That would allow 1T msg/word. With a dumb spreading
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: RICHARD LOOSEMORE I cannot even begin to do justice, here, to the issues involved in solving the high dimensional problem of seeking to understand the meaning of text, which often involve multiple levels of implication, which would normally be accomplished by some sort of search of a large semantic space You talk as if an extension of some current strategy will solve this ... but it is not at all clear that any current strategy for solving this problem actually does scale up to a full solution to the problem. I don't care how many toy examples you come up with, you have to show a strategy for dealing with some of the core issues, AND reasons to believe that those strategies really will work (other than I find them quite promising). Not only that, but there at least some people (to wit, myself) who believe there are positive reasons to believe that the current strategies *will* not scale up. ED PORTER I don't know if you read the Shastri paper I linked to or not, but it shows we do know how to do many of the types of implication which are used in NL. What he shows needs some extensions, so it is more generalized, but it and other known inference schemes explain a lot of how text understanding could be done. With regard to the scaling issue, it is a real issue. But there are multiple reasons to believe the scaling problems can be overcome. Not proofs, Richard, so you are entitled to your doubts. But open your mind to the possibilities they present. They include: -1-the likely availability of roughly brain level representational, computational, and interconnect capacities within the several hundred thousand to 1 million dollar range in seven to ten years. -2-the fact that human experience and representation does not explode combinatorially. Instead it is quite finite. It fits insides our heads. Thus, although you are dealing with extremely high dimensional spaces, most of that space is empty. There are know ways to deal with extremely high dimensional spaces while avoiding the exponential explosion made possible by such high dimensionality. Take the well know Growing Neural Gas (GNG) algorithm. It automatically creates a relative compact representation of a possibly infinite dimensional space, by allocated nodes to only those parts of the high dimensional space where there is stuff, or, if resource are more limited, where the most stuff is. Or take indexing, it takes one only to places in the hyperspace where something actually occurred or was thought about. One can have probabilitistically selected hierarchical indexing (something like John Rose suggested) which make indexing much more efficient. -3-experiential computers focus most learning, most models, and most search on things that actually have happened in the past or on things that in many ways are similar to what has happened in the past. This tends to greatly reduce representational and search spaces. When such a system synthesizes or perceives new patterns that have never happened before the system will normally have to explore large search spaces, but because of the capacity of brain level hardware it will have considerable capability to do so. The type of hardware that will be available for human-level agi in the next decade will probably have sustainable cross sectional bandwidths of 10G to 1T messages/sec with 64Byte payloads/msg. With branching tree activations and the fact that many messages will be regional, the total amount of messaging could well be 100G to 100T such msg/sec. Lets assume our hardware has 10T msg/sec and that we want to read 10 words a second. That would allow 1T msg/word. With a dumb spreading activation rule that would allow you to: active the 30K most probably implications; and for each of them the 3K most probable implications; and for each of them the 300 most probable implications; and for each of them the 30 most probable implications. As dumb as this method of inferencing would be, it actually would make a high percent of the appropriate multi-step inferences, particularly when you consider that the probability of activation at the successive stages would be guided by probabilities from other activations in the current context. Of course there are much more intelligent ways to guide activation that this. Also it is important to understand that at every level in many of the searches or explorations in such a system there will be guidance and limitations provided by similar models from past experience, greatly reducing the amount of or the number of explorations that are required to produce reasonable results. -4-Michael Collins a few years ago had was many AI researches considered to be the best grammatical parser, which used the kernel trick to effectively match parse trees in, I think it was, 500K dimensions. By use of the Kernel trick the actual computation usually was performed in a small subset of these dimensions
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Matt: Semantic models learn associations by proximity in the training text. The degree to which you associate snake and rope depends on how often these words appear near each other Correct me - but it's the old, old problem here, isn't it? Those semantic models/programs won't be able to form any *new* analogies, will they? Or understand newly minted analogies in texts? And I'm v. dubious about their powers to even form valid associations of much value in the ways you describe from existing texts. You're saying that there's a semantic model/program that can answer, if asked,: yes - 'snake, chain, rope, spaghetti strand' is a legitimate/ valid series of associations/ yes, they fit together (based on previous textual analysis) ? or: the odd one out in 'snake/ chain/ cigarette/ rope is 'cigarette'? I have yet to find or be given a single useful analogy drawn by computers (despite asking many times). The only kind of analogy I can remember here is Ed, I think, pointing to Hofstader's analogies along the lines of xxyy is like . Not exactly a big deal. No doubt there must be more, but my impression is that in general computers are still pathetic here. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71683316-d0bd3c
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
--- Ed Porter [EMAIL PROTECTED] wrote: We do not know the number and width of the spreading activation that is necessary for human level reasoning over world knowledge. Thus, we really don't know how much interconnect is needed and thus how large of a P2P net would be needed for impressive AGI. But I think it would have to be larger than say 10K nodes. In complex systems on the boundary between stability and chaos, the degree of interconnectedness per node is constant. Complex systems always evolve to this boundary because stable systems aren't complex and chaotic systems can't be incrementally updated. In my thesis ( http://cs.fit.edu/~mmahoney/thesis.html ) I did not estimate the communication bandwidth. But it is O(n log n) because the distance between nodes grows as O(log n). For each message sent or received, a node must also relay O(log n) messages. If the communication protocol is natural language text, then I am pretty sure our existing networks can handle it. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71684400-910726
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
--- Mike Tintner [EMAIL PROTECTED] wrote: Matt: Semantic models learn associations by proximity in the training text. The degree to which you associate snake and rope depends on how often these words appear near each other Correct me - but it's the old, old problem here, isn't it? Those semantic models/programs won't be able to form any *new* analogies, will they? Or understand newly minted analogies in texts? And I'm v. dubious about their powers to even form valid associations of much value in the ways you describe from existing texts. You're saying that there's a semantic model/program that can answer, if asked,: yes - 'snake, chain, rope, spaghetti strand' is a legitimate/ valid series of associations/ yes, they fit together (based on previous textual analysis) ? Yes, because each adjacent pair of words has a high frequency of co-occurrence in a corpus of training text. or: the odd one out in 'snake/ chain/ cigarette/ rope is 'cigarette'? Yes, because cigarette does not have a high co-occurrence with the other words. I have yet to find or be given a single useful analogy drawn by computers (despite asking many times). The only kind of analogy I can remember here is Ed, I think, pointing to Hofstader's analogies along the lines of xxyy is like . Not exactly a big deal. No doubt there must be more, but my impression is that in general computers are still pathetic here. This simplistic vector space model I described has been used to pass the word analogy section of the SAT exams. See: Turney, P., Human Level Performance on Word Analogy Questions by Latent Relational Analysis (2004), National Research Council of Canada, http://iit-iti.nrc-cnrc.gc.ca/iit-publications-iti/docs/NRC-47422.pdf -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71685861-05fe0f
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Mike -Original Message- From: Mike Tintner [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 8:25 PM To: agi@v2.listbox.com Subject: Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research] MIKE TINTNER Isn't it obvious that the brain is able to understand the wealth of language by relatively few computations - quite intricate, hierarchical, multi-levelled processing, ED PORTER How do you find the right set of relatively few computations and/or models that are appropriate in a complex context without massive computation? MIKE TINTNER How then does the brain rapidly pull relevant large-object-moving shapes out of memory? (There are obviously more operations involved here than just shape search, but that's what I want to concentrate on). Now this is where I confess again to being a general techno-idiot (although I suspect that in this particular area most of you may be, too). My confused idea is that if you have a stack of shapes, there are ways to pull out/ spot the relevant ones quickly without sorting through the stack one by one. I think Hawkins suggests something like this in ON INtelligence. Maybe you can have thoughts about this. ED One way is by indexing some thing by its features, but this is a form of a search, which if done completely activates each occurrence of each feature searched for, and then selects the one or more pattern with the best activation score. Others on the list can probably name other methods Another used in perception is to hierarchically match inputs against patterns that represent given shapes under different conditions. MIKE TINTNER (Alternatively, the again confused idea occurs that certain neuronal areas, when stimulated with a certain shape, may be able to remember similar shapes that have been there before - v. loosely as certain metals when heated, can remember/ resume old forms) Whatever, I am increasingly confident that the brain does work v. extensively by matching shapes physically, (rather than by first converting them into digital/symbolic form). And I recommend here Sandra Blakeslee's latest book on body maps - the opening Ramachandran quote - ED there clearly is some shape matching in the brain. MIKE TINTNER P.S. One important feature of shape searches by contrast with digital, symbolic searches is that you don't make mistakes. IOW when we think about a problem like getting the box out of a house, all our ideas, I suggest, will be to some extent relevant. They may not totally solve the problem, but they will fit some of the requirements, precisely because they have been derived by shape comparison. When a computer blindly searches lists of symbols by contrast, most of them of course are totally irrelevant. ED Yes, but there are a lot of types of thinking that cannot be done by shape alone, and shape is actually much more complicated than shape. There is shape, and shape distorted by perspective, and shape changed by bending, and shape changed by size. There is shape of objects, shape of trajectories, 2d shapes, 3d shapes. There are visual memories, where we don't really remember all the shapes, but instead remember the types of things that were their and fill in most of the actual shapes. In sum, it's a lot more complicated that just finding a matching photograph. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71691780-efaeb1attachment: winmail.dat
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
RICHARD LOOSEMORE= I'm sorry, but this is not addressing the actual issues involved. You are implicitly assuming a certain framework for solving the problem of representing knowledge ... and then all your discussion is about whether or not it is feasible to implement that framework (to overcome various issues to do with searches that have to be done within that framework). But I am not challenging the implementation issues, I am challenging the viability of the framework itself. ED PORTER= So what is wrong with my framework? What is wrong with a system of recording patterns, and a method for developing compositions and generalities from those patterns, in multiple hierarchical levels, and for indicating the probabilities of certain patterns given certain other pattern etc? I know it doesn't genuflect before the alter of complexity. But what is wrong with the framework other than the fact that it is at a high level and thus does not explain every little detail of how to actually make an AGI work? RICHARD LOOSEMORE= These models you are talking about are trivial exercises in public relations, designed to look really impressive, and filled with hype designed to attract funding, which actually accomplish very little. Please, Ed, don't do this to me. Please don't try to imply that I need to open my mind any more. Th implication seems to be that I do not understand the issues in enough depth, and need to do some more work to understand you points. I can assure you this is not the case. ED PORTER= Shastri's Shruiti is a major piece of work. Although it is a highly simplified system, for its degree of simplification it is amazingly powerful. It has been very helpful to my thinking about AGI. Please give me some excuse for calling it trivial exercise in public relations. I certainly have not published anything as important. Have you? The same for Mike Collins's parsers which, at least several years ago I was told by multiple people at MIT was considered one of the most accurate NL parsers around. Is that just a trivial exercise in public relations? With regard to Hecht-Nielsen's work, if it does half of what he says it does it is pretty damned impressive. It is also a work I think about often when thinking how to deal with certain AI problems. Richard if you insultingly dismiss such valid work as trivial exercises in public relations it sure as hell seems as if either you are quite lacking in certain important understandings -- or you have a closed mind -- or both. Ed Porter - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71696956-846847
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Richard Loosemore= None of the above is relevant. The issue is not whether toy problems set within the current paradigm can be done with this or that search algorithm, it is whether the current paradigm can be made to converge at all for non-toy problems. Ed Porter= Richard, I wouldn't call a state of the art NL parser that matches parse trees in 500K dimensions a toy problem. Yes, it is much less than a complete human brain, but it is not a toy problem. With regard to Hecht-Nielsen's sentence completion program it is arguably a toy problem, but it operates extremely efficiently (i.e., converges) in an astronomically large search space, with a significant portion of that search space having some arguable activation. The fact that there is such efficient convergence in such a large search space is meaningful, and the fact that you just dismiss it, as you did in your last email as a trivial publicity stunt is also meaningful. Ed Porter - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71705619-d121f2
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Matt, IN my Mon 12/3/2007 8:17 PM post to John Rose from which your are probably quoting below I discussed the bandwidth issues. I am assuming nodes directly talk to each other, which is probably overly optimistic, but still are limited by the fact that each node can only receive somewhere roughly around 100 128 byte messages a second. Unless you have a really big P2P system, that just isn't going to give you much bandwidth. If you had 100 million P2P nodes it would. Thus, a key issue is how many participants is an AGI-at-Home P2P system going to get. I mean, what would motivate the average American, or even the average computer geek turn over part of his computer to it? It might not be an easy sell for more than several hundred or several thousand people, at least until it could do something cool, like index their videos for them, be a funny chat bot, or something like that. Ed Porter -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 8:51 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] --- Ed Porter [EMAIL PROTECTED] wrote: We do not know the number and width of the spreading activation that is necessary for human level reasoning over world knowledge. Thus, we really don't know how much interconnect is needed and thus how large of a P2P net would be needed for impressive AGI. But I think it would have to be larger than say 10K nodes. In complex systems on the boundary between stability and chaos, the degree of interconnectedness per node is constant. Complex systems always evolve to this boundary because stable systems aren't complex and chaotic systems can't be incrementally updated. In my thesis ( http://cs.fit.edu/~mmahoney/thesis.html ) I did not estimate the communication bandwidth. But it is O(n log n) because the distance between nodes grows as O(log n). For each message sent or received, a node must also relay O(log n) messages. If the communication protocol is natural language text, then I am pretty sure our existing networks can handle it. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71708450-da8cab
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Matt, In addition to my last email, I don't understand what your were saying below about complexity. Are you saying that as a system becomes bigger it naturally becomes unstable, or what? Ed Porter -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 8:51 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] --- Ed Porter [EMAIL PROTECTED] wrote: We do not know the number and width of the spreading activation that is necessary for human level reasoning over world knowledge. Thus, we really don't know how much interconnect is needed and thus how large of a P2P net would be needed for impressive AGI. But I think it would have to be larger than say 10K nodes. In complex systems on the boundary between stability and chaos, the degree of interconnectedness per node is constant. Complex systems always evolve to this boundary because stable systems aren't complex and chaotic systems can't be incrementally updated. In my thesis ( http://cs.fit.edu/~mmahoney/thesis.html ) I did not estimate the communication bandwidth. But it is O(n log n) because the distance between nodes grows as O(log n). For each message sent or received, a node must also relay O(log n) messages. If the communication protocol is natural language text, then I am pretty sure our existing networks can handle it. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71710422-50e2fa
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
On Thursday 29 November 2007, Ed Porter wrote: Somebody (I think it was David Hart) told me there is a shareware distributed web crawler already available, but I don't know the details, such as how good or fast it is. http://grub.org/ Previous owner went by the name of 'kordless'. I found him on Slashdot. - Bryan - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71712384-417a60
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: Richard Loosemore= None of the above is relevant. The issue is not whether toy problems set within the current paradigm can be done with this or that search algorithm, it is whether the current paradigm can be made to converge at all for non-toy problems. Ed Porter= Richard, I wouldn't call a state of the art NL parser that matches parse trees in 500K dimensions a toy problem. Yes, it is much less than a complete human brain, but it is not a toy problem. This is a toy problem. Parsing is a deep problem? Do you understand the relationship between parsing NL and extracting semantics? Do you understand what this great NL parser would do if confronted with a syntactically incorrect but contextually meaningful sentence? Has it been analysed to see what its behavior is on ambiguous sentences? Could it learn to cope with someone speaking a pidgin version of NL, or would someone have to write an entire grammar for the language before the system could even start parsing it? Can it generate syntactically correct sentences that express an idea? Can it cope with speech errors, recgnising the nature o fteh error and backfilling, or does it just collapse with no viable parse? Would the parser have to be completely rewritten in the future when someone else finally solves the problem of representing the semantics of language? Finally, if you are impressed by the claim about 500K dimensions then what can I say? Can you explain to me in what sense it matches parse trees in 500K dimensions, and why that is so impressive? Perhaps I am being unnecessarily hard on you, Ed. I don't mean to be personally rude, you know, but it is sometimes exhausting to have someone trying to teach you how to suck eggs Richard Loosemore With regard to Hecht-Nielsen's sentence completion program it is arguably a toy problem, but it operates extremely efficiently (i.e., converges) in an astronomically large search space, with a significant portion of that search space having some arguable activation. The fact that there is such efficient convergence in such a large search space is meaningful, and the fact that you just dismiss it, as you did in your last email as a trivial publicity stunt is also meaningful. Ed Porter - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71714474-5576ff
Re: [agi] RE:P2P and/or communal AGI development [WAS Hacker intelligence level...]
On Monday 03 December 2007, Mike Dougherty wrote: I believe the next step of such a system is to become an abstraction between the user and the network they're using. So if you can hook into your P2P network via a firefox extension, (consider StumbleUpon or Greasemonkey) so it (the agent) can passively monitor your web interaction - then it could be learn to screen emails (for example) or pre-chew either your first 10 google hits or summarize the next 100 for relevance. I have been told that by the time you have an agent doing this well, you'd already have AGI - but i can't believe this kind of data mining is beyond narrow AI (or requires fully general adaptive intelligence) Another method of doing search agents, in the mean time, might be to take neural tissue samples (or simple scanning of the brain) and try to simulate a patch of neurons via computers so that when the simulated neurons send good signals, the search agent knows that there has been a good match that excites the neurons, and then tells the wetware human what has been found. The problem that immediately comes to mind is that neurons for such searching are probably somewhere deep in the prefrontal cortex ... does anybody have any references to studies done with fMRI on people forming Google queries? - Bryan - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71715011-399ee5
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed Porter wrote: RICHARD LOOSEMORE= I'm sorry, but this is not addressing the actual issues involved. You are implicitly assuming a certain framework for solving the problem of representing knowledge ... and then all your discussion is about whether or not it is feasible to implement that framework (to overcome various issues to do with searches that have to be done within that framework). But I am not challenging the implementation issues, I am challenging the viability of the framework itself. ED PORTER= So what is wrong with my framework? What is wrong with a system of recording patterns, and a method for developing compositions and generalities from those patterns, in multiple hierarchical levels, and for indicating the probabilities of certain patterns given certain other pattern etc? I know it doesn't genuflect before the alter of complexity. But what is wrong with the framework other than the fact that it is at a high level and thus does not explain every little detail of how to actually make an AGI work? RICHARD LOOSEMORE= These models you are talking about are trivial exercises in public relations, designed to look really impressive, and filled with hype designed to attract funding, which actually accomplish very little. Please, Ed, don't do this to me. Please don't try to imply that I need to open my mind any more. Th implication seems to be that I do not understand the issues in enough depth, and need to do some more work to understand you points. I can assure you this is not the case. ED PORTER= Shastri's Shruiti is a major piece of work. Although it is a highly simplified system, for its degree of simplification it is amazingly powerful. It has been very helpful to my thinking about AGI. Please give me some excuse for calling it trivial exercise in public relations. I certainly have not published anything as important. Have you? The same for Mike Collins's parsers which, at least several years ago I was told by multiple people at MIT was considered one of the most accurate NL parsers around. Is that just a trivial exercise in public relations? With regard to Hecht-Nielsen's work, if it does half of what he says it does it is pretty damned impressive. It is also a work I think about often when thinking how to deal with certain AI problems. Richard if you insultingly dismiss such valid work as trivial exercises in public relations it sure as hell seems as if either you are quite lacking in certain important understandings -- or you have a closed mind -- or both. Ed, You have no idea of the context in which I made that sweeping dismissal. If you have enough experience of research in this area you will know that it is filled with bandwagons, hype and publicity-seeking. Trivial models are presented as if they are fabulous achievements when, in fact, they are just engineered to look very impressive but actually solve an easy problem. Have you had experience of such models? Have you been around long enough to have seen something promoted as a great breakthrough even though it strikes you as just a trivial exercise in public relations, and then watch history unfold as the great breakthrough leads to absolutely nothing at all, and is then quietly shelved by its creator? There is a constant ebb and flow of exaggeration and retreat, exaggeration and retreat. You are familiar with this process, yes? This entire discussion baffles me. Does it matter at all to you that I have been working in this field for decades? Would you go up to someone at your local university and tell them how to do their job? Would you listen to what they had to say about issues that arise in their field of expertise, or would you consider your own opinion entirely equal to theirs, with only a tiny fraction of their experience? Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71711822-0e911b
Re: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]ED Yes, but there are a lot of types of thinking that cannot be done by shape alone, and shape is actually much more complicated than shape. There is shape, and shape distorted by perspective, and shape changed by bending, and shape changed by size. There is shape of objects, shape of trajectories, 2d shapes, 3d shapes. There are visual memories, where we don't really remember all the shapes, but instead remember the types of things that were their and fill in most of the actual shapes. In sum, it's a lot more complicated that just finding a matching photograph. Ed, I am not suggesting that shape matching is everything, merely that it is central to a great many of the brain's operations - and to its ability to search rapidly and briefly and locate analogical ideas (and if that's true, as I believe it is, then, sorry, AGI's stuckness is going to continue for a long time yet). The reason I'm replying though is a further thought occurred to me. Essentially I've been suggesting that the brain has some means to locate matching shapes quickly in very few operations where a digital computer laboriously searches through long lists or networks of symbols in a great many operations. One v. crude idea for the mechanism I suggested was that neuronal areas somehow retain memories of shapes, which can be stimulated by similar incoming shapes - so that analogies can be drawn with extreme rapidity, more or less on the spot. [Spot checks] It's occurred to me that this may well happen over and over throughout the body related brain areas. The same body areas that today feel stiff / expanded/ cold , felt loose/ contracted/ warm yesterday. The same hand that was a ball, and many other shapes, is now a fist. So perhaps these memories are all somehow laid on top of each in the same brain areas..Map upon map upon map .Just an extremely rough idea, but I think it does go some way to showing how shape matching could indeed be extremely rapid and effective in the brain, by contrast with computers' blind, disembodied search. It follows BTW re your points above, that the same brain areas will also retain many morphic variations on the same basic shapes - objects/cups seen say moving, from different angles, zooming in and out etc. And if it's true, as I believe, that the brain uses loose, highly flexible templates for visual object perception - then that too should mean that it will easily and rapidly be able to connect closely related shapes as in snake/ chain/ rope/ spaghetti strand. Analogies and perception are interwoven for the brain. Blakeslee makes a good deal of the brain using flexible, morphic body maps. Thanks for your reply. Further thoughts re mechanisms welcome. As Blakeslee points out, this whole area is just beginning to open up. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71724560-1bc574
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
Ed, Well it'd be nice having a supercomputer but P2P is a poor man's supercomputer and beggars can't be choosy. Honestly the type of AGI that I have been formulating in my mind has not been at all closely related to simulating neural activity through orchestrating partial and mass activations at low frequencies and I had been avoiding those contagious cog sci memes on purpose. But your expose on the subject is quite interesting and I wasn't that aware that that is how things have been being done. But getting more than a few thousand P2P nodes is difficult. Going from 10K to 20K nodes and up, getting more difficult to the point of being prohibitively expensive to being impossible or extremely lucky. There are ways to do it but according to your calculations the supercomputer mayt be more of a wise choice as going out and scrounging up funding for that would be easier. Still though (besides working on my group theory heavy design) exploring the crafting and chiseling of an activation model you are talking about to the P2P network could be fruitful. I feel that through a number of up front and unfortunately complicated design changes/adaptations that the activation orchestrations could be improved thus bringing down the message rate requirements, reducing activation requirements, depths and frequencies, through a sort of computational resource topology consumption, self-organizational design molding. You do indicate some dynamic resource adaption and things like intelligent inference guiding schemes in your description but it doesn't seem like it melts enough into the resource space. But having a design be less static risks excessive complications... A major problem though with P2P and the activation methodology is that there are so many variances in the latencies and availability that serious synchronicity/simultaneity issues would exist that even more messaging might be required. Since there are so many variables in public P2P, empirical data also would be necessary to get a gander on feasibility. I still feel strongly that the way to do AGI P2P (with public P2P as core not augmental) is to understand the grid, and build the AGI design based on that and what it will be in a few years, instead of taking a design and morphing it to the resource space. That said, there are finite designs that will work so the number of choices is few. John _ From: Ed Porter [mailto:[EMAIL PROTECTED] Sent: Monday, December 03, 2007 6:17 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] John, You raised some good points. The problem is that the total number of messages/sec that can be received is relatively small. It is not as if you are dealing with a multidimensional grid or toroidal net in which spreading tree activation can take advantage of the fact that the total parallel bandwidth for regional messaging can be much greater than the x-sectional bandwidth. In a system where each node is a server class node with multiple processors and 32 or 64Gbytes of ram, much of which is allocable to representation, sending messages to local indices on each machine could fairly efficiently activate all occurrences of something in a 32 to 64 TByte knowledge base with a max of 1K internode messages, if there was only 1K nodes. But in a PC based P2P system the ratio of nodes to representation space is high and the total number of 128 byte messages/sec than can be received is limited to about 100, so neither methods of trying to increase number of patterns than can be activated with the given interconnect of the network buy you as much. Human level context sensitivity arises because a large number of things that can depend on a large number of things in the current context are made aware of those dependencies. This takes a lot of messaging, and I don't see how a P2P system where each node can only receive about 100 relatively short messages a second is going to make this possible unless you had a huge number of nodes. As Richard Loosemore said in his Mon 12/3/2007 12:57 PM post. It turns out that within an extremely short time of the forst word being seen, a very large numbmer of other words have their activations raised significantly. Now, whichever way you interpret these (so called priming) results, one thing is not in doubt: there is massively parallel activation of lexical units going on during language processing. With special software, a $10M dollar supercomputer cluster with 1K nodes, 32TBytes of Ram, and a dual ported 20Mb infiniband interconnect send about 1
RE: [agi] What are the real unsolved issues in AGI [WAS Re: Hacker intelligence
From: Richard Loosemore [mailto:[EMAIL PROTECTED] Top three? I don't know if anyone ranks them. Try: 1) Grounding Problem (the *real* one, not the cheap substitute that everyone usually thinks of as the symbol grounding problem). 2) The problem of desiging an inference control engine whose behavior is predictable/governable etc. 3) A way to represent things - and in particular, uncertainty - without getting buried up to the eyeballs in (e.g.) temporal logics that nobody believes in. Take this with a pinch of salt: I am sure there are plenty of others. But if you came up with a *principled* solution to these issues, I'd be impressed. Thanks Richard for listing these. I have thought about 1 and 3 more than 2. But I am not sure I understand any of them fully enough to comment as I can tell from some of your other emails that many people have made speculations and declarations in guerilla warfare fashion afterwards yet to be found amidst the jungle causing frustration and anxiety to those who have truly dove deep into the conundrums. I think that you have some emails on the Grounding Problem way back. Also I feel that I have some good ideas on #3... including analogies. But inference control engines can be tough. That is more of a #1 I think. John - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=71731204-c8cf46