[agi] Motivational Systems again [WAS Re: Hacker intelligence level]
Mike Tintner wrote: RL:However, I have previously written a good deal about the design of different types of motivation system, and my understanding of the likely situation is that by the time we had gotten the AGI working, its motivations would have been arranged in such a way that it would *want* to be extremely cooperative. You do keep saying this. An autonomous mobile agent that did not have fundamentally conflicting emotions about each and every activity and part of the world, would not succeed and survive. An AGI that trusted and cooperated with every human would not succeed and survive. Conflict is essential in a world fraught with risks, where time and effort can be wasted, essential needs can be neglected, and life and limb are under more or less continuous threat. Conflict is as fundamental and essential to living creatures and any emotional system as gravity is to the physical world. (But I can't recall any mention of it in your writings about emotions). I think the way to resolve your questions is to analyze each one for hidden assumptions. First: you mention emotions many times, but I talk about *motivations*, not emotions. The thing we call emotions is closely related, but it is by no means the same, and it confuses the issues a lot to talk about one when we should be talking about the other. For example, I am writing this in a cafe, and after finishing the above paragraph I reached around and picked up my bagel and took a bit. Why did I do that? My motivation system has a set of things that are its current goals, and it did a smooth switch from having the [write down my thoughts] motivation in control to having the [take a bite of food] motivation in control. [takes sip of tea] But I feel no emotions at the moment. [bite]. Although, a short time ago I was trying to drive up a steep hill here in town, to get to an orchestra rehearsal, and had to abandon the attempt when it turned out that the road had a layer of ice underneath the snow, so that people were getting stuck on the hill, with wheels spinning. Person behind me hit the horn when they saw me taking a long time to get turned around safely, and hearing the horn made me feel a short burst of anger toward the idiot: that was an emotion, but it was not a motivation, except insomuch as I may have felt inclined to do something like turn and give them a signal of some kind. Motivations and emotions are linked, but it is more complex than you paint it, and although I often talk about a motivational/emotional system, it is the motivational part that is causally important. Now, your statement about how an AGI would not succeed and survive unless it felt conflict ... you are talking about creatures that use the standard biological design for a motivational system, which are forced to compete against other creatures using nothing but tooth and claw and wits. It is important to see that all of the conditions that force biological creatures to have to (a) use only the standard motivational system that nature designed, and (b) compete with other creatures using the same motivation system in an evolutionary context, DO NOT APPLY to AGI systems. This is such an obvious point that i find it difficult to begin explaining it. There are no selection pressures, no breeding, no limited lifespan, no conflict for food when there is the ability to engineer as much as necessary. There would quite likely be only one, or a limited number of AGIs on the whole planet, with no uncontrolled growth in their numbers. On and on the list goes. Every one of the factors that would cause a situation in which an AGI had to compete in order to succeed and survive are missing. The fundamental mistake (which many, many people make) is to simply assume that when an AGI is built, it will be dropped into the current design of world without substantially changing it: this is what I have called the Everything just the same, but with robots scenario. It makes no sense. They would change the world - immediately - so as to make all that compete to succeed nonsense a thing of the past. In the rest of your text, below, I will just highlight the places where you do the same thing: No one wants to be extremely cooperative with anybody. Of course: people are built with motivational systems specifically designed to NOT be especially cooperative. So what? Everyone wants and needs a balance of give-and-take. (And right away, an agent's interests and emotions of giving must necessarily conflict with their emotions of taking). An agent's? What agent? Designed with what kind of motivational system? And did you assume a human-similar one? Anything approaching a perfect balance of interests Balance of interests? Sounds like an evolutionary pressure kind of idea what evolution would that be? between extremely complex creatures/ psychoeconomies with extremely complex
RE: Re[2]: [agi] Self-building AGI
From: Dennis Gorelik [mailto:[EMAIL PROTECTED] There are programs that already write source code. The trick is to write working and useful apps. Many of the apps that write code basically take data and statically convert it to a source code representation. So a code generator may allow you to design a template existing in say XML and the generator then converts that to a source code structure. I'm not aware of ones that do more than that although I assume that there are experimental models. The most important part in writing useful apps is not about writing code. It's about gathering/defining requirements and designing the system. This is true in general but there are many apps whose innovativeness depends on writing the code better as the code is pushing state of the art in competition with other companies producing similar products. I think AGI falls into this category and just gathering requirements and filling in the blanks with standard coding isn't enough mainly due to the resource demands of AGI. More intelligent development environments (as well as many other tools) can help to build AGI, but development environments cannot build AGI by itself. If you look at nanotechnology one of the goals is to build machines that build machines. Couldn't software based AGI be similar? 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=71166132-535eca
Re[4]: [agi] Self-building AGI
John, If you look at nanotechnology one of the goals is to build machines that build machines. Couldn't software based AGI be similar? Eventually AGIs will be able to build other AGIs, but first AGI models won't be able to build any software. - 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=71194112-033f71
Re: [agi] Self-building AGI
Well... Have you ever tried to understand the code created by a decompiler? Especially if the original language that was compiled isn't the one that you are decompiling into... I'm not certain that just because we can look at the code of a working AGI, that we can therefore understand it. Not without a *LOT* of commentary and explanation of what the purpose of certain constructions/functions/etc. are. And maybe not then. Understanding a working AGI may require a deeper stack than we possess, or a greater ability to handle global variables. And when code is self-modifying it gets particularly tricky. I remember one sort routine that I encountered that called a short function in assembler. The reason for that call was a particular instruction that got overwritten with a binary value that depended on the parameters to the call. That instruction was executed during the comparison step of the loop, which was nowhere near the place where it was modified. It was a very short routine, but it took a long time to figure out. And it COULDN'T be translated into the calling language (FORTRAN). Well...a translation of sorts was possible, but it would have been over three times as long (with a separate loop for each kind of input parameter, plus some overhead for the testing and switching). Which would mean that some programs then wouldn't fit in the machine that was running them. It would also have been slower. Which means more expensive. Current languages don't have the same restrictions that Fortran had then, they've got different ones. I think the translator from actual code into code for humans would be considerably more complicated than an ordinary compiler, if the original code was written by an AI. Perhaps even it not. (Most decompilers only handle the easy parts of the code. Sometimes that's over 90%, but the code that's left can be tricky...particularly since most people no longer learn assembler. It's been perhaps 3 decades since I knew the assembler of the computer I was programming.) I don't think an optimizing AI would use any language other than assembler to write in, though perhaps a stylized one. (Not MIX or p-code. Possibly Parrot or jvm code. Possibly something created specially for it to use for it's purpose. Something regular, but easily translated into almost optimal assembler code for the machine that it was running on.) FWIW, most of this is just my ideas, without any backing of expert in the field since I've never built a mechanical translator. Dennis Gorelik wrote: Ed, 1) Human-level AGI with access to current knowledge base cannot build AGI. (Humans can't) 2) When AGI is developed, humans will be able to build AGI (by copying successful AGI models). The same with human-level AGI -- it will be able to copy successful AGI model. But that's not exactly self-building AGI you are looking for :-) 3) Humans have different level intelligence and skills. Not all are able to develop programs. The same is true regarding AGI. Friday, November 30, 2007, 10:20:08 AM, you wrote: Computers are currently designed by human-level intellitences, so presumably they could be designed by human-level AGI's. (Which if they were human-level in the tasks that are currently hard for computers means they could be millions of times faster than humans for tasks at which computers already way out perform us.) I mention that appropriate reading and training would be required, and I assumed this included access to computer science and computer technology sources, which the peasants of the middle age would not have access. So I don't understand your problem. -Original Message- From: Dennis Gorelik [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 1:01 AM To: agi@v2.listbox.com Subject: [agi] Self-building AGI Ed, At the current stages this may be true, but it should be remembered that building a human-level AGI would be creating a machine that would itself, with the appropriate reading and training, be able to design and program AGIs. No. AGI is not necessarily that capable. In fact first versions of AGI would not be that capable for sure. Consider middle age peasant, for example. Such peasant has general intelligence (GI part in AGI), right? What kind of training would you provide to such peasant in order to make him design AGI? - 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=71202814-4efdc4
Re: Re[2]: [agi] Lets count neurons
--- Dennis Gorelik [EMAIL PROTECTED] wrote: Matt, Using pointers saves memory but sacrifices speed. Random memory access is slow due to cache misses. By using a matrix, you can perform vector operations very fast in parallel using SSE2 instructions on modern processors, or a GPU. I doubt it. http://en.wikipedia.org/wiki/SSE2 - doesn't even mention parallel or matrix. It also doesn't mention that one instruction performs 8 16-bit signed multiply accumulates in parallel, or various other operations: 16 x 8 bits, 8 x 16 bits, 4 x 32 bits (int or float), or 2 x 64 bit (double) in 128 bit registers. To implement the neural network code in the PAQ compressor I wrote vector dot product code in MMX (4 x 16 bit for older processors) that is 6 times faster than optimized C/C++. There is an SSE2 version too. Actual difference in size would be 10 times, since your matrix is only 10% filled. For a 64K by 64K matrix, each pointer is 16 bits, or 1.6 bits per element. I think for neural networks of that size you could use 1 bit weights. -- 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=71210692-be60c4
RE: Re[2]: [agi] Self-building AGI
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. Ed Porter -Original Message- From: Dennis Gorelik [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 9:42 PM To: agi@v2.listbox.com Subject: Re[2]: [agi] Self-building AGI Ed, 1) Human-level AGI with access to current knowledge base cannot build AGI. (Humans can't) 2) When AGI is developed, humans will be able to build AGI (by copying successful AGI models). The same with human-level AGI -- it will be able to copy successful AGI model. But that's not exactly self-building AGI you are looking for :-) 3) Humans have different level intelligence and skills. Not all are able to develop programs. The same is true regarding AGI. Friday, November 30, 2007, 10:20:08 AM, you wrote: Computers are currently designed by human-level intellitences, so presumably they could be designed by human-level AGI's. (Which if they were human-level in the tasks that are currently hard for computers means they could be millions of times faster than humans for tasks at which computers already way out perform us.) I mention that appropriate reading and training would be required, and I assumed this included access to computer science and computer technology sources, which the peasants of the middle age would not have access. So I don't understand your problem. -Original Message- From: Dennis Gorelik [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 1:01 AM To: agi@v2.listbox.com Subject: [agi] Self-building AGI Ed, At the current stages this may be true, but it should be remembered that building a human-level AGI would be creating a machine that would itself, with the appropriate reading and training, be able to design and program AGIs. No. AGI is not necessarily that capable. In fact first versions of AGI would not be that capable for sure. Consider middle age peasant, for example. Such peasant has general intelligence (GI part in AGI), right? What kind of training would you provide to such peasant in order to make him design AGI? - 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/?; - 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=71219386-c7a577
RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research]
John, I tested Exeter, NH to LA at 5371kbs download, and 362Kbs upload. Strangelly my scores were slightly slower to NYC. Just throwing out ideas, for example, AGI-at-home PC's in the net could crawl the web looking for reasonable NL text. Use current NL tools to guess parse and word sense. For each word in text, send it and it surrounding text, Part of speech labeling, surrounding parse tree, and word sense guess, to another P2P node that specializes in that word in similar contexts and separately another P2P node that specializes in similar parse trees. These specialist node could then develop statistical models for word senses based on clustering or other technique. Then over time the statistical models would get send down to the reading nodes, and this EM cycle could be constantly repeated. Of course, without the cross-sectional bandwidth of proper AGI hardware, you are going to be severely limited from doing a lot of the things you would really like to be able to do. But I think you should be able to come up with pretty good word sense models. Ed Porter -Original Message- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 2:55 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] Ed, That is probably a good rough estimate. There are more headers for the more frequently transmitted smaller messages but a 16 byte header may be a bit large. Here is a speedtest link - http://www.speedtest.net/ My Comcast cable from Denver to NYC tests at 3537 kb/sec DL and 1588 kb/sec UL much larger than the calculations 256kb/sec. The variance between tests to the same location is quite large on the DL side but UL is relatively stable. Saturating either DL or UL would impact the other. You can get higher efficiencies if you use UDP transmission without message serialization. Also you can do things like compression, only sending changes, etc.. Distributed crawling with NL learning fits the scenario well since nodes download at higher speeds, process the download into a smaller dataset, then UL communicate the results to the server or share with peers. When one peer shares with many peers you hit the UL limit fast though so it has to be managed. And you have to figure out how the knowledge will be spread out - server centric, shared, hybrid... As the knowledge size increases with peer storage you have to come up with distributed indexes. John -Original Message- From: Ed Porter [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 12:06 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] John, Thanks. I guess that means and AGI-at-home system could be both up- loading and receiving about 27 1K msgs/sec if it wasn't being used for anything else and the networks weren't backed up in its neck of the woods. Presumably the number for say 128Byte messages would be say, roughly, 8 times faster (minus some percent for the latency associated with each message, so lets say roughly about 5 times faster or 135msg/sec. Is that reasonable? So, it seems for example it would be quite possible to do estimation/maximilation type NL learning in a distributed manner with a lot of cable-box connected PC's and a distributed web crawler. Ed Porter -Original Message- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 12:33 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] Hi Ed, If the peer is not running other apps utilizing the network it could do the same. Typically a peer first needs to locate other peers. There may be servers involved but these are just for the few bytes transmitted for public IP address discovery as many(or most) peers reside hidden behind NATs. DNS names also require lookups but these are just for doing the initial match of hostname to IP address, if DNS is used at all. We're just talking basic P2P, one peer talking to one other peer, nothing complicated. As you can imagine P2P can take on many flavors as the number of peers increases. John -Original Message- From: Ed Porter [mailto:[EMAIL PROTECTED] Sent: Friday, November 30, 2007 10:10 AM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] John, Thanks. Can P2P transmission match the same roughly 27 1Kmsg/sec rate as the client to server upload you discribed? Ed Porter -Original Message- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Thursday, November 29, 2007 11:40 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] OK for a guestimate take a half-way decent cable connection say Comcast on a good day with DL of 4mbits max and UL of 256kbits max with an undiscriminated protocol,