Re: [agi] It is more important how AGI works than what it can do.
oops, i meant 1895 ... damn that dyslexia ;-) ... though the other way was funnier, it was less accurate!! On Sat, Oct 11, 2008 at 8:55 AM, Ben Goertzel [EMAIL PROTECTED] wrote: I'm only pointing out something everybody here knows full well: embodiment in various forms has, so far, failed to provide any real help in cracking the NLU problem. Might it in the future? Sure. But the key word there is might. To me, you sound like a guy in 1985 saying So far, wings have failed to provide any real help in cracking the human flight problem ben g -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] Nothing will ever be attempted if all possible objections must be first overcome - Dr Samuel Johnson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] It is more important how AGI works than what it can do.
I'm only pointing out something everybody here knows full well: embodiment in various forms has, so far, failed to provide any real help in cracking the NLU problem. Might it in the future? Sure. But the key word there is might. To me, you sound like a guy in 1985 saying So far, wings have failed to provide any real help in cracking the human flight problem ben g --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] It is more important how AGI works than what it can do.
Dave, Well, I thought I'd described how pretty well. Even why. See my recent conversation with Dr. Heger on this list. I'll be happy to answer specific questions based on those explanations but I'm not going to repeat them here. Simply haven't got the time. Although I have not been asked to do so, I do feel I need to provide an ex post facto disclaimer. Here goes: I am aware of the approach being taken by Stephen Reed in the Texai project. I am currently associated with that project as a volunteer. What I have said previously in this regard) is, however, my own interpretation and opinion insofar as what I have said concerned tactics or strategies that may be similar to those being implemented in the Texai project. I'm pretty sure my interpretations and opinions are highly compatible with Steve's views even though they may not agree in every detail. My comments should NOT, however, be taken as an official representation of the Texai project's tactics, strategies or goals. End disclaimer. I was asked by Dr. Heger to go into some of the specifics of the strategy I had in mind. I honored his request and wrote quite extensively (for a list posting -- sorry 'bout that) about that strategy. I have not argued, nor do I intend to argue, that I have an approach to AGI that is better, faster or more economical than approach X. Instead, I have simply pointed out that NLU and embodiment problems have proven themselves to be extremely difficult (indeed, intractable to date). I, therefore, on those grounds alone, believe (and it's just an OPINION, although I believe a well-reasoned one) that we will get to a human-beneficial AGI sooner (and, I guess, probably, therefore, cheaper) if we side-step those two proven productivity sinks. For now. End of argument. I'm not trying to sell my AGI strategy or agenda to you or anyone else. Like many people on this list who have an opinion on these matters, I have a background as a practitioner in AI that goes back over twenty years. I've designed and written narrow-AI production (expert) system engines and been involved in knowledge engineering using those engines. The results of my efforts have saved large corporations millions of dollars (if not billions, over time). I can assure you that most of the humans who saw these systems come to life and out-perform their own human experts, were pretty sure I'd succeeded in getting a human into the box. To them, it was already AGI. I'd gotten a computer to do something only a human being (their employee) had theretofore been able to do. And I got the computer to do it BETTER and FASTER. Of course, these were mostly non-technical people who didn't understand the technology (in many cases had never even heard of it) and so, to them, there was a bit of magic involved. We, here, of course know that was not the case. While the stuff I built back in the 1980's and 1990's may not have been snazzy, wiz-bang AI with conversational robots and the whole Sci-Fi thing, it was still damn impressive and extremely human-beneficial. No NLU. No embodiment. I don't claim to have a better way to get to AGI, just a less risky way. Based on past experience (in the field). I have never intended to criticize any particular AGI approach. I have not tried to show that my approach is conceptually superior to any other approach on any specific design point. Indeed, I firmly believe that a multitude of vastly different approaches to this problem is a good thing. At least initially. As far as OCP's approach to embodiment is concerned, again it's neither the specifics nor the novelty of any particular approach that concerns me. The efficacy of any approach to the embodiment problem can only be determined once it has been tried. I'm only pointing out something everybody here knows full well: embodiment in various forms has, so far, failed to provide any real help in cracking the NLU problem. Might it in the future? Sure. But the key word there is might. When you go to the track to bet on a horse, do you look for the nag that's come in last or nearly last in every previous race that season and say to yourself, Hey, I have a novel betting strategy and, regardless what history shows (and the odds-makers say), I think I can make a killing here by betting the farm on that consistent loser! Probably not. Why? Because past performance, while not a guarantee of future performance, is really the only thing you have to go on, isn't it? Cheers, Brad P.S. Back in the early 1970's I once paid for a weekend of debauchery in Chicago from the proceeds of my $10 bet on a 20-to-1 horse at Arlington Park race track because I liked the name, She's a Dazzler. So it can happen. The only question is: How much do you want to bet? ;-) David Hart wrote: Brad, Your post describes your position *very* well, thanks. But, it does not describe *how* or *why* your AI system might
Re: [agi] It is more important how AGI works than what it can do.
Hi Brad, An interesting point of conceptual agreement between OCP and Texai designs is that very specifically engineered bootstrapping processes are necessary to push into AGI territory. Attempting to summarize using my limited knowledge, Texai hopes to achieve that boostrapping via reasoning over commonsense knowledge which has been acquired via a combination of expert-system data entry and unsupervised learning. OCP hopes to achieve that boostrapping via a combination of embodied interactive learning and reasoning supplemented with narrow-AI NL components (wordnet, RelEx semantic comprehension, RelEx NLgen, etc.). Of course, each project has their own reasons for believing that their approach is the most tractable and the least likely to become stuck in the AI-rabbitholes of the past. I believe that surface comparisons of most modern AGI-oriented designs cannot be used to make 'likelihood to proceed faster than others' predictions with sufficient confidence to weave convincing arguments over an email medium. So, making assertions about a design being 'better, faster, cheaper, less risky, etc.' are okay, if those assertions are clearly opinions (being backed up in writing is good, but that generally requires paper and book length treatment) and agreements to disagree are arrived at readily (without resorting to digressions about straw men to undermine others positions). The goal of this structure for this aspect of list discussion is to create an atmosphere where everyone can learn as much as possible about competing AGI designs. I think we're all saying effectively the same thing here, so we should be able to agree to agree on this point. IMO, it's more productive to highlight the reasons why your [insert AGI design here] system might work, rather than obsessing on the flaws of other designs. E.g, it's really not useful to repeatedly press the fact that past [grossly insufficient] attempts at NLU and embodiment have been abject failures, since *ALL* past attempts at AGI have fallen short of the mark, including knowledge-based expert-system with reasoning-bolted-on approaches. Furthermore, if all of science and engineering used the conservative logic that past performance [...] is really the only thing you have to go on, then we'd still be stuck with Victorian-level science and technology, since all of the great leaps where past performance WASN'T the best indicator would have been missed. On to a positive argument for the OCP design, the simple explanation for why embodiment in various forms has, so far, failed to provide any real help in cracking the NLU problem, is that all past attempts at embodiment have been incredibly crude and grossly insufficient. The technologies that might allow for fine realtime motor control and perception (including proprioception, or even hacks like good inverse kinematics, and other subtleties) in real or virtual settings have simply not yet been sufficiently developed. Any roboticist or virtual world programmer can confirm this assertion. One aspect of OCP development focuses on this issue and is working with the realXtend developers to enhance OpenSim to provide sufficient functionality to enable ever more sophisticated perception-action-reasoning loops (we'd also like to work with robot simulation and control software at some later stage); this work will likely be written up in a paper sometime next year. -dave On Sat, Oct 11, 2008 at 9:52 PM, Brad Paulsen [EMAIL PROTECTED]wrote: Dave, Well, I thought I'd described how pretty well. Even why. See my recent conversation with Dr. Heger on this list. I'll be happy to answer specific questions based on those explanations but I'm not going to repeat them here. Simply haven't got the time. Although I have not been asked to do so, I do feel I need to provide an ex post facto disclaimer. Here goes: I am aware of the approach being taken by Stephen Reed in the Texai project. I am currently associated with that project as a volunteer. What I have said previously in this regard) is, however, my own interpretation and opinion insofar as what I have said concerned tactics or strategies that may be similar to those being implemented in the Texai project. I'm pretty sure my interpretations and opinions are highly compatible with Steve's views even though they may not agree in every detail. My comments should NOT, however, be taken as an official representation of the Texai project's tactics, strategies or goals. End disclaimer. I was asked by Dr. Heger to go into some of the specifics of the strategy I had in mind. I honored his request and wrote quite extensively (for a list posting -- sorry 'bout that) about that strategy. I have not argued, nor do I intend to argue, that I have an approach to AGI that is better, faster or more economical than approach X. Instead, I have simply pointed out that NLU and embodiment problems have proven themselves to be extremely difficult (indeed,
Re: [agi] It is more important how AGI works than what it can do.
Dr. Matthias Heger wrote: Brad Pausen wrote The question I'm raising in this thread is more one of priorities and allocation of scarce resources. Engineers and scientists comprise only about 1% of the world's population. Is human-level NLU worth the resources it has consumed, and will continue to consume, in the pre-AGI-1.0 stage? Even if we eventually succeed, would it be worth the enormous cost? Wouldn't it be wiser to go with the strengths of both humans and computers during this (or any other) stage of AGI development? I think it is not so important what abilities our first AGIs will have. Human language would be a nice feature but it is not necessary. Agreed. And nothing in the above quote indicates otherwise. I'm only arguing that we should not spend scarce resources now (or ever, really) trying to implement unnecessary features. Both human-level NLU and human-like embodiment are, in my considered opinion, unnecessary for AGI 1.0. It is more important how it works. We want to develop an intelligent software which has the potential to solve very different problems in different domains. This is the main idea of AGI. Imagine someone thinks he has build an AGI. How can he convince the community that it is in fact AGI and not AI? If he shows some applications where his AGI works then this is an indication for the G in his AGI but it is no proof at all. I agree. Even a turing test would be no good test because given n questions for the AGI I can never be sure whether it can pass the test for further n questions. Ah, I see you've met my friend Mr. David Hume. AGI is inherently a white box problem not a black box problem. A chess playing computer is for many people a stunning machine. But we know HOW it works and only(!) because we know the HOW we can evaluate the potential of this approach for general AGI. Brad, for this reason I think your question about whether the first AGI should have the ability for human language or not is not so important. If you can create a software which has the ability to solve very different problems in very different domains than you have solved the main problem of AGI. Actually, I disagree with you here. There is really no need to create a single AGI that can solve problems in multiple domains. Most humans can't do that. We can, more easily, I believe, coordinate a network of AGI agents that are, each, experts in a single domain. These experts would be trained by human experts (as well as be able to learn from experience) and would be able to exchange information across domains as needed (need being determined, perhaps, by an expert supervisor AGI agent). None of these agents, alone, would qualify as AGI (because they are narrow-domain experts). The system in which these AGI experts function would, however, constitute true AGI. Your reply makes it sound like I have a question about whether the first AGI should have human language ability. I have no question about this. What I have is an informed opinion. It is this: requiring solution of an AI-complete problem (human-level NLU) is the kiss of death for the success of any AGI concept. If we let go of this strategy and concentrate on making non-human-like intelligences (using already-proven AI strategies that do not rely on NLU or embodiment and that leverage the strengths of the only non-human intelligence we have at present), I believe we will get to much more powerful AGI much sooner. My concept of AGI holds that creating many different domain experts using proven, narrow-AI technology and, then, coordinating a vast network of these domain experts to identify/solve complex, cross-domain problems (in real-time and concurrently if necessary) will, in fact, result in a system that has a problem-solving capability greater than any single human being. Without requiring human-level NLU or embodiment. It will be more robust (massive redundancy, such as that found in biologically-evolved systems is the key here) than any human being, be quicker than any human being and be more accurate than any human being (or, especially, organization of human beings -- have you ever tried to get an error in you HMO medical records corrected?). For example, in the (near, I hope) future when you feel sick, you will sit down at your computer and call up a medical practitioner (GP) AGI agent (it doesn't really matter from where, but assume from the Internet). This will be the same GP AGI agent anyone else anywhere in the world would call up (except, of course, each human is invoking a, localized, instance of the GP AGI agent). Note that you're ahead of the game already. You didn't have to wait two weeks to get an appointment (at 7AM in the morning). You didn't have to go to a remote location (the doctor's office or clinic). The visit to your doctor is already much less stressful, a medical benefit in and of itself. Once the GP AGI responds, you will only have
Re: [agi] It is more important how AGI works than what it can do.
Brad, Your post describes your position *very* well, thanks. But, it does not describe *how* or *why* your AI system might achieve domain expertise any faster/better/cheaper than other narrow-AI systems (NLU capable, embodied, or otherwise) on its way to achieving networked-AGI. The list would certainly benefit from any such exposition! On a smaller point of clarification, the OCP 'embodied' design will not attempt to simulate deep human behavior, but rather kluge good enough humanesque and non-humanesque embodiment to provide *grounding* for good enough solutions in a wide variety of situations (sub-adult performance in some situaitons and better-than-genius performance in others) including NLU, types of science that require massive information synthesis and creative leaps in thinking (inlcuding in non-everday-human contexts such as nanoscopic quantum scales or macroscopic relativistic scales), plus other interesting areas such as industry, economics, public policy, arts, etc. -dave --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
