>> Obviously, I don't agree with this phrasing, but if you replace "a miracle >> happens here" with "there is research to be done here, on the level of a >> good PhD thesis" then I agree with you.
Hmm. Looks like my joke Gantt chart doesn't have the wide currency here that I thought that it had. I assume that my previous explanation makes it clear that we're in agreement. >> Second, over the past few years, I've become more and more convinced that >> discovery systems, while they do "learn", are not the type of learning that >> I think is necessary for AGI. > Here we have a fundamental philosophical disagreement. But neither of us can prove the other wrong and there are numerous people in both camps . . . . >> There seems to be an issue of definitional disagreement here, as the PLN >> inference component within Novamente specifically does handle reasoning by >> analogy. >> I'm not sure why you feel analogical inference is particularly difficult, as >> opposed to e.g. deductive or abductive inference. Analogical inference *once you have found the analogy* is handled by PLN and is not "particularly difficult, as opposed to e.g. deductive or abductive inference." Finding the analogies is a real trick, is heavily based on existing knowledge, and really *is* in the "a miracle happens here" realm. PLN inference just resolves the problem after it is set up and *in this instance* is subject to many of the complaints that we all commonly level against narrow AI. >> scale-invariance of knowledge, ways of determining and exploiting >> encapsulation and modularity of knowledge without killing useful "leaky" >> abstractions, and "memory" design. (oh my!) > When we drilled down on these issues before, Mark, they seemed to me to come > down to matters of taste regarding software implementation choices, rather > than fundamental AGI issues. To use your phrase -- "Here we have a fundamental philosophical disagreement." The fact that you don't see these as fundamental AGI issues looks to me like a minor version of the AIXI solution (assuming infinite resources, we can . . . ). This *is* of lot the basis of why I believe Novamente is still very much in the throes of "a miracle happens here". Your low-level stuff is great -- but I don't see it as capable of scaling up (and certainly not as effectively as possible). You may well prove me wrong. You may well find quick ways to enhance Novamente so that it isn't a problem. I just believe that it is a show-stopper and currently unaddressed. THE REST I think that we're pretty much in agreement on. I certainly was not addressing you when I said "Thus, those blindly insisting that Novamente is the be-all-and-end-all and that all other approaches should be abandoned are not doing any of us a service. I want to see Novamente go forward but we shouldn't put all of our eggs in one basket." I haven't seen you say that (for others) while I do believe that *YOU* should be putting all of your eggs in the Novamente basket because that's the only way to make progress. My objection is to the "me-too" fan-boys that are stepping on other alternatives because "the problem is obviously solved except for a bit of work and some minor details". ----- Original Message ----- From: Benjamin Goertzel To: [email protected] Sent: Monday, November 12, 2007 10:54 AM Subject: Re: [agi] What best evidence for fast AI? Hi, First, Novamente is a discovery system (and a *really* good one). The other parts of it's design, however, are not fully fleshed out and there are huge "a miracle happens here" holes. Obviously, I don't agree with this phrasing, but if you replace "a miracle happens here" with "there is research to be done here, on the level of a good PhD thesis" then I agree with you. Second, over the past few years, I've become more and more convinced that discovery systems, while they do "learn", are not the type of learning that I think is necessary for AGI. Here we have a fundamental philosophical disagreement. Novamente can certainly tease out patterns from large quantities of data but it isn't fully designed (at this point) to do anything like reasoning by analogy, for example. There seems to be an issue of definitional disagreement here, as the PLN inference component within Novamente specifically does handle reasoning by analogy. I'm not sure why you feel analogical inference is particularly difficult, as opposed to e.g. deductive or abductive inference. Ben does have some plans for this but, my opinion is that, he is still in the realm of "a miracle happens here" on this subject. Actually, IMO, this is **not** one of the areas in need of most future research within the Novamente framework. It's something that is pretty straightforward within Novamente. Third, and I've said this before, there are some fundamental engineering features (scale-invariance of knowledge, ways of determining and exploiting encapsulation and modularity of knowledge without killing useful "leaky" abstractions, etc.) that aren't implemented yet in Novamente that really need to be implemented much earlier rather than later. Also, I have a lot of questions about Novamente's "memory" design. When we drilled down on these issues before, Mark, they seemed to me to come down to matters of taste regarding software implementation choices, rather than fundamental AGI issues. In particular, I think that Novamente's foray into learning in a virtual world is either going to be incredibly useful or a rather large bust because it is precisely the type of learning that Novamente hasn't specialized in before this point. Well, our initial foray involves using MOSES to do embodied procedure learning based on a combination of reinforcement signals, imitative cues and user corrections. This is actually quite similar to stuff we did before, using MOSES to learn to play fetch and tag in a 2D world, for example. What is different here, from these prior learning experiments, is that we're using MOSES in a context of an agent with a memory (an AtomTable), which does a little bit of reasoning based on this memory to guide its learning; and we have a mini version of the Novamente goals, feelings and action selection framework in there. So, while it's true that we have not done any commercial projects involving embodied learning, we've done a few academic-style prototype projects of this nature before. A number of people on this list seem to regard Ben as almost a deity or a prophet. Ben is intelligent, creative, has a solid background, and gets to work hard in the field so he looks a lot better than most everyone else. It also means that he has polished his ideas and eliminated the most obvious problems. This does not, however, mean that he has a provably correct path. Yes, I do not claim to have a provably correct path. I really think that is an overly strict criterion. For sure, to create an AGI using Novamente, there is further research to be done, not just engineering. Novamente may lead to AGI (with *a lot* more hard work). Yes, it will take a lot more hard work. My current estimate is about 6 years of full-time work for a dedicated team of 10-15 really great, appropriately trained AI engineers, who would be doing a combination of research and software engineering and teaching and testing. This estimate could be an underestimate or an overestimate. It sure ain't off by an order of magnitude though. Personally, as I've said, I believe that it is *a path* but one which will be overtaken and passed by a shorter, easier path (just as I believe that brain emulation is a path that will be overtaken and passed by a shorter, easier path). This might be the case. I just haven't seen that shorter, easier path clearly articulated anywhere. Frankly, I have spent some time looking for it, and haven't found it. But this has gotten rather long so I should sum up . . . . Novamente has great promise -- but part of the reason why it has such great promise is because so much of it *hasn't* been fully determined yet. The design is still open enough that it can be stretched to fit many things. This flexibility is intentional, and I consider it a strength of the design. We still have a lot to learn, so I took great pains to construct a design with the flexibility to be adapted based on learning done in the course of developing and teaching the system. The problem is that stretching it in some directions may/probably will make it less adept at other things (jack of all trades/master of none) Well, the human brain is arguably a jack of all trades and master of none. So I think this is okay for a first-pass AGI system. and it may well be (and this is my primary complaint) that it is *so* general that, while it could serve as the basis of an AGI, it is far more complicated than necessary to do so (just as a bird's biology is not necessary for flight). I am pretty sure that this is the case. However, I think that the simpler, more elegant design may become apparently only AFTER we have achieved a pretty powerful level of AGI with the current, more flexible design. Thus, those blindly insisting that Novamente is the be-all-and-end-all and that all other approaches should be abandoned are not doing any of us a service. I want to see Novamente go forward but we shouldn't put all of our eggs in one basket. Well Mark, even I am not suggesting that the world should put all of its AGI resources into developing Novamente and should ignore all other approaches. If I had $100M I would put it all into developing Novamente. If I had $1B I would not, I would fund a bunch of alternate approaches too ;-) As I have neither, I'm gonna get back to work now... -- Ben G ------------------------------------------------------------------------------ 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=8660244&id_secret=64175323-aef3dd
