Steve, Not all of us suffer from such a "malady". The text-based AI I am in the process of building now is designed to learn through *interactive conversation*, not by merely passively reading large amounts of error-ridden text. It also does not make the assumption that anything it learns is ever finally, completely correct and accurate. I agree with your assessment that models are required, but I disagree with you that natural language cannot be used to build a model. In fact, I believe that *constructing, updating, and querying a person's internal model is precisely the function of language*. (Can you think of another reason for it to have evolved in the first place?) People give each other instructions all the time via language, and they get around the problems you describe by iteratively refining the model based on interactive correction, which is a mechanism I am building into my system. Lacking the ability to interact with the real world is a legitimate shortcoming, but once I have the system put together, I will have a framework for integrating evidence generated from real-world experience into the model via very similar mechanisms to those used for integrating evidence generated from conversation.
In this email, I provide a very rough, high-level sketch of a system I am building, focusing on why I think my system is an exception to your criticisms. I am certain it is full of the rounding errors and overly general statements you describe. But were you to take a sufficient interest, and were I to find it desirable to convey the design to you at a given level of detail, we could take turns between explanations and clarifying questions until you were able to explain the design back to me in your own words in a way that demonstrates to me that the model you have constructed in your own head is isomorphic, at the targeted level of abstraction, to the one I hold in my own head, at which point I can be satisfied that you understand fully what it is that I am saying. Likewise, if you had criticisms that applied to that model, you could then communicate them to me via the same mechanisms. This is a process that is commonly followed by educators and tutors around the world; it's nothing new, and there is no reason a machine could not follow the same process, because it basically consists of cumulative trial and error with a stopping condition. Jim's understanding is sound. Where I disagree with him is in the need for a new language. My system already takes English sentences and agglomerates their meanings into a complex model. It already uses English semantics to match referents within that model, and it can take a wrongly understood sentence and re-evaluate its interpretation. I am still working on the conversational back-and-forth, but it is primarily a matter of conversational mechanics; the ability to update and revise the model based on new, potentially contradictory or more precise inputs is already there. There is still an enormous amount of work to be done, and the system is far from demo ready, but I already have the proof of concept in hand. The system can already describe things back to the user in its own words, and answer basic questions about what it has been told, despite starting with zero knowledge of anything except English grammar and the meanings of so-called "closed" word classes (e.g. the prepositions, to which new words are very rarely added). It is my belief that the reason children pick up on language so easily despite the so-called poverty of the stimulus <https://en.wikipedia.org/wiki/Poverty_of_the_stimulus> problem is that they already have internal models for the environment to which language is being mapped, drastically reducing the dimensions of the problem space. If I am correct, then once the modeling framework is complete and integration of experiential evidence is functioning effectively - the experimentation which you talk about in a later email than the one I am replying to - it will be possible to eliminate even this predigested bootstrapping knowledge of grammar and closed word classes, by pointing the system's experiential evidence integration routines at the problem of language understanding itself. On Sun, Nov 8, 2015 at 4:16 PM, Steve Richfield <[email protected]> wrote: > Jim, > > You seem to be suffering from a malady that many others on this forum > appear to be suffering from. The primary symptom is the belief that simple > observation of HIGHLY noise-ridden text can lead machines to useful > epiphanies. > > The problem is that pretty much everything people write is, in a word, > wrong. Overbroad statements (and natural languages can NOT precisely > bracket statements), statements made based on unstated presumptions (and no > one can list all their presumptions), statements made based on faulty > models (and ALL models are faulty, as every physicist knows), confused > meanings or words (most words have multiple meanings), lies made for > economic or other gain, etc., etc., etc. It is difficult to find ANY clear > statements that are unquestionably accurate in all of their potential > meanings. > > OK, so suppose you accept the above and simply want to do the best you > can. Still, you can NOT approach the capabilities of an average person, > because we have a real world in which to test the many possible > interpretations of what we read, whereas a machine can only accept, reject, > or assign a probability with NO other information. > > I used to believe as you do - until 2001 when I became very sick. It took > me 3 months of every conscious minute to figure out what was wrong and what > might be done about it. Sure I got most of my information from books or the > Internet, but there remained several very different models in which this > all fit, which I resolved with phone calls to key researchers, and a little > of my own primary experimentation to sort the flyshit from the pepper. > Still, I remained unsure until what should work did work to cure my > condition. > > When I went back to figure out how the Internet might be reorganized to > reduce my search from 3 months to a few minutes, the structure of DrEliza > emerged, and later my patent. > > The problem is that to usefully solve problems you need MODELS, yet > natural language (and most human thinking) predates this concept and only > provides INFORMATION. Sometimes you can construct a model from information, > but this is RARE. Making models requires highly qualified geniuses who can > get their arms around entire fields and synthesize models that fit ALL > known observations. I have done this in several narrow areas, but it will > take a LOT more of this to transform the Internet to a model-based system > from which an AI/AGI can usefully address problems of all types. > > Further, natural language is highly granular - there are far more things > that you can NOT say than things you CAN say, so people without even > thinking about it round to the nearest syntactically expressible meaning in > EVERYTHING they say or write. This "rounding" completely destroys any > ability to construct accurate models. > > Some languages have weak workarounds to this, e.g. German with its > concatenated words, or Arabic where they alter spellings for emphasis, but > these measures only slightly reduce their granularity. > > All in all, each person here must either find a way to jump from > information to models, or abandon this quest. Sure, information can help > people solve a problems whose solution has already been stated, but there > are already plenty of experts around who do this quite well. It is the > UNsolved problems that are interesting, and it is these UNsolved problems > that can NOT EVER be solved by automated means based on people's writings. > > How can I say not EVER when writings continue to accumulate? Because > society's problems also continue to accumulate, so as people find solutions > to past unsolved problems, even more new problems emerge to replace them. > > I hereby proclaim your apparent quest to be theoretically unachievable for > the many reasons outlined above. Sure it would be of astronomical value, > like the methodology to change lead into gold that so many people put so > much effort into, but why waste your time unless/until you can find SOME > way around the above-listed barriers. > > Steve > ===================== > > On Sun, Nov 8, 2015 at 9:00 AM, Jim Bromer <[email protected]> wrote: > >> Steve, >> It will take me some time to reply carefully so let me respond to >> something I feel strongly about. >> >> >>And because it is not an all encompassing >> language of communication it could be used to test the 'emergence' of >> insight that could arise if enough preparatory work had been done, >> even if I haven't figured out how that could be done without the >> artificial referent language. >> >> >> >> >There is a VAST chasm between being able to define language >> constructions and meanings, and "insight". >> > >> >> I believe there is a vast chasm between 'simple associations' or 'simple >> correlations' or associations derived from 'neural networks' and conceptual >> integration. Sophisticated artificial conceptual integration would make >> 'insight' feasible and simple examples across a wide range of subject >> matter should arise fairly quickly. But since AI programs are only capable >> of the simplest examples of 'insight' then declarations about the chasm >> between AI and 'insight' are expected. So I totally disagree with you about >> this. I feel that your feelings about this are historically accurate but >> have little to do with the potential near-future. As I say, I do not recall >> hearing about an AI program that is capable of learning via conversation >> except for extremely simple domains. I feel that I have a solution for this >> problem but the trial and error process of getting from where I am now and >> where I think I can get is so overwhelming a challenge that my decision to >> use the artificial referent para-language makes a sense. >> >> Jim Bromer >> >> On Sun, Nov 8, 2015 at 11:19 AM, Steve Richfield < >> [email protected]> wrote: >> >>> Jim, >>> >>> FINALLY - SOMEONE who wants to discuss PRACTICAL implementations of >>> TAI/TAGI. >>> >>> Continuing... >>> On Sun, Nov 8, 2015 at 7:14 AM, Jim Bromer <[email protected]> wrote: >>> >>>> After I wrote that message I realized that I had tried to start >>>> discussions about an artificial language that could be used to shape a >>>> general AI program before. Many of these discussions were side tracked >>>> when people started talking about Esperanto or about lambda calculus >>>> based artificial languages and stuff like that. That is not what I am >>>> thinking of. >>>> >>> >>> You mean, having syntax like: >>> >>> *When that I write "xxxx" I mean "yyy".* >>> >>> to define idioms, for more subtle things like: >>> >>> *Consider that when I write "," I may mean ";".* >>> >>> which expresses potential alternative interpretations of future >>> writings? >>> >>>> >>>> The artificial language could be used with video or audio or other >>>> kinds of IO environments, but I would use it along side of an attempt >>>> to get the AI program to learn to use a natural language. >>> >>> >>> I did a VERY similar thing in a FORTRAN/ALGOL/BASIC compiler I once >>> wrote for Remote Time-Sharing Corp. It started out as a very simplistic >>> metacompiler, to which I fed it a description of a more capable >>> metacompiler, in which language I fed it a description of an optimizing >>> metacompiler. >>> >>> This could easily be done in a language like English, where a >>> rule-driven system like I have been discussing here has rules whose >>> function is to introduce new rules. >>> >>> >>>> One of the >>>> dreams of old AI was that if you started instructing the program to >>>> learn using the artificialities of some kind of language it would >>>> eventually have enough information for genuine learning to emerge. >>>> >>> >>> The think that seems to be the killer here is erroneous learning of >>> various sorts. Superstitious learning is theoretical unavoidable. Once you >>> get something erroneous into such a system, it becomes difficult/impossible >>> to get it out. A VERY simple demonstration comes in trying to use Dragon >>> NaturallySpeaking's speech input to correct its errors in your dictation. >>> As you would expect it makes errors in trying to correct the errors, and >>> this often compounds to overwhelm any hope of setting things right. >>> >>> Add to that not knowing exactly what a computer got wrong, or even being >>> able to recognize that the computer got something wrong, and you can see >>> how difficult/impossible it is to correct wrongly "learned" rules. >>> >>> >>>> This never really worked. Why not? Partly because computers were not >>>> powerful enough in the old days >>> >>> >>> And still aren't - unless you use my patented LFU methodology. >>> >>> >>>> and, in my opinion, AI researchers had >>>> not appreciated the necessity of sophisticated data integration >>>> methods for some reason. (Old computer systems might one day be shown >>>> to have been potentially powerful enough to run some future program >>>> but they were not powerful enough to entertain the trial and error >>>> process that would have been required using experimental programs of >>>> the day. >>> >>> >>> The advantage in LFU is about the same as the advantage of a modern PC >>> over an old vacuum tube clunker, so yes, they could have done a LOT more >>> way back then. >>> >>> The "cycle time" of an IBM-709 computer was 12 microseconds, and most >>> instructions took two cycles - one to access and interpret the instruction, >>> and one to access and operate on the operand. >>> >>> >>>> For example, with better conceptual integration methods a >>>> future efficient AI program might be used on an old computer system >>>> just to show that it could be run on it.) >>>> >>> >>> No, except for a few in the Computer Museum's display in Cupertino they >>> have all been melted down for their scrap metal, and the Museum won't turn >>> them back on. >>> >>>> >>>> So the artificial referent language would not be a complete language >>>> (of communication) like Esperanto wants to be. And it would not be a >>>> logically sound language like lambda calculus wants to be. It could be >>>> used to establish referents from the IO data environment. It would >>>> need to be capable of denoting a distinction between how those data >>>> objects can be used. For example in natural language there is an >>>> important distinction between syntax and semantics. So if I used this >>>> referent language with a natural language IO then one of the >>>> artificialities would be to distinguish syntactic relations from >>>> semantic relations. On the other hand, this distinction is not always >>>> necessary, desired or clear cut. To explain this, many (or maybe most) >>>> (what I think are) desirable syntactic relations are based on some >>>> semantic conditions. But then again there is no reason not to design >>>> the artificial language to be able to represent relations that are >>>> mixes of semantics and syntax. >>>> >>> >>> Leaving a stupid computer to untangle such messes is probably a mistake. >>> However, it would be fairly easy to provide a mechanism for people to >>> specify such things. >>> >>>> >>>> As I see it, the main problem with language based AI has been the lack >>>> of a really good conceptual integration solution. >>>> >>> >>> This broad statement could be said about ANYTHING people haven't yet >>> seen a way to make work - like AGI. >>> >>>> >>>> One of the reasons I write to groups like this is that I want to get >>>> some ideas about how an idea might work. >>> >>> >>> Same here. >>> >>> >>>> But when I wrote about an >>>> artificial para-language before I wasn't really sure it I even wanted >>>> to use it. I finally have come to the conclusion that it makes a lot >>>> of sense. I can use it to speed up tests about my AI/AGI theories but >>>> then I could also test those theories with more relaxed instructions. >>>> So the artificial para-referent language would not a all encompassing >>>> language of communication, it would not be a logically sound language >>>> other than to denote semantic and syntactic references and relations >>>> based on mixes of semantic and syntactic references. It could also >>>> denote relations that I think would be important to a text-based >>>> AI/AGI program. Because the logic of the method would not be tight and >>>> a contradicting case would not (always) lead to an artificially >>>> reported error, the AI methods would have to do some learning for >>>> itself. So the para-referent language would not sidetrack the whole >>>> effort because if the AI methods have to have the potential to exhibit >>>> some genuine learning. And because it is not an all encompassing >>>> language of communication it could be used to test the 'emergence' of >>>> insight that could arise if enough preparatory work had been done, >>>> even if I haven't figured out how that could be done without the >>>> artificial referent language. The benefit is that I could use it to >>>> test and develop my AI theories. I am really excited by this idea this >>>> time. >>>> >>> >>> There is a VAST chasm between being able to define language >>> constructions and meanings, and "insight". >>> >>> *Steve* >>> ====================== >>> >>>> Jim Bromer >>>> >>>> >>>> On Sat, Nov 7, 2015 at 10:22 PM, Jim Bromer <[email protected]> >>>> wrote: >>>> > I was just working on my latest p=np? idea and I hit up against method >>>> > that is either in np or is otherwise extremely inefficient. So I have >>>> > to come to the conclusion that the human mind is not capable of SAT in >>>> > p. >>>> > >>>> > So then how do we figure how to deal with so many complicated >>>> > situations? Of course I still don't know because so many situations >>>> > seem similar to a SAT problem. The mind must be able to detect many >>>> > different things that are going on at once or which might be useful to >>>> > recall from memory to deal with a situation. But still, there is >>>> > nothing in my own introspective analysis of my thinking which looks >>>> > anything like a p=np process. >>>> > >>>> > So what is wrong with AI? One thing that AI has been consistently >>>> > lacking is the ability to learn through conversation. My feeling is >>>> > that this is not just a problem with communication but a learning >>>> > problem as well. In other words AI is not able to truly learn except >>>> > in a few special cases. Most of those special cases are examples of >>>> > narrow AI but there are others where the learning that takes place >>>> > isn't necessarily like other narrow AI but where the domain of >>>> > learning is so restricted that it is narrow in the sense that the >>>> > applicability of the method is limited. >>>> > >>>> > Then I started thinking of an artificial language which can refer to >>>> > situations or objects in the IO data environment and which can be used >>>> > to instruct a program as it is running. I think this is an unusual >>>> > idea. >>>> > >>>> > One of the characteristics about programming methods that seem to >>>> > catch on with programmers is that they can be used in a very simple >>>> > manner and in more complicated programming. I think an artificial >>>> > language which could be used to instruct a computer to notice objects >>>> > in the IO data environment and which could also be used to refine >>>> > those instructions using this artificial language with the references >>>> > that it had previously established has a lot of potential. And it can >>>> > help us become more clear about what is needed to make better AGI >>>> > programs. >>>> > Jim Bromer >>>> >>>> >>>> ------------------------------------------- >>>> AGI >>>> Archives: https://www.listbox.com/member/archive/303/=now >>>> RSS Feed: >>>> https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac >>>> Modify Your Subscription: https://www.listbox.com/member/?&; >>>> Powered by Listbox: http://www.listbox.com >>>> >>> >>> >>> >>> -- >>> Full employment can be had with the stoke of a pen. Simply institute a >>> six hour workday. That will easily create enough new jobs to bring back >>> full employment. >>> >>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>> <https://www.listbox.com/member/archive/rss/303/24379807-653794b5> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >> Modify <https://www.listbox.com/member/?&;> Your Subscription >> <http://www.listbox.com> >> > > > > -- > Full employment can be had with the stoke of a pen. Simply institute a six > hour workday. That will easily create enough new jobs to bring back full > employment. > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/23050605-2da819ff> | > Modify > <https://www.listbox.com/member/?&;> > Your Subscription <http://www.listbox.com> > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
