[moved from the SL4 list] Ben Goertzel: BG> Cliff,
BG> Regarding your idea for an AGI application... BG> I think it's an interesting one, but a very hard one, suitable only for BG> AGI's at the near-human stage of general intelligence. Understanding the BG> pragmatics of what humans are trying to do at their PC's, is a tough BG> problem, requiring sophisticated modeling of human psychology.... First, I'm not too sure that's the case, at least given a limited domain such as word processing, browsing or email. Aside from modeling what the user does, I think modeling what the computer does and communicating that to the user more meaningfully than in hardcoded (and often inaccurate or misleading) dialog boxes would have value. In any case, the idea is for the user model and the "understanding" of the user's communication to evolve based in part on user direction, not just through observation and (confirmed) intervention (which would probably be harder). In other words, in the case of a word processing application the package would come with the "basic concepts" of document commands (creating, opening, editing saving), editing actions (typing, inserting, deleting, moving), formatting actions, and document contents (chapters, TOCs, paragraphs, words). (These concepts having some sort of "referents", i.e. not being hollow phrases.) Where a current user would need to write a "macro", a user of this package would instead communicate with the package about the word processing program in terms of those abstractions. "Copy this list of bullets and make each one a new chapter heading" doesn't require a sophisticated user model, nor that sophisticated a model of the word processing software... Where I see the proto-AGI elements coming in is in allowing the language the user uses to instruct the computer to evolve, and in interpreting that language, which requires some sophistication in applying analogies. An example conversation: "What I just did is "expanding" a list of bullets. Now expand this list." (gestures with mouse) "Now expand this paragraph." (I envision a CopyCat-like exploration of possible ways to apply the "expand" concept to a paragraph arriving at the ~90% strong conclusion that the sentences of a paragraph most closely correspond to each bullet in a bulleted list.) " "Collapse" is the opposite of "expand". Now collapse chapters 3-7." The above to me seems reasonably doable; more valuable and somewhat more complex would be: "I want to do something like expanding this (gestures) list, but instead of chapters, I want paragraphs starting with each bullet point italicized, the rest of the paragraph normal, with a marker for me to fill out later. (done) In fact, "expand" is more general than I first said; it can mean expanding into chapters, into paragraphs like just now, and maybe other ways. If I don't specify, assume paragraphs." [later] "Expand this list." (expands into paragraphs) "Whoops, I meant into chapters." (corrects last action) The above examples imply a high degree of personal customization, that is, the software over time builds a language pretty exclusive to a particular user's style of use and communication. This can have drawbacks as well. The other thing implied by the above is some natural language processing, but need not be as sophisticated as it may appear above...alternately, a limited-vocabulary software-specific language might be feasible. BG> There's only one way I can envision your product being doable in the short BG> term, and that's if it were installed on a HUGE number of computers all at BG> once, so that there was a vast amount of statistical data from which it BG> could make inferences about what a given user might be trying to do. In BG> other words, maybe it could work in the short term if M$ were to integrate BG> it into Windows ;-) Or if it were somehow applicable to browsing or email. BG> I think that there's a limited subset of your proposal that is more feasible BG> in the very short term though, and that's AI-driven system administration. BG> Stephanie Forrest and her students at UNM have done some interesting work on BG> "computer immune systems", which are simple statistical narrow-AI-ish BG> systems for intrusion detection. But there's a huge amount more that can be BG> done along these lines. The initial market would be companies owning server BG> farms, which are very complex to administer, rather than individual PC BG> owners.... Computer/server security is definitely an area where ongoing, realtime detection is at least as valuable as the "simple" things (firewalls, virus detection), and where AI could have value. -- Cliff >> -----Original Message----- >> From: [EMAIL PROTECTED] [mailto:owner-sl4@;sl4.org]On Behalf Of Cliff >> Stabbert >> Sent: Friday, November 08, 2002 11:53 PM >> To: [EMAIL PROTECTED] >> Subject: re: early AGI apps >> >> >> I wrote: >> CS> I think the Novamente approach -- more narrowly focused commercial >> CS> efforts -- are a very good approach to funding currently. I do have >> CS> a long-simmering but vague-in-details idea for some AI tech that >> CS> Novamente at its current stage may or may not be suited to, which >> CS> if implemented as a software package could be quite popular. >> >> Ben Goertzel: >> BG> Well, feel free to voice the details, either on the list or >> via private >> BG> e-mail, [EMAIL PROTECTED] ;) >> >> Peter Voss: >> > I'm interested in any and all potential early applications for >> AGI - both to >> > evaluate the performance of our a2i2 system, and for possible >> > implementation. >> >> I don't have much time at the moment to sketch this in greater detail, >> and my own limited notes on this are at the time unavailable. So what >> follows is *very* vague, brief and blue-sky...but it may spark some >> ideas. >> >> The main thrust is the use of AI in GUIs. There are two main aspects >> to that -- modeling the computer and modeling the user -- as well as >> some other features. As a product, I envision something that sits "on >> top" of the OS, or encapsulates it, and is always the "outermost" >> control the user can speak to -- the ultimate arbiter, the >> presidential hotline, and where the user communicates *about* what the >> computer does. (Another, somewhat less ambitious, implementation >> would be specific to and sit on top of a software package such as >> Microsoft Office). >> >> ------------- >> >> I'm sure all of us have run into problems like the following: I was >> installing a piece of software under Windows. I have my PC set up >> with all apps on E:, so I chose E: as the install drive. It wouldn't >> let me proceed because, it said, the E: drive was full. I checked >> under Explorer -- and as I thought, I had plenty of space. >> >> So what was going on? I'm not certain -- it may have been that the >> installation program was looking at total minus used space and going >> by the 2GB value for total space (this was under Win95); it may have >> been using variables that were too small and getting overflow...it may >> have been any number of glitches. >> >> But it got me thinking -- we have two different programs (the >> installer and Explorer) telling me two different things. We have an >> "out of space" dialog box that some programmer went to the trouble of >> building, and the out-of-space check they coded for it... >> >> When you work on a modern GUI, it's really in a sense like you're >> having a dialogue. You tell it to do things, and it comes back with >> "dialog boxes" ("disk too full", etc.). >> >> The problem is that you are having a dialogue with a schizophrenic >> amnesiac, and what is more, one who communicates in a very inflexible, >> rigid, repetitive way. >> >> As a user, I want a single, coherent, consistent conversation that >> builds -- from question to question and day to day. A conversation >> that makes sense. >> >> ------------- >> >> In essence what I want as a user is that the computer "understands" >> what it is telling me. Well, what does that mean? That it must have >> some internal representation of itself, that symbols such as "hard >> drive" and "free space" must be not just strings of characters, but >> have some semantic *value* -- some *referent*. >> >> I.e., the software should have an internal model of the computer in >> which terms such as the above are meaningfully linked to actual values >> from that model. >> >> Such a model, with some associated pseudo-natural-language processing, >> may be the "simpler" aspect of the software -- perhaps amenable to >> established technologies (DBs, natural query languages, expert >> systems, etc.). >> >> ------------- >> >> When I tell the computer to save my file and the disk *is* full, it >> should come up with a more meaningful set of choices - it should have >> some basic understanding of why I want to save a file and e.g. offer >> me - do you want to clear some space or save to D: instead? >> >> Now, I have seen the above -- a dialog box asking me whether I want to >> clear space -- but it, again, was something some programmer separately >> designed and built code for. >> >> As a user, I want the computer to "know" that I want to save files, >> that if space is unavailable in one place I may want to save it >> elsewhere, or free space up. >> >> More fundamentally, I want to be able to teach it such things. >> >> ------------- >> >> This implies the software needs to have an internal evolvable model of >> me, the user, as well. >> >> In the context of Microsoft Office, it would have some ideas of "what >> I do" (create new documents and edit existing ones, type and format >> stuff, print them, save them, file them, search them). It should >> learn and suggest things on its own (create "wizards" on the fly -- >> I type "Dear George" and it pops up with "another letter to George, >> 'ey? Usual headings? File under 'Presidential correspondence'?") >> >> The model should evolve both under internal (user-confirmed) guesses >> and under user direction. The software should always have available >> -- no matter what the user is doing -- a way for the user to interrupt >> and tell it to "go meta". By this I mean a way for the user to tell >> the software "Now watch what I'm doing...whenever I do X, I want you >> to do Y and Z." or "See that dialog that popped up? Always click "No" >> on that one." >> >> Here's where the model gets more complex, because it needs to deal >> with analogies. "See how I'm taking each of the sentences from this >> paragraph and making them bullet points? Do it for the rest of the >> paragraph." is relatively simple. "Do the same thing to this line >> chart." is more complex (what is "the same thing"? Does each series >> (line) get split into its own chart? Each year?). This is also the >> point in the model where an ongoing dialogue becomes most important -- >> being able to correct the assumptions and guesses the software makes, >> such corrections getting folded back into the model. >> >> ------------- >> >> The language used to describe the internal structures of the computer >> and its software, actions associated with them, etc. should allow for >> sharing between users. In other words, once one user has "told" the >> system about what PhotoShop does (via interactive querying by the >> software, details too long to go into), others wouldn't need to. >> >> ------------- >> >> So, vague and scatter-shot. There's a lot left out of the above and >> when I can dig up all my notes and have the time, I might put together a >> more coherent and detailed presentation of this; the AGI list is >> probably a more suitable venue. >> >> The reason I see this as a viable application for (early) AGI is >> because on the one hand it requires limited, quantifiable, shareable >> knowledge (domain expertise about software combined with the ability >> to intercept both user events and OS API calls) and on the other, >> provides an "evolution-driving" environment: the user's requirements. >> >> There is IMO a widespread need for such tech in user interfaces, thus >> plenty of people who would want to use it. And possibly, if this or >> some variant on it was structured in the right way (perhaps dealing >> with the browsing experience, say), a huge number of users could be >> leveraged to drastically accelerate the evolution of the software. >> >> >> -- >> Cliff >> ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/
