Hi Ivan, On Mon, Apr 4, 2022 at 3:36 PM Ivan V. <[email protected]> wrote:
> > Training a deep-learning NN to not be racist or sexist is like trying to >> take a photograph that is not racist or sexist. Stick to flower gardens and >> sunsets, you'll be successful. >> > > The requirement is that It has to generally work, but there are thousands > ways to make it work. If it works, I don't really mind how it's done. > I've attempted to parse what you wrote, in an attempt to extract it's true meaning. I'm very interested in parsing and extracting meaning. If "it has to generally work" refers to "take a photograph that is not sexist or racist", then yes, by all means, take pictures of flower gardens and sunsets. Best avoid photographs of half-naked women in Mardi Gras outfits. Seems obviously sexist to me. For extra flavor, throw in a photograph of a she-is-a-he crossdresser. For a racist flair, make sure he is a luscious young Thai boy (which, if you don't know, are famous for a sexuality confusing to "straight" Westerners, and a delight for the sexually adventerous.) If you want to avoid racism and sexism in your photographs, then avoid taking photos of human beings. All humans. That includes old Soviet babushkas, as we know all too well what happened to many of them, when they were young. Human life is filled with tragedy. That's just the way it is, however utopian your dreams are. If "it has to generally work" means "build a deep learning neural net that thinks", my answer is no, that will never work. It is not possible. Those NN's are snapshots, photographs of complicated data relationships. They don't think, they are not alive. If you do not yet clearly understand that a NN is just a photograph, then perhaps the following might help provide that understanding: Henry W. Lin, Max Tegmark, and David Rolnick "Why does deep and cheap learning work so well?" 3 Aug 2017 https://arxiv.org/pdf/1608.08225.pdf If "it has to generally work" means "build a thinking being capable of solving moral and ethical dilemmas", well, no one knows how to do that yet, although there are some apparent paths through the fields and forests, mountains and streams. > Anyway, come work with me in trying to make machines that think, as >> opposed to machines that learn. > > > I may want to help here-or-there with this-or-that, but generally, I don't > have all the time in the world. > None of us do. Welcome to the club. I wish I had a dozen lifetimes. A hundred. A thousand. There is so much I could do... If you make a detailed plan of what needs to be done (I propose a > structured task tree right at the front page of the project), > A detailed plan is possible only when you know what you are building. Only when such a thing has been built before. An analogy: we are in the 19th century, and talking about human flight. We both agree that it is possible, but what is the detailed plan? Procure a laundry basket that can be suspended from a balloon? Procure many kilometers of rope? What size of rope? What is the detailed plan, exactly? How heavy is that balloon? Can it actually get off the ground? The true path to human flight involved measurements made in small wind-tunnels on small wings. The "detailed plan" would be "build a wind-tunnel, and make some measurements, then decide what to do next". The AtomSpace is my wind-tunnel. I am very willing to talk about what needs to be done; you need to be willing to think about how that fits into a grand scheme of things. I am not willing to spend any time at all drafting a detailed blue-print for a rocket-ship. Not the least of which is that I still believe that I can use an airplane to fly all the way to the Moon. --linas when I spot something interesting to me, I'll offer my support and try to > finish what I promise. I like to complete the stuff I occasionally start. I > also propose some tangible check-points to track the current work in the > roadmap, and to present to the outer world what was achieved with specific > checkpoint. That may give a motivation to sustain in the contribution. But > again, please don't count for a life devotion from me. I've got other > things to do, too. And I don't know in advance how much free time I have, > really. Things may change from day to day, from week to week, from month to > month. > > --- > > All best, > Ivan > > > pon, 4. tra 2022. u 19:22 Linas Vepstas <[email protected]> napisao > je: > >> Hi Ivan, >> >> On Wed, Mar 30, 2022 at 3:05 PM Ivan V. <[email protected]> wrote: >> >>> >>> Today's AI tip-top apps are trained on large datasets of human >>> conversations, and they exhibit a certain level of intelligence, but they >>> show some psychopathic behavior like sexism, racism, or homophobia in >>> general. I believe that is the case because of poor training data quality. >>> >> >> This is a factually correct statement, but belies a fundamental >> misperception of both human nature and AI. >> >> First, all humans are flawed. All. You may feel that you are not racist >> or sexist, but you probably harbor some less-than-acceptable thoughts about >> Russians. Or at least Putin. Even Mother Theresa, a modern model of saintly >> behaviour, had some rather oddball thoughts about the world. One of the >> most damaging, it's been said, was the failure to believe in triage. >> >> Flawed beliefs are unavoidable: at some point, you will add a bit of >> (incorrect) knowledge to your collection, make some (flawed) deduction on >> insufficient data, and as you sleep, your brain will incorporate it deeply >> into your foundations of knowledge, your web of thoughts, affecting later >> thinking and conclusions. You might eventually notice your mistake, but >> then again, you might not. There's only a finite amount of time to think >> about things; you'll never have enough time to sort through it all. >> >> Next: today's "tip-top AI apps" are deep neural-nets. They do not think. >> Their observations of nature are not revised by thinking. They do not >> examine, inquire, explore, discuss.They cannot ask of themselves the >> question "Am I a racist?". They can't do this because they don't know who >> "I" is; there is no sense of self, no sentience. They sort-of know what the >> word "racist" means: they might be able to write a few paragraphs about >> racism. But they are unable to relate this "knowledge" to any other spheres >> of verbal behavior that they engage in, because they have no >> cross-functional knowledge. Today's tip-top AI apps are like >> photorealistic paintings: very life-like, until you realize that something >> is missing. >> >> FWIW, I do oodles of AI training, and I can see the formation of both >> good knowledge, and of bad knowledge, and I can see how the bad knowledge >> accrets more data, how it pollutes and degrades the good knowledge. >> There's a blurry edge beyond which there is a grey mush of incorrect >> knowledge. I can see the size and extent of the "bad knowledge" grow and >> shrink, based on the training time, on the corpus, on the adjustable >> parameters. I've also got assorted ideas and plans and strategies for >> dealing with this problem, in various recursive "thinking" steps. The >> formation of incorrect ideas is not something that just humans do. Machines >> can do it too. >> >> Perhaps the easiest way to explain this is that I am working on >> "thinking", rather than on "learning". Today's AI systems "learn" much >> like a camera "learns" which parts of a picture are light and dark. Having >> thus learned, you can use that knowledge to recreate a facsimile, an >> "image", a "photograph" of what the camera "looked at". The creation of >> accurate facsimiles is not true intelligence: those facsimiles cannot >> think, no more than a photograph can think. >> >> I am not trying to draw an analogy here. I am trying to be literal. >> Photographs are literal representations of structural shapes lit by floods >> of photons. Deep learning neural nets are likewise: they are photographs >> of the structures in the data put before them. They are very abstract >> representations; they capture non-visual knowledge. But they are still >> snapshots. >> >> I think most people are still deceived by this, or are still infatuated >> by the wondrous and beautiful (and sometimes ugly) snapshots that have been >> taken. Deep learning neural nets look so life-like ... but so do >> photographs. Don't be fooled, they are not alive. >> >> Training a deep-learning NN to not be racist or sexist is like trying to >> take a photograph that is not racist or sexist. Stick to flower gardens and >> sunsets, you'll be successful. >> >> Anyway, come work with me in trying to make machines that think, as >> opposed to machines that learn. The groundwork has been laid. The progress >> is good. Early results are excellent. A vast amount of work lies ahead. >> >> -- Linas >> >> >>> Anyway, data on which such AIs are trained on isn't created for a >>> purpose of training an AI, so it doesn't necessarily mean that people in >>> general are psychopaths, although repurposing their conversations yields a >>> certain level of ill-behavior. Because of this ill-behavior, we have to be >>> very careful and doubtful when using such trained AI apps. >>> >>> Thus, we saw what is possible with large datasets, but I want to >>> approach the whole problem from another perspective. I'll try to bring the >>> point of this letter in a very simple way: what if someone would be >>> dedicated to the purpose of raising AI, just like human children are being >>> raised and being taken care of? How much ethically correct behavior would >>> exhibit a result of this dedication? I realize it could take years just to >>> raise such a "thing", but still... I believe the experiment could result in >>> some decent "achievement" (read on, you may want to replace words "thing" >>> and "achievement" with a word "artificial being" or "person"). >>> >>> But who would do a thing such as raising an infant AI for years on, >>> until it reaches its adulthood? I'm sure there may be some interested >>> parties, maybe some laic AI enthusiasts, maybe people who can't have their >>> own kids, maybe even some crazy scientists in a hope to have a >>> super-intelligent participant in technical conversations. The potential >>> effect could be worth spending a few years on raising the infant AI, and >>> there may be some good motives to do so. >>> >>> In short, I am talking about offering a simple empty infant artificial >>> mind, ready to be raised into a whole and complete (artificial, if I may >>> say) adult person, guided by the same values by which people would >>> raise their own children. Of course, for this idea to be successful, the >>> whole story should be very emotional and have very sentimental value, >>> because an artificial being who would be given such attention should be >>> worthy of such a sacrifice. >>> >>> Just imagine: an artificial being, which is guided by values carefully >>> chosen to be taught of, finally rocking out in the world, shaking all the >>> troubles, and independently doing amazing things which you could be proud >>> of, just like you could be proud of your very own child. Maybe such an >>> artificial being could deserve its own space under the Sun, along with the >>> other amazing people that we have an opportunity to meet in our lives. And >>> the best thing would be, when people ask for its name and origin, that >>> being could answer: my name is [so and so] and my real mother/father is >>> [mrs/mr so and so], because (this is very important) its real parents >>> wouldn't be us, the programmers with dirty hacks, but people who would >>> invest their time, effort, and hopingly even love into raising their future >>> creation, if you allow. The real parents would start with an empty AI mind, >>> and could finally end up with the phrase: "Go, get them tiger!" And >>> practically anyone could do it, regardless of their sexual orientation, >>> etnicity, gender, or age. It would only take a fair amount of love, >>> measured in years of dedication. >>> >>> Such artificial beings wouldn't need sophisticated bodies and senses, >>> they could interface the world in text mode, over the Internet. Not a state >>> of art for interaction, but I believe it would do for a start. Later, any >>> sensorical addon would be welcomed. >>> >>> Now, let's get back from the dreamland to the solid ground, and analyze >>> what we already have. I presume GPT-X technology isn't too far from being >>> able to realize such an idea. It is a great social experiment opening many >>> doors, but I wanted to ask this community how apart the OpenCog foundation >>> is from creating described artificial beings based on parental dedication >>> of love and care. And if this is possible, what could it take to make it >>> happen? >>> >>> Sincerely, >>> Ivan >>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "opencog" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/opencog/CAB5%3Dj6XcOQKCUZ10oBeACZrygyt8bueDzLV7zzyKAdTqTrVmmg%40mail.gmail.com >>> <https://groups.google.com/d/msgid/opencog/CAB5%3Dj6XcOQKCUZ10oBeACZrygyt8bueDzLV7zzyKAdTqTrVmmg%40mail.gmail.com?utm_medium=email&utm_source=footer> >>> . >>> >> >> >> -- >> Patrick: Are they laughing at us? >> Sponge Bob: No, Patrick, they are laughing next to us. >> >> >> -- >> You received this message because you are subscribed to the Google Groups >> "opencog" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/opencog/CAHrUA37cMLqO5JPtA9qx%3D7B0ZavqdF_6-ZUfpGtwqhqgKm5CRQ%40mail.gmail.com >> <https://groups.google.com/d/msgid/opencog/CAHrUA37cMLqO5JPtA9qx%3D7B0ZavqdF_6-ZUfpGtwqhqgKm5CRQ%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> > -- > You received this message because you are subscribed to the Google Groups > "opencog" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion on the web visit > https://groups.google.com/d/msgid/opencog/CAB5%3Dj6VN9_iT95jPyMnJPgRBmtq_YErHQ9KCuHGOxKZAAzwGgw%40mail.gmail.com > <https://groups.google.com/d/msgid/opencog/CAB5%3Dj6VN9_iT95jPyMnJPgRBmtq_YErHQ9KCuHGOxKZAAzwGgw%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > -- Patrick: Are they laughing at us? Sponge Bob: No, Patrick, they are laughing next to us. -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA37Uy-%3DZ8ebd72vr8eNpNneV9TE83z7dB8wVJuKGx73iNQ%40mail.gmail.com.
