Hi Aaron, Matt, *>Aaron:So maybe "programming" will eventually come to mean "demonstrating a task for a robot" *>So it seems clear that software that learns by example is the way to go, rather than hand-coding each possible task.* *etc.
Absolutely that's the direction! Programmers - sorry dudes, nobody will need millions of "experts" to copy-paste and adjust a few parameters of the same web-database-css-html code for the next dummy web site project. Even better, the machine could learn without demonstrating it explicitly - a lot of it **is* *already "demonstrated" and explained in all kinds of records - video, books, diagrams, manuals, novels. And will be in the interactions with the machine, and with its own interactions when it's playing and exploring on its own or with other "machine friends" and human. It will be in its childhood experiences. >Let's assume that we have cheap, powerful hardware (enough to model >human brains) and that we solved the hard AI problems like language, >vision, art, robotics, and modeling human behavior. Sorry, but I'll object again, none of those is a hard problem, at lest for some AGI or natural GI agents. Maybe except the robotics - it is the hardest of them, more precisely the mechanical part is the most expensive to be produced in current economy. All the rest is trivial and coming to automation really soon, much sooner than your terribly wrong calculations suggest, and for your surprise, people won't be very happy about losing their jobs in current social-economical system - I'll put that in another publication, though. *Regarding the "hard" problems - for most people even elementary tasks like drawing or playing musical instruments or dance (I mean real dancing, not slow shaking) or acting, or basic Calculus or speaking a non-native language without an awful accent or making correct sentences as long as this one etc. etc. are not just hard problems - they are impossible for them to cope with. Most humans are helpless and the range of new skills that they can acquire is pathetic. That's not a mechanical problem - they just can't get it and their "petaflop" brain can't deal with it and will not deal with it, even if you flood it with 99999 "Plenta-Bytes" of music, pictures and dance moves of which the real information content is 1000 bits. They will not get it, because their brain is too stiff and poor and because it has intrinsic bottlenecks and tiny little thresholds that it cannot overcome and cannot incrementally build or record a chain of meaningful operations. On the other hand, musical prodigies just see the musical correlations at a glance - even without practice, immediately after sitting behind the instruments. The same goes for other arts. They can see it and reproduce it or improvise and develop a piece, starting from a given example, without need for explicit instruction - how to play an instrument is encoded in its design, in the design of human body, and in the acoustics and the operation of our brain. They can learn and invent extensions in an hour, a day or a week things that others can't understand for a lifetime. They can also explain to the not-talented how they do it, what they should do - what notes to play, how to practice, what to draw first, what do they are supposed to hear. Unfortunately that won't help - the untalented ones will not get it, even though it's encoded in 10, 50, 100 or 1000 bits.* Some humans have a true or close to true general intelligence, but the majority - don't really, human general intelligence is a big fat myth and a lie - humans early reach terrible thresholds and their brains specialize. Most people learning abilities are rather like of a semi-programmed "robots" that can't learn new things and see world as "magic", even if you demonstrate them and explain it step-by-step. > > - *From:* Matt Mahoney <[email protected]> > > > - *To:* "[email protected]" <[email protected]> > > > - *Subject:* Re: [agi] More on Baxter - "the robot pc" > > > - *Date:* Mon, 7 Jan 2013 18:26:52 -0500 > > > What is your estimate of the cost of training a robot to clean houses? > Assume you have enough computing power (say, 10 petaflops and 1 > petabyte), but no software. Or maybe you can name existing software > that might be useful. > In order for a robot to watch someone clean, it first has to be > programmed to see. Humans learn to see after training on about 1 > petabyte of high resolution video. Assume you can download all of > Youtube, although this might not be all you need and it will take a > lot of effort just to discover this fact. In order for a robot to ask > questions and understand the answers, it must also have a language > model grounded in sensory experience and action. (Watson was a $30 > million effort and was not grounded). You also have to add hearing, > speech recognition, and motor control. It takes a baby a year to learn > to walk. How long will it take to train a robot? > I assume that knowledge can be copied to identical robots. There will > be some training cost paid to human teachers for any job that is even > slightly different. How much will it cost to train robots to do all > the work that humans do? > We can already train robots to perform repetitive motions by guiding > its manipulator. This isn't the same thing as a human picking up an > object, guided by vision and tactile feedback. That is still a mostly > unsolved problem. Matt, I challenge you that a thinking machine that will learn to see, to speak and to interact using the "unsolved" problem for multi-modal learning, including tactile, in virtual worlds, can be run even on a Pentium Dual Core E5200, even in 32-bit mode. Well, let it be overclocked a bit, say 3.33 GHz. The stock speed of 2.5 GHz is benchmarked at up to several GFLOPS, regarding: http://browser.primatelabs.com/geekbench2/344462 If there's some modest GPU - why not use it also, say a GeForce 8600 GT. It will learn to see through web cameras and watching recorded videos using any other cameras. And in order to be "fair", it should has a threshold of at least one year to learn to say the first word etc. (there are other milestones). >I assume that knowledge can be copied to identical robots. There will >be some training cost paid to human teachers for any job that is even >slightly different. How much will it cost to train robots to do all >the work that humans do? Absolute nonsense, not that I'm saying something new, but that's utterly non-AGI-sh like many of your posts. What you say is "computers can ever do only what you know exactly how to tell them how exactly to do and code it exactly each step of what it will do with every specific situations". Zero learning and adaptation - why do you need grounding then? That limits your programs to what your tiny working-memory can hold - a few bites. As of the work that humans are paid to do - a lot of it is for doing nothing - a bunch of inefficient time-wasting or useless role playing games to make the fake social system work as it is and keep the status-quo. An elaboration goes also to another article, though. ....* Todor "Tosh" Arnaudov ....* * .... Twenkid Research:* http://research.twenkid.com .... *Self-Improving General Intelligence Conference*: http://artificial-mind.blogspot.com/2012/07/news-sigi-2012-1-first-sigi-agi.html *.... Todor Arnaudov's Researches Blog**: * http://artificial-mind.blogspot.com More on this - in an article. > > - *From:* Matt Mahoney <[email protected]> > > > - *To:* "[email protected]" <[email protected]> > > > - *Subject:* Re: [agi] More on Baxter - "the robot pc" > > > - *Date:* Mon, 7 Jan 2013 18:26:52 -0500 > > > - > > > - > > On Mon, Jan 7, 2013 at 4:25 PM, Aaron Hosford <[email protected]> wrote: > > > > Reality is the best source to consult about reality. Why write a book > about how to clean a house if your audience can just watch someone doing > it, and ask appropriate questions when it gets stumped? And once a robot > has learned a task, the information can be copied without further human > demonstration. So maybe "programming" will eventually come to mean > "demonstrating a task for a robot" and cease having anything to do with > writing code. > What is your estimate of the cost of training a robot to clean houses? > Assume you have enough computing power (say, 10 petaflops and 1 > petabyte), but no software. Or maybe you can name existing software > that might be useful. > In order for a robot to watch someone clean, it first has to be > programmed to see. Humans learn to see after training on about 1 > petabyte of high resolution video. Assume you can download all of > Youtube, although this might not be all you need and it will take a > lot of effort just to discover this fact. In order for a robot to ask > questions and understand the answers, it must also have a language > model grounded in sensory experience and action. (Watson was a $30 > million effort and was not grounded). You also have to add hearing, > speech recognition, and motor control. It takes a baby a year to learn > to walk. How long will it take to train a robot? > I assume that knowledge can be copied to identical robots. There will > be some training cost paid to human teachers for any job that is even > slightly different. How much will it cost to train robots to do all > the work that humans do? > We can already train robots to perform repetitive motions by guiding > its manipulator. This isn't the same thing as a human picking up an > object, guided by vision and tactile feedback. That is still a mostly > unsolved problem. > -- -- ------------------------------------------- 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
