On 2/24/19, Matt Mahoney <[email protected]> wrote: > Colin, I think the source of our disagreement is that we have very > different ideas about what we mean by AGI. To you, AGI is an autonomous > agent that learns on its own by doing science (experiments). It is > completely general and can work in any field like a real human. To me, AGI > means automating anything we might have to otherwise have to pay a human to > do.
It seems like you are both right: AGI should be an autonomous agent the learns on its own by doing science (experiments) AND AGI should be automating anything we might have to otherwise pay a human to do. (Now all that remains is the trivial task of development .... :) > > You believe (I think) that AGI is not even possible in conventional > computers. I tend to agree. First, transistors use too much power, about a > megawatt for a human brain sized neural network. We might achieve tens of > kilowatts using neuromorphic computing at the physical limits of > minituration. > > Second, the brain is optimized for reproductive fitness, not universal > learning, something Legg proved is not even mathematically possible. > Instead we are born with 10^9 bits of knowledge encoded on our DNA. That is > half of what we know as adults, and that knowledge took 3 billion years to > program at the rate of one bit per generation. For example, you cannot > learn to remember a 20 digit permutation on a screen and immediately recall > it back no matter how much you practice, which a something a gorilla can do > because its DNA is different. > > Third, we do not even want autonomy. Then we would have to deal with human > limitations like emotions and the need to sleep, take vacations, and get > paid. We work around these limitations and train humans to specialize in a > million different fields because that is how an organization gets work > done. > > I don't expect AGI to look anything like a human. We already automate 99% > of work using specialized machines that are vastly better at their jobs > than humans could ever be. We don't want autonomy. We want to be in > control. We want to asymptomatically approach 100% as the cost of the > remaining human portion rises at 3-4% per year as it has for centuries. AGI > is not a robot revolution. AGI is more productivity with less effort using > machines that can see and understand language, sense but not feel, know > what we want without wanting, and recognize and predict human emotions > without having any. > > On Sun, Feb 24, 2019, 1:56 AM Colin Hales <[email protected]> wrote: > >> Matt: >> >> "When you put millions of these specialists together you have AGI." >> >> No you don't!!!! Says who? (1) Where's the proof? (2) Where's the >> principle that suggests it? You have neither of these things. Even if you >> had both these things you'd still have to build AGI and test it assuming >> they are false in order to do the science properly. >> >> And you do not get to 'define it' this way. >> >> This is SCIENCE. You have an artificial version of a natural general >> intelligence when you've built one and it (yes the AGI itself, not your >> or >> anyone else's blessing) AUTONOMOUSLY proves it. Like fire. Flight and a >> million other things. >> >> I can think of 1000 things that such a specialist-narrow-AI-collection >> doesn't cover (like everything that science does not know, but could find >> out), and that a natural general intelligence can autonomously learn >> (find >> out), but that 'collection of specialist narrow AI' lacks .... along with >> an ability to _autonomously_ learn it, which is also something natural >> general intelligence (yes us) can do. And even worse: such specialist >> collections have ZERO intelligence, not because it doesn't know >> something, but because it lacks the _autonomous_ bit. A real AGI can know >> absolutely NOTHING and yet have non-zero intelligence because it includes >> a >> means to autonomously find out. Like us. >> >> So that collection cannot be an "artificial version of a natural general >> intelligence". >> >> End of story. >> >> Real AGI is defined by how it autonomously handles what it doesn't know >> (its ignorance!), not by what we bestow on it. If we bestow a means for >> learning X, that too does not make it AGI. A real but artificial version >> of >> a natural general intelligence has to _autonomously_ learn how to learn >> something it does not know, like us. To prove it you first prove it does >> NOT know it. Then, later, after learning, when it proves it does, >> autonomously, it gets a stab at the AGI gold medal. >> >> Meanwhile? Nothing wrong with what's being done. Powerful. Useful. >> Impressive. All good. Rah rah rah, carry on. BUT: Its AUTOMATION based on >> natural general intelligence, not AGI and it has ZERO intellect. >> >> [image: AGI.JPG] >> >> The entire AGI project is foundered on this basic fact of the science. >> >> And all we ever get here is the endless echo chamber of "if only we can >> program enough computers" (= AUTOMATION) AGI will magically appear. >> Rubbish. It was rubbish 65 years ago and we've done nothing but prove >> more >> completely it is still rubbish. >> >> Enough. >> >> I know you'll never face it. Forget it. >> >> >> >> On Sun., 24 Feb. 2019, 11:08 am Matt Mahoney, <[email protected]> >> wrote: >> >>> OpenCog is one open source effort. But real progress in AI like Google, >>> Siri, Alexa etc. is not just software. It's hundreds of petabytes of >>> data >>> from the 4 billion people on the internet and the millions of CPUs >>> needed >>> to process it. It's not just something you could download and run. >>> >>> I realize it's not AGI yet. We are still spending USD $83 trillion per >>> year for work that machines can't do yet. There are still incompletely >>> solved problems in vision, language, robotics, art, and modelling human >>> behavior. That's going to take lots more data and computing power. The >>> theoretical work is mostly done, although we still lack good models of >>> humor and music and much of our own DNA. >>> >>> If you want to make progress, choose a narrow AI problem. When you put >>> millions of these specialists together you have AGI. Don't try to do it >>> all >>> yourself. You can't. >>> >>> On Sat, Feb 23, 2019, 12:33 PM Ed Pell <[email protected]> wrote: >>> >>>> All, why is there no open source AGI effort? >>>> >>>> Ed Pell >>>> >>> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* >> / AGI / see discussions <https://agi.topicbox.com/groups/agi> + >> participants <https://agi.topicbox.com/groups/agi/members> + delivery >> options <https://agi.topicbox.com/groups/agi/subscription> Permalink >> <https://agi.topicbox.com/groups/agi/T69c9e23ba6b51be9-M908fd539511a9ae0f08c9a65> >> ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T69c9e23ba6b51be9-M4108310fea95babd1d8c29be Delivery options: https://agi.topicbox.com/groups/agi/subscription
