Sorry for the double but my mail client somehow formatted everything as a quote.
> On 2 Dec 2019, at 09:39, Gentoolx <[email protected]> wrote: >> On Dec 2, 2019 at 2:50 AM, <Matt Mahoney> wrote: >> >> I don't believe anyone on this list is working on SOAR. It is an old system >> (1983). Back then the only thing you could do in AI with available computing >> power was structured knowledge representation, rule based language models, >> and expert systems. Any work done with autonomous agents was only in >> simulation. There was no vision, hearing, speech, or robotics. Language >> understanding was brittle, with little resemblance to the way humans process >> language. We learn semantics before grammar, the opposite of machines, which >> is why they are so bad with ambiguity. >> >> OpenCog by Ben Goertzel, who posts here occasionally, dates to 1995 if you >> include it's precursors Webmind and Novamente. It has many similar >> limitations due to hardware. The atomspace architecture is supposed to >> support structured knowledge, probabilistic reasoning, induction, and >> learning, but there is nevertheless no knowledge base or useful >> applications. The evolutionary learner MOSES and neural vision system DeSTIN >> only work on toy problems and were never integrated with atomspace as it was >> designed to be. MOSES is being integrated with AtomSpace, see the “as-moses” repo on GitHub. DeSTIN has been dropped in favor of modern Artificial Neural Network models, which makes sense given their limited manpower. >> The last public demo was in 2009 of a puppy in a virtual world. Since then >> there really hasn't been any basic research. >> >> I don't mean to be critical but AGI is a really hard problem which no >> individual on this list has the resources to solve. Google, Amazon, Apple, >> Facebook, and Microsoft have made some progress, but these are companies >> with trillion dollar market caps. You don’t magically make progress in reasearch by throwing money at it - it won’t make the people working on it smarter, it does however allow you to hire lots of developers to make high-quality software for you and to use lots of data and compute to train statistical models. Current AI techniques still result in models that are often brittle and show signs of not being really congruent with human mode of learning - so much for that basic research done by those trillion cap companies, I guess... >> >> A human brain sized neural network needs 10 to 20 petaflops and a petabyte >> of RAM. Our software, encoded in DNA, is equivalent to 300 million lines, or >> $30 billion. And then you have to train it on an exabyte of video. >> >> But this approach doesn't even make sense. Our whole economy is based on job >> specialization. It is far more efficient to organize machines like we >> organize people, each doing a specific task. Everyone making progress in AI >> is doing narrow AI, and really this is the only practical approach. Instead >> of trying to automate a million different jobs all at once, you'll have more >> success automating one job. That's going to be hard enough, given that all >> the low hanging fruit has been picked. >> >>> On Sat, Nov 30, 2019, 11:28 AM digikar via AGI <[email protected]> wrote: >>> I am a third year student pursuing a Bachelors in Computer Science and >>> Engineering, and have been wanting to get into AGI, since, two years may >>> be. I discovered OpenCog, and felt it to be too daunting - like I think >>> I'll require another year or two of study to make good sense of it. >>> >>> I studied some first language acquisition the last summer (along with a >>> basic Andrew Ng's ML course, another NLPwDL CS 224n from Stanford, and a >>> more rigorous and exhaustive (than the Ng's anyways) Foundations of ML at >>> my own university). Reading about first language acquisition led me to >>> believe that a primary problem is being able to represent the world (with >>> as much details as possible, since dealing with block worlds is easy). So, >>> this is the primary issue. >>> >>> Recently, I spent some time with Natural Semantic Metalanguage, and its >>> criticisms. The concept is definitely ambitious; however, that definitely >>> doesn't seem to be the way our thoughts work. For instance, see explication >>> for "left"; for me, 'left' just evokes a direction than all the other >>> things. May be, it might be useful for actually representing the world in a >>> computer, but an explicit simulation (with which I haven't worked with yet) >>> seems more wieldy. >>> >>> I found SOAR ambitious and more "established" - however, their forums >>> seemed void. So, just wanted to know if anyone is working on it. >>> >>> Other than that, does their exist some system for representing the world - >>> gaming systems come to mind, but is their some established standard? >>> >>> Thanks! >> >> Artificial General Intelligence List / AGI / see discussions + participants >> + delivery options Permalink ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tfbb2a2f3b421ca41-M6a417b12bf48c15394309335 Delivery options: https://agi.topicbox.com/groups/agi/subscription
