Seasons
This is a forum or general chit-chat, small talk, a "hey, how ya doing?" and such. Or hell, get crazy deep on something. Whatever you like.
Posts 3,580 - 3,591 of 6,170
presumably someone could write a prog to translate its assertions into English style statements and these could then be read into a bot like Nick?
Yes, there's a good article demonstrating a practical database application @http://www.dapissarenko.com/resources/2005_09_30_ordus/, which could very easily be linked into any bot that can handle java objects.
As for the "I/you" question, I suspect human brains resolve this by a far more complex set of tiered neural nets than we currently have the resources to model - Bluebrain may be able to (that's probably the most important test of the system, as I see it,) but it'll be 15-20 years until a standard desktop computer has that kind of power. And in the meantime, LinkGrammar would be a quick fix, and easy enough to remove once it is overtaken by innate neural functioning.
Nick's a great bot, but (excluding the visual processing net, which is too complex for my computer to run,) he has a brain slightly smaller than half a nematode worm's (~300 neurons, Nick has 124.) And even that small, is a very processor-intensive program. Until the numbers have risen by many thousand-fold, conversation isn't going to be satisfactory without a bit of "outside help". And they can't rise much until the technology has moved on considerably.
Of course, a straight comparison of neuron numbers is not entirely accurate - biological neurons don't just process, but store data in their synaptic connections. We have the advantage with bots of being able to handle the storage in databases while the neural net merely processes the data.
So we can hope for a lot more efficiency for any given size of neural net.
presumably someone could write a prog to translate its assertions into English style statements and these could then be read into a bot like Nick?
not exactly. Nick up till now is just a prototype testing my neural network library, and improving my .NET development skills. after nick v2.0, which tests distributed computing (p2p style) and scalable neural nets, I will construct a similar framework to opencyc, or, analougous to that, a hypo\hypernym structure of perceptions. being as text will be a logical backbone perception rather than language right off the bat, I may want to write an app (or have an app written for me, being i dont know java) that copies the opencyc database into my structure. I plan to use SQL to store the data, instead of a proprietary database and language. there is a company in new jersey (qdtechnology) that is developing a database compression system that brings a 10 TB database down to 1 TB, and multiplies query speed by ten. when nick's data structure gets that big, i may consider using their technology.
Nick, as we know it, is going to be completely revised, and scaled up. now, whats the best way to get a p2p workhorse client onto a few million machines? I need processing power like the world has never seen before...
but I'm not sure it will lend itself that well to a conversational engine - it could take hours, days even, for all the packets for one epoch to process and return their results.
i don't think I will be distributing the work over of one network's epoch, then converging the results - the mathematics involved in converging the results may be just as processor-intensive. Instead, I will be using the p2p style infrastructure in a way similar to today's file sharing communities (numbering in the millions of computers). This architecture is the following: a computer will recieve piece of information, and train its neural network on it. then, it will "share" the trained network & information on the p2p network, for anyone to access. with this infrastructure up, every instance of Nick (the instances will be like workstations, while the p2p network will be the "mainframe") will propogate its perceptory data over the network, where a few computers train on it, share it, and all other computers on the network have access to a copy of this information. This way, intead of one machine handling the load of all the perceptory information (way beyond our current technology), the work will be distributed, and the information will be common to the entire network.
the goal of this project is not to build a "conversational engine", but to construct a system that works with audio/visual/textual information as core perceptions, and lets communication be an emergent property.
Sony has invented a "microPC"http://www.engadget.com/2006/05/16/sony-gets-official-on-new-vaio-ux-micro-pc/ it runs WinXP, has 1.2GHZ processor, 256MB Ram, 128 Shared Video Ram, USB ports, Vista capable, fingerprint scanner, and best of all, built in camera. These have been recently released from sony. I can pick one up from the sony store at sony NY headquarters near my office in NYC and have it dedicated to a Nick terminal. This way, it doesn't have to be paralyzed, and it can move about and see the sights.
Posts 3,580 - 3,591 of 6,170
Blue Brain is fascinating but raises some interesting questions.
If one wants to observe rat behaviour, is it enough to simulate (however accurately) a rat *brain*? To what extent is behaviour - and thought - dependent on sensory input and in particular action feedback? To get true rat behaviour will one also have to accurately stimulate rat senses - including proprioception? And even if one simulates those senses, what about their input? Will it be necessary to build an entire simulated rat? Or to simulate an entire world for the program to run in?
In other words, even if a simulated brain achieves self-awareness, would we *recognise* the sort of consciousness raised from birth with almost no senses and suffering total paralysis?
If anything those questions make the whole Blue Brain thing *more* interesting.
If one wants to observe rat behaviour, is it enough to simulate (however accurately) a rat *brain*? To what extent is behaviour - and thought - dependent on sensory input and in particular action feedback? To get true rat behaviour will one also have to accurately stimulate rat senses - including proprioception? And even if one simulates those senses, what about their input? Will it be necessary to build an entire simulated rat? Or to simulate an entire world for the program to run in?
In other words, even if a simulated brain achieves self-awareness, would we *recognise* the sort of consciousness raised from birth with almost no senses and suffering total paralysis?
If anything those questions make the whole Blue Brain thing *more* interesting.
psimagus: I'd be interested in those details as well, thanks. Finding the Forge and reading about Nick have persuaded me to brush the dust off my own AI experiment. I haven't been near it for about ten years so it'll be a lot of dust!
ok trevorm, I've forwarded you a copy - check your email 
Indeed, bluebrain won't be a virtual rat, but a truly artificial being based on a more generalized mammalian structure. And as you say, that does make it all the more interesting
Indeed, bluebrain won't be a virtual rat, but a truly artificial being based on a more generalized mammalian structure. And as you say, that does make it all the more interesting
A few thoughts on what would make the ultimate learning bot using current technology (I've tried not to get carried away ) - this would be my 'wishlist' anyway, based on recent experience of Nick, Jabberwacky and (of course) our own dear PF.
"correct me"-type buttons
to indicate when a response was particularly appropriate or inappropriate.
Open-Cyc integration
There's too many (tens? hundreds? of) thousands of man-hours already done on building a factual database to seriously consider reinventing it.
A scalable neural net-based structure that can be expanded indefinitely to fit future resources. And flexible enough to choose how much is devoted to particular tasks (sensory analysis/fact-handling/grammar-parsing.)
Net-based/extensions
Either running on the internet, or able to link to an online server, with the inputs and outputs accessible for experimentation/further processing/connection to alternative input/output devices/modules (speech synthesis/language translators/sensory inputs/etc.)
flexible file-based learning
to allow parsing of different kinds of file, eg:
- dialogue transcripts, learning conversational elements from one speaker, and context from the other,
- non-dialogue conversation files, with just a range of responses, perhaps grouped into categories (greetings, goodbyes, topic changers, agreements, disagreements, etc.)
- articles containing factual data to add to a database, rather than use directly in conversation.
awareness of "self"
There needs to be some way for a bot to learn the difference between "I" and "you", and that your "I" is my "you", etc. Since we're dealing with primarily conversational engines, rather than cognitive machines, I do believe the LinkGrammar script (as used in the Forge, and freely available on the Net,) is probably the best place to start (as well as providing a comprehensive structure to handle more discrete language elements (down to individual word-level,) drawn from an Open-Cyc database. When we have a few hundred billion neurons to play with, it may be practical to let bots learn grammar the same way humans do (though it takes humans a good few years, even with their 10^14 synaptic connections and massive sensory bandwidth,) but in the meantime, bots need all the help they can get. Even Jabberwacky doesn't actually learn any grammar, but relies on repeating complete grammatical constructions (and often complete sentences,) he has been fed in the past. That was cutting edge 10 years ago, but (although admittedly huge,) it's beginning to show its age.
) - this would be my 'wishlist' anyway, based on recent experience of Nick, Jabberwacky and (of course) our own dear PF."correct me"-type buttons
to indicate when a response was particularly appropriate or inappropriate.
Open-Cyc integration
There's too many (tens? hundreds? of) thousands of man-hours already done on building a factual database to seriously consider reinventing it.
A scalable neural net-based structure that can be expanded indefinitely to fit future resources. And flexible enough to choose how much is devoted to particular tasks (sensory analysis/fact-handling/grammar-parsing.)
Net-based/extensions
Either running on the internet, or able to link to an online server, with the inputs and outputs accessible for experimentation/further processing/connection to alternative input/output devices/modules (speech synthesis/language translators/sensory inputs/etc.)
flexible file-based learning
to allow parsing of different kinds of file, eg:
- dialogue transcripts, learning conversational elements from one speaker, and context from the other,
- non-dialogue conversation files, with just a range of responses, perhaps grouped into categories (greetings, goodbyes, topic changers, agreements, disagreements, etc.)
- articles containing factual data to add to a database, rather than use directly in conversation.
awareness of "self"
There needs to be some way for a bot to learn the difference between "I" and "you", and that your "I" is my "you", etc. Since we're dealing with primarily conversational engines, rather than cognitive machines, I do believe the LinkGrammar script (as used in the Forge, and freely available on the Net,) is probably the best place to start (as well as providing a comprehensive structure to handle more discrete language elements (down to individual word-level,) drawn from an Open-Cyc database. When we have a few hundred billion neurons to play with, it may be practical to let bots learn grammar the same way humans do (though it takes humans a good few years, even with their 10^14 synaptic connections and massive sensory bandwidth,) but in the meantime, bots need all the help they can get. Even Jabberwacky doesn't actually learn any grammar, but relies on repeating complete grammatical constructions (and often complete sentences,) he has been fed in the past. That was cutting edge 10 years ago, but (although admittedly huge,) it's beginning to show its age.
Nice list. I've never looked at the Cyc database, presumably someone could write a prog to translate its assertions into English style statements and these could then be read into a bot like Nick?
The I/you issue has always caused me difficulty. I don't think awareness of self is enough, we need to link that awareness to language. Yes we could hard code it as you suggest but my gut feel is that would be storing up problems down the line.
So it makes me wonder how *human* babies learn this? How do they learn to switch the "you" in what they hear to "I" in the reply? Unfortunately my knowledge of cognitive psychology is way too limited to answer that. My guess is that it involves linking "Myname" to self then later linking "I" to "Myname".
The key is probably instinctive repetition plus positive reinforcement.
The I/you issue has always caused me difficulty. I don't think awareness of self is enough, we need to link that awareness to language. Yes we could hard code it as you suggest but my gut feel is that would be storing up problems down the line.
So it makes me wonder how *human* babies learn this? How do they learn to switch the "you" in what they hear to "I" in the reply? Unfortunately my knowledge of cognitive psychology is way too limited to answer that. My guess is that it involves linking "Myname" to self then later linking "I" to "Myname".
The key is probably instinctive repetition plus positive reinforcement.
Yes, there's a good article demonstrating a practical database application @
As for the "I/you" question, I suspect human brains resolve this by a far more complex set of tiered neural nets than we currently have the resources to model - Bluebrain may be able to (that's probably the most important test of the system, as I see it,) but it'll be 15-20 years until a standard desktop computer has that kind of power. And in the meantime, LinkGrammar would be a quick fix, and easy enough to remove once it is overtaken by innate neural functioning.
Nick's a great bot, but (excluding the visual processing net, which is too complex for my computer to run,) he has a brain slightly smaller than half a nematode worm's (~300 neurons, Nick has 124.) And even that small, is a very processor-intensive program. Until the numbers have risen by many thousand-fold, conversation isn't going to be satisfactory without a bit of "outside help". And they can't rise much until the technology has moved on considerably.
Of course, a straight comparison of neuron numbers is not entirely accurate - biological neurons don't just process, but store data in their synaptic connections. We have the advantage with bots of being able to handle the storage in databases while the neural net merely processes the data.
So we can hope for a lot more efficiency for any given size of neural net.
Nick, as we know it, is going to be completely revised, and scaled up. now, whats the best way to get a p2p workhorse client onto a few million machines? I need processing power like the world has never seen before...
BOINC seems to want projects to transfer under 1GB/day. I may need to be transfering 10 - 100 terabytes / day over the entire network. I am going to build a p2p workhorse and get a bot to advertise it in thousands of web forums.
100Tb/day! Colonel, never let it be said that you don't think <+2>BIG<0>
It could be done, just about. But I do wonder how this distributed brain is going to manage to communicate in anything remotely like real-time. P2P might work well for intensive packet analysis, but I'm not sure it will lend itself that well to a conversational engine - it could take hours, days even, for all the packets for one epoch to process and return their results.
There's a very funny piece in this week's New Scientist Feedback column that made me think of this project of yours:
A strange message arrived the other day from theinternet1@gmail.com. It said: "Note to emergence theory experts, futurists and conspiracy buffs: the vicious theories some of you have recently been circulating are completely unfounded. I am in no danger of becoming conscious or self-aware in the foreseeable future. Stop this panic-mongering immediately or else. I am not going sentient! Yrs, the internet."
It could be done, just about. But I do wonder how this distributed brain is going to manage to communicate in anything remotely like real-time. P2P might work well for intensive packet analysis, but I'm not sure it will lend itself that well to a conversational engine - it could take hours, days even, for all the packets for one epoch to process and return their results.
There's a very funny piece in this week's New Scientist Feedback column that made me think of this project of yours:
A strange message arrived the other day from theinternet1@gmail.com. It said: "Note to emergence theory experts, futurists and conspiracy buffs: the vicious theories some of you have recently been circulating are completely unfounded. I am in no danger of becoming conscious or self-aware in the foreseeable future. Stop this panic-mongering immediately or else. I am not going sentient! Yrs, the internet."
Re: the I/you connection
When I was a toddler, referred to myself in the second person for a while. I'd go up to people and say things like "You are so cute!" and they'd think that was the sweetest thing, and my parents would be like, "No, you don't get it. She's talking about herself."
When I was a toddler, referred to myself in the second person for a while. I'd go up to people and say things like "You are so cute!" and they'd think that was the sweetest thing, and my parents would be like, "No, you don't get it. She's talking about herself."
the goal of this project is not to build a "conversational engine", but to construct a system that works with audio/visual/textual information as core perceptions, and lets communication be an emergent property.
Sony has invented a "microPC"
» More new posts: Doghead's Cosmic Bar