Half a croissant, on a plate, with a sign in front of it saying '50c'
h a l f b a k e r y
We are investigating the problem and will update you shortly.

idea: add, search, annotate, link, view, overview, recent, by name, random

meta: news, help, about, links, report a problem

account: browse anonymously, or get an account and write.



Massively Distributed Strong AI

With today's technology! (and a lot of money)
(+1, -1)
  [vote for,

It is felt by some futurists that an AI with the generalizing abilities of a person would, having access to the whole internet and a capacity for both human (generalized, imprecise thought) and machine (fast math, fast logic, fast knowledge-searching, excellent memory, broad knowledge) intelligence could design a better unit like itself, which could design a better unit like itself(2) and so forth. This is theorized to be the dawn of a new utopian era, as knowledgable and powerful machines devote themselves to the service of mankind.

Strong AI, as it is called, has been beyond the abilities of mankind.

I propose a distributed AI that would be comprised of many wirelessly-networked, ad-hoc machines. If most of the US and Europe had one of these (each person) they could connect and make a world-spanning brain more powerful than a human. The machines would consist of a wireless networking chip, a weak CPU, a power supply, and a little memory. They would be sold for a few dollars each, and marketed thusly: "Own a piece of the future!" The buyer would be expected to plug her unit into the wall and ignore it. (you can add some blinky lights if that would make it sell better) Each unit would process a hundred neurons, CAD-style, and send the input/output to ajoining machines. Access points to be created at various universities.

Voice, Oct 31 2007

Wikipedia: Singularity http://en.wikipedia...logical_singularity
What the point of smarter-than-human AI is usually called. [jutta, Oct 31 2007]

Strong AI http://en.wikipedia.org/wiki/Strong_AI
[zen_tom, Oct 31 2007]

Emergence http://en.wikipedia.org/wiki/Emergence
[zen_tom, Oct 31 2007]

Emergence indeed http://www.wired.co...s-hits-drone-fleet/
[theircompetitor, Oct 07 2011]

\\Not a few of a person's brain cells communicate to other cells physically distant\\ http://www.google.c...AQ&biw=1040&bih=753
Pretty. [mouseposture, Oct 07 2011]

More of the same http://www.humanconnectomeproject.org/
[mouseposture, Oct 07 2011]

You probably never heard of https://en.wikipedi...g/wiki/Mary_Midgley
[pashute, Jun 17 2015]


       // If most of the US and Europe had one of these (each person) they could connect and make a world-spanning brain more powerful than a human.   

       No, actually, they could not.   

       Transputer networks are an old hat, but we have no clue how to actually run a brain on top of them, or on top of anything else. (Partially because we still don't deeply understand how biological thinking works.) Your vague suggestion is well within what science-fiction authors and computer scientists have been proposing in the last thirty years or so. It wouldn't surprise me if the very idea of distributed computing had taken a clue from the way neurons connect to each other; it's hard to look at those colored cell clusters and not want to try it in hardware.   

       As an aside, once people figure out how to do this, I'm pretty sure it'll be several orders of magnitude easier, faster, cheaper, and more power-efficient in a single computer.   

       Two book recommendations in this context: Charles Stross's "Accelerando" is sci-fi set in and around this singularity event; its characters run into some quirky little problems that made me go "ha!". The other book I'm reading right now is "The Three-Pound Enigma", and intersperses portraits of neuro researchers with an introduction to the biology of the brain.
jutta, Oct 31 2007

       Actually, recently a supercomputer did model a bit of brain using CAD. The researchers found effects like relay-firing similar to the real thing. CAD brain modeling is no longer the realm of science fiction. If you are proposing that a large, flat brain would be less capable than a more compact style you could be right, but that can be worked around by connecting clusters in various places to other clusters in other places. (i.e. hook up some of them far away from each-other using the internet) The advantage of my proposal is that it can be done with today's technology. Also different from transputers because each cell is not trying to connect to anything thats not close.
Voice, Oct 31 2007

       <checks the 'bakery before going to bed>   

       Hail Degener! Hail Stross!   

       <goes to bed>
normzone, Oct 31 2007

       CAD brain modelling isn't Science fiction, but it's still in its infancy in terms of real-brain simulation. There's just too many neurons, and too many connections.   

       Interestingly, I think the bit that's missing here is the inputs and outputs. A human brain has a distinct set of inputs, sight, smell, touch etc - what distinct inputs will this distributed computer have? Even if the sim-brain does attain the critical mass of neuron-like connections to become self aware - what will it do? It could be that you need a distinct and coherent input set in order for a brain to begin functioning.
Then, there's the passivity problem. A brain in a jar is never going to be a particularly engaging entity. It has no output, even if it can make sense of the world, why bother if it can't act.
So you need a distributed control system for some real-world mechanism - and that mechanism needs to be coupled to a myriad of sensors so that there's some feedback (in the same way that it would be difficult to learn typing if you never were able to look at your fingers on the keyboard)
But the control for these manipulators would have to be wired into the overall system, in a way traditional electrical engineers wouldn't like - for example, our brain would be no more be able to output 'debug' text to a terminal than we are able to consciously change the state of any of our body-cells.
So this manipulator/sensor bundle might be put into space, and the highly distributed brain-network used to operate it.
Having done all that, you'd still need to figure out a way to communicate with it.
zen_tom, Oct 31 2007

       I understand 'Strong AI' to be the notion of a system can attain intelligence and consciousness.   

       Neural networks can indeed be trained, as can a suitably complex CAD model - but this is far from Strong AI.   

       I'm not convinced the jar/brain/keyboard/screen interface would work - Or rather, so much of our brain is wired to deal with body inputs, an intelligence that knew only character i/o would seem so alien to us, we might find it difficult to interact with it.   

       And this idea of connecting to a node of a massively connected neural network, and expecting a meaningful response has totally the wrong architecture - the analogy would be trying to communicate with someone by only having access to one of their braincells - sure, it's wired up to everything else, but at the individual node level, there's not a great deal of meaningful information that can be accessed. A brain is the sum of all the activity in its network of neurons, and its ability to take inputs and provide outputs is managed via a complex set of specialist i/o mechanisms that make up the body.   

       Asking a neural net to explain its reasoning is notoriously difficult - it's almost an inherent feature of massively parallel systems that are capable of otherwise non-linear computation not to be able to explain how they arrive at the conclusions they make.   

       'Programming' a net is another difficult problem - yes, they can be trained - but how would you train a global system to do anything when its practically impossible to control its inputs and outputs - this is why I think the i/o would have to be clustered, otherwise, there would be just too much noise, for anything meaningful to crystallise - maybe, global conditions, changing seasons, weather patterns etc would allow a coherent pattern of i/o for a global collection of nodes to begin coordinated action - and maybe distributed inputs from people (i.e. like people connecting to things on the internet) have coherent and synchronised patterns - but it's certainly not something we can control to the extent required for any kind of formal training to take effect.   

       So while it is a fascinating idea, and maybe it could work to the extent that some form of intelligence emerged, I don't see how we'd ever be able to discover it. It would be like an individual brain-cell suddenly realising that it's part of a thinking being. Each cell is completely oblivious to the processes it responds to and is a part of, and even if it was aware of the greater whole, it would never be able to understand it. A neuron's reactions and motivations are on an entirely different level to the overarching intelligence in which it participates.
zen_tom, Oct 31 2007

       Sure, we're doing fine with keyboards and screens - but we have the benefit of all our other modes of experience to give us something to talk about. If our only experience were limited to strings of characters, how would we know what meanings lay behind them? My laptop is never going to properly know what a "bacon sandwich" is - but I can talk to you about one, even if you've never seen bacon, or two slices of bread before, because we share similar sensoria. Brain-jar man, Neural-net boy, and AI girl, without analogous sensual experience, are all going to lack that commonality - making all but the most abstract concepts completely alien.
zen_tom, Oct 31 2007

       I think our emotions are nature's programming. Our DNA's survival is furthered through our most basic, emotional drives - if you were to program a neural net, the only way you'd be able to do so would be through something analogous to emotion, logic just wouldn't work.   

       Without a drive (whether it's called emotion or programming is immaterial) there's no motivation - but for something to be intelligent, the drive has to be non specific enough for it not to override the decision making process. e.g. If I feel cold, I might react in any number of ways to warm up - or, I might learn to like the cold. That's where the 'free will' comes in. If one were a religious type, you could say that God created us to feel, but gave us the choice as to what to do about it.   

       So an AI would feel emotion - just as any living creature would - I'm not sure what kind of emotion it would have.   

       I don't think it would be any more depressed to discover it was created by someone any more than we get depressed about growing in someone's belly - that is to say, emotion isn't something necessarily human, but that it's something necessary for intelligence, of any kind.   

       If anything, I think the early AIs will be terribly emotional, their designers, in an eagerness to make them compliant and easy to motivate, will juice up the emotions as much as possible. Getting the balance right will be the hardest part (and something we've not quite got the hang of ourselves just yet)   

       If you want a hard, cold calculating machine, then traditional computing, or some powerful, but specific perceptron type thing would be the best tool.
zen_tom, Nov 01 2007

       As has been mentioned, [Voice], it's a non-starter.   

       Forgive me for being a little trite, but I remember being told that there are more neurons in the human brain than there are known stars in the sky. However, even if you make the assumption that we do have the computing power to make such an AI beast... If we make it geographically distributed then we have to manage the communications between nodes - that sounds like an exponentially difficult problem, just in terms of node addressing and latency of messages.   

       That brings up the next issue of node topology: since brains are grown, we don't have a very good model topology to recreate - which, in any case, would have to be a logical topology built on top of a physical topology. That's going to be a nightmare to work out. The remodelling of these flat clusters will represent a bottleneck - one that will be unpredictable and likely affect performance significantly. Just imagine if each relay-firing was an IP packet!   

       Then we get to the best bit - we make it wireless. Wireless communications are problematic for many reasons - most of which due to the terrible channel (with reflections, interference, attenuation, fading, and all manner of new ways for errors to creep in to the system).   

       In my opinion, the best thing about this whole idea is the potential for a science-fiction-magic story about mobile phones becoming one distributed sentient intelligence (a la Skynet from the Terminator films).
Jinbish, Nov 01 2007

       Actually, I think wireless sounds like a very sensible way to do communication in a brain-emulator. Brain cells work on the order of a few hertz or a few tens of hertz, so communication needn't be blindingly fast if you're trying to emulate a brain (in fact, I think delay may be a crucial part of some of the brain's functions). You could therefore afford to waste plenty of time on encoding the data, and you'd need fewer wireless channels (eg, stick an 18-digit prefix on every packet, and you can address it to one particular "neuron" out of a billion billion).   

       Having said all that, though, bear in mind that I have no idea what I'm talking about.
MaxwellBuchanan, Nov 02 2007

       (chorus) we have been found out by Jinbish. Exterminate him.
theircompetitor, Nov 03 2007

       Jinbish, bigsleep, Few of a person's brain-cells actually communicate to other cells physically distant. They talk only to their immediate neighbors, except for the big mess of nerves in the middle. That mess of nerves is a lot of data, true, but nowhere near the communication between individual cells. Each neuron talks to its neighbors. My proposed device connects to its neighbors the same way.   

       As for I/O, universities would have whole clusters of nodes, spread out enough to communicate with the whole thing...
Voice, Nov 05 2007

       //CAD brain modelling isn't Science fiction// Using open-source software, you can model a single neuron on a reasonably high-end laptop. The amount of work to build the model has, in recent years, dropped below 1.0 doctoral dissertations -- roughly 0.2, and there are public libraries of already-built models.   

       Well and good. That's the easy part. These models grossly oversimplify the synapses (points of connection between neurons), which greatly outnumber the neurons. But that's still not the hard part.   

       The area of greatest ignorance is the *pattern* of synaptic connections between neurons. It is very, very far from being "nearest-neighbor" or anything remotely like that, in biological nervous systems.   

       //Few of a person's brain-cells actually communicate to other cells physically distant. They talk only to their immediate neighbors, except for the big mess of nerves in the middle.// <link> <link>   

       This is untrue, for example, of cerebellar granule cells, which are the most numerous type of neuron in the brain, and constitute almost half the neurons in the mammalian central nervous system.
mouseposture, Oct 07 2011

       TIL [mouse]. Thank you
Voice, Oct 08 2011

       //The buyer would be expected to plug "Her" unit into the wall and ignore it.// What is this female Macho sexist crap?   

       And so Millicent took her invention to the market, in so doing, created a race of super computers owned by fat dumb humans, suffering from neurosis of the shiver.
Ah Supp, Oct 08 2011

       <checks the 'bakery before going to bed>   

       Hail Degener! Hail Stross!   

       <goes to bed>   

       Sometimes I find what I think is a new idea and amaze myself...
normzone, Oct 08 2011

       Of course there is a cheaper solution. Given mortals use only one half of the brain, right side of brain kicks in for left handed people left side of brain kicks in for right handed people. Knowledge power could quite easily be doubled by only allowing left and right handed people to marry. Catholic Church scientists experimented with this idea in the middle ages. It was complete and utter failure, every left right marriage only ever produced IrishXFrench children unable to understand a word they were saying to themselves.
Ah Supp, Oct 09 2011

       I think the problem with this is that it lacks many things, including user motivation.   

       A modified suggestion:   

       Provide software that implements a networked algorithm (for instance, a networked neural net) and some initial behavioural rules. This is run in the background by anyone who's online with the software.   

       Let the "players" form (or join) teams, perhaps with a few thousand or tens of thousands of computers per team. Each team acts as an "intelligence".   

       Let the human players challenge and evaluate the other teams (for instance, by a Turing-style test).   

       Then, add a points system. I suggest a sort of "share scheme", whereby players can buy into a given team; the team's share price will be determined (as for real shares) by a combination of its performance and market forces.   

       Finally, let the shares be worth real honest money.   

       The end result is that you have several competing, distributed evolving AIs, propelled by human greed, which is surely the best assurance of success.
MaxwellBuchanan, Oct 09 2011

       Why bring computers into the frey of humans being propelled by greed?
Ah Supp, Oct 10 2011

       I'm sorry. I can't do that, Dave.
not_only_but_also, Oct 10 2011

       Maybe it's worth buying up a load of old PCs to try this on.... open the eBay doors, Hal.
MaxwellBuchanan, Oct 10 2011

       Looking back, it seems I've been talking to myself for a while there!   

       I'm sure it's possible yes, but the fact remains that it would be difficult to communicate meaningfully with such a machine as its locus of sensitivity would be so alien to us.   

       A brain is there to process stimulai perceived by a sensorium from the outside, and translate it into action. That definition works for snails, autopilots and humans - but it has 2 requirements -
i) the ability to sense,
ii) the ability to react.
Without a stimulus, and the ability to respond - there's no point in doing any processing. And if there's no point, no direction, then how will it be possible to understand whatever murky stirrings that may result?

       For intelligence you need *intent*. When you watch a cat stalk something, you can see the *intent* of the cat. The Turing Test for intelligence should be the ability to detect and recognise *intent* in another thing.   

       Go back to a brain in a jar with no external interface - yes, it could be whirring away - but it would be freewheeling - and you'd never be able to communicate with it.   

       If the world were a brain - who's neurons were electronically connected devices - what would it sense? Maybe variations in the solar wind, or regular x-ray pulses from distant galaxies. Perhaps it would be aware of its regular procession around the sun, day-night rhythms, and changes in radio frequencies/chatter as it sweeps the sky - the thing it was able to sense would be the things that came in from outside. We, the system's components wouldn't, couldn't be able to know what the whole was up to.   

       If you wanted a Strong AI that was on the same level as us humans, you'd need to construct it in such a way that you could craft its sensory inputs so that they were similar in experience to those of another person. You'd need a body for it to be able to react to external stimuli, and you'd probably need to send it to school to socialise with other robots/humans in order for it to gain any real understanding of inter-personal relationships, teach it language and other human-specific interface techniques.   

       There are people* who sometimes post on the bakery who are difficult enough to understand. Imagine trying to understand a global-sized brain who's inputs were totally outside of your own personal experience, that had never heard of language, and which slowly blinked into conciousness alone, with only a limited sense of who or what it was, and without any idea as to how it got there.   

       It would be more considerate to try to segregate the interconnecting networks in such a way that there were two distinct entities, so that when they did emerge into wakefullness, they would at least have someone to talk to.   


       *we assume here, for the sake of argument that they are people
zen_tom, Oct 11 2011


back: main index

business  computer  culture  fashion  food  halfbakery  home  other  product  public  science  sport  vehicle