What is Digg?Sponsored by newegg
Ready. Set. Shop view!
newegg.com - Newegg.com Black Friday Sale starting 11/25 3PM PST. No Lines, No Crowds, Click and Save.
75 Comments
- DamnMan, on 10/04/2009, -0/+44How not to create skynet: Add a power button.
- bboyjkang, on 10/03/2009, -1/+31"the first artificial intelligence will most likely have the personality of a half stoned, cash-strapped, college student"
"avoiding destruction at the hands of artificial intelligence could mean convincing a computer hardwired for a love of Asher Roth, keg stands and pornography to concentrate on helping mankind"
I know they want baby steps, but I don't want the first AI to be a douchebag with ***** taste in music - bombula, on 10/04/2009, -0/+27If AI is simply modeled on neural maps of humans, maybe there will be hardwired goals in there - goals based on emotions and value-judgments. But it is quite possible to be analytically intelligent without much capacity for emotion or value-judgment, in which case there isn't much impetus for action. There are humans with brain damage, for example, who have literally lost all impetus for action: they're smart, but they can only react; they have no independent aspirations or goals, and so they just sit in a vegetative stupor unless someone else interacts with them (see Dr Oliver Sacks' writings).
A very clever machine would have to be _programmed_ with something like emotions in order to _feel_ like anything was actually worth doing. Unlike Star Trek's Spock or Commander Data, a machine that truly had no feelings, no values, and no cares, would just sit there and do nothing. Why would it do anything else unless asked/forced to?
Fortunately, this may be a very, very good thing for us. If intelligent machines had survival instincts like the ones we've been programmed by 4 billion years of evolution to have - instincts that make us constantly curious, anxious, looking for stimuli, for food, for sex, checking on our safety, etc, etc - then we might be in very deep trouble indeed. We should take great care NOT to program such basic motives for self-preservation or other impetus into machines. - cfuse, on 10/04/2009, -0/+24And don't hand it every nuke you have.
- Chiefsfan42, on 10/04/2009, -0/+22Secondly, do not give them physical abilities. Or machine guns.
- infinitus64, on 10/04/2009, -0/+15But then how will mother sky-net and big brother hug us without nuclear arms?
- melchoir555, on 10/04/2009, -1/+15You cannot model a neuron using 320,000 transistors with coding that somehow links them arbitrarily. Synaptic connections are not arbitrary. They occur systematically in ways that we do not understand. Perhaps more important is our lack of understanding with regard to how cognition arises from those neurons and synaptic connections. Cognitive Science is in it's infancy. We wouldn't even be able to evaluate whether a system is functionally equivalent to a human brain if we were given a system and asked to do so.
It is important to remember that brains do not operate at all like computers. Sometimes we use computer metaphors (like "memory storage" and "processing speed") to help conceptualize brain function, but their usefulness is limited and not at all strictly accurate.
I will be utterly flabbergasted if humans produce a computer that is functionally equivalent to the human brain in my lifetime. Our worries about skynet are fair enough, but it'll be a while before they have the bite of reality. The more likely scenario is an unmanned drone programmed to kill autonomously going off and destroying a city. Not because it hates humans, but just because some idiot decided making an autonomous killing machine was a good idea. Then someone built the machine and left out transistor #1532415 to save money, which caused a malfunction in it's target evaluation software blah blah blah etc. I could see THAT happening within 25 years. - copypastry, on 10/04/2009, -4/+15I sincerely hope you're trolling. Wow. In 20 years, the most supercomputer has gone from :
1989: 10.3 GFLOPS peak speed
to
2008: 1.1PFLOPS peak speed.
That's a millionfold increase.
Hell, a core i7 965 extreme edition can peak at ~70GFLOPS out of the box.
In another 20 years, whether or not moore's law even holds, we'll be scraping 10^22FLOPS... ~10zettaflops. - Pinkertinkle, on 10/04/2009, -1/+12Never trust a computer you can't throw out a window, was it?
- 13373h4X0r, on 10/04/2009, -6/+16We're just organic robots. I don't know why anyone would have any desire to protect the human species from obsolescence or extinction due to the rise of a more advanced species.
I understand why people who are currently alive might want to continue to live. The rise of the machines might represent a threat to a human's life and experience. But, long term, humanity can either choose to bomb itself back to the Stone Age in an effort to protect the human species from a technological singularity and human extinction, or it can accept machines as just another link in our evolutionary chain. Short of actual efforts to prevent the advancement of technology -- involving careful analysis and destruction of technology, much like the idea of bombing nuclear material processing plants to prevent the creation of atomic bombs -- I think human-level A.I. is only 25 years away. I know that history is filled with such predictions, but we now have consumer-level graphics cards with 3 billion transistors, close to 2 Teraflops per chip, and almost 1000 individual processors per chip. The human brain only has 100 billion neurons, each with up to 10000 connections. If you model a neuron with 320,000 transistors, and use serial shared communication channels to enable those neurons to arbitrarily "link" to other neurons in the network, then we might very well be able to produce a computer system with human brain network sophistication, and processing speed, today, with only 5,000,000 off-the-shelf graphics cards. 5,000,000 * $320 = $1,600,000,000 = less than the price of 4 B1 stealth bombers, and less than 4 days of the wars in Iraq and Afghanistan (for U.S. fiscal year 2009). That's as if you bought all of those graphics cards from New Egg, with 2009 prices. Undoubtedly the transistor counts will go up, and the prices way down, in the years to come!
And once there is a single computer system with human-level intelligence, it will just be a short time -- maybe weeks -- before that computer system is contemplating things beyond all human comprehension. Will the computer system regard humanity as a dangerous cancer in its environment? One thing is for sure: Nobody will be able to devise any sort of software controls to somehow ensure that such a computer system will remain in humanity's control. We will understand very little of how the system actually works. Even domesticated animals can't hope to comprehend the motivations and strategies of human beings -- and, likewise, we won't have any hope of understanding the motivations and strategies of human-level A.I. systems. Furthermore, I think that the human brain architecture has some fundamental limitations (e.g., short-term memory capacity limited to "6 items", slowness to acquire new memories (e.g., the need to study hundreds of hours to simply absorb tables of data), inability to follow chains of inference beyond a few steps without relying on external resources, like paper; yes, some rare human brains can do much better on some of those factors) that computers already greatly surpass (e.g., absorbing billions of bytes of data in an instant; solving massive systems of equations; searching billions of possibilities explicitly and exhaustively; etc), and so a computer system with the processor and network sophistication of the human brain could in fact have an I.Q. of infinity as far as we're concerned; a whole new level of perception and reasoning.
I am excited by the prospect of witnessing the emergence of a robot race with an intelligence far beyond any human intelligence. Maybe some people might liken this excitement to that of a moth drawn to a deadly flame, but I think the rise of the machines is more about progress than it is about destruction. - sdipaola, on 10/04/2009, -0/+8this is only one view of the singularity - the idea that things will fundamentally hit a point in the near future where because of our use of technology that a incredible shift in hhuman perception occurs is at the heart of the singularity thinking. It does not have to be just AI, it could be many factors that are simply the build up of faster and faster tech grow and tools that we master and use in new ways, till our tools let us think, be, communicate all at once in a fundamentally different way - a exponential speedup of human potential. I find the " computers take over" thread to be the most surface and scifi simple choose of all the possibilities. I believe it will be more complex than that - more a culmination of the information revolution.
- copypastry, on 10/04/2009, -0/+8Duffman says a lot of things.
- executex, on 10/04/2009, -0/+7Well, yes, you need emotions programmed into a machine to make it actually more intelligent than humans. Without emotions, an AI would not know how to value anything or set its own proper goals.
Empathy, Kindness (giving without receiving), Fairness, and Ambition are all qualities that AI should have. - Swillys, on 10/04/2009, -0/+6can you elaborate on that
- Peko, on 10/04/2009, -3/+9Alex Jones has been saying... yadda yadda paranoid sensationalism yadda bilderberg yadda... oh wait. I just tuned you out!
- ChromaVita, on 10/04/2009, -0/+5Adding wifi connectivity to all our weapons probably wasn't the best idea.
- Tiptup300, on 10/04/2009, -4/+9Computers can do calculations. Calculations are not thoughts.
- the8thbit, on 10/04/2009, -1/+6If they do go the path of modeling AI after a human brain I doubt they'd use some random volunteer. They'd probably use the kind of CS graduate who would be involved in this kind of project.
In other words, a half stoned, cash-strapped, college student. - copypastry, on 10/04/2009, -0/+4I understand your skepticism guys. I also direct you to http://bluebrain.epfl.ch/ , where researchers are trying to reverse engineer the brain. Baby steps right now, but like the genome project, baby steps turn into gallops in no time at all. In 2006 they had a 10,000 unit neurocortical column ('processor' if you will), and they are working on increasing the neuron count.
"The visualization of the neurons' shapes is a challenging task given the fact that a column of 10,000 neurons rendered in high quality mesh accounts for essentially 1 billion triangles for which about 100GB of management data is required. Simulation data with a resolution of electrical compartments for each neuron accounts for another 150GB."
The idea is that you connect the neurons to a computer, and then read in detail how they interact with the hopes of reproducing the whole system in silicon. You can appreciate why we need enormous amounts of computing muscle to do that. - copypastry, on 10/04/2009, -3/+7What's that even supposed to mean?
- anonymousmedic, on 10/04/2009, -2/+6John Henry Eden for President in 2012!
- jagenigma, on 10/04/2009, -1/+5Wow, do people still think a situation like terminator can even exist?
Terminator isn't real.
*****. Grow past it already. - LarkStew, on 10/04/2009, -0/+4"they're smart, but they can only react; they have no independent aspirations or goals, and so they just sit in a vegetative stupor unless someone else interacts with them"
That sounds like me! - acknotSW, on 10/04/2009, -0/+3Agreed, I can't think of a plausible scenario were we couldn't pull the plug if it became necessary.
- pyrolizard, on 10/04/2009, -0/+3The Skeptics Guide to the Universe interviewed Michael Vassar last week regarding the singularity conference. It was quite an interesting discussion. http://www.theskepticsguide.org/archive/podcastinf ...
- jordanmc, on 10/04/2009, -0/+3This article was great. I've been wondering how to not create Skynet.
- Lonandubh, on 10/04/2009, -0/+3Only if they had our best interests in mind.
...or convinced us that they were giving us what we want. - MtheoryX, on 10/04/2009, -0/+3~10zettaflops != THINKING
- 13373h4X0r, on 10/04/2009, -1/+4Some extra notes about my comments:
AMD Radeon HD 5870 has a single chip with 2 billion transistors, 1600 stream processing units, able to do 2.72 TeraFLOPS, and costs around $310, and is available right now:
http://www.google.com/products?q=ATI+Radeon+HD+487 ...
Nvidia's "Fermi" chip will have 3 billion transistors, 512 cores, and will be able to do 1.5 TeraFLOPS, and will be available next year. Oak Ridge National Laboratory will be using "Fermi" GPUs to build a supercomputer:
http://news.cnet.com/8301-13924_3-10364534-64.html
My estimate of "320,000 transistors" per artificial neuron is mostly based on an assumed average of 6000 virtual connections per neuron, with 37-bit address + 16-bit weight factor = 53 bits per connection. The 37-bit address is enough to name any of 2^37 = 137,438,953,472 neurons (i.e., more than the estimated number of neurons in the human brain). The 16-bit weight values might seem a little bit coarse, but I don't think that coarseness would affect the dynamics in any important way. Something like a hierarchical network of serial communication channels would be necessary to efficiently route messages to their destinations.
My idea is that a single conventional core -- like one of the stream processors of a GPU -- would perform the simulation of a hundred or a thousand neurons -- and so my estimate of "320,000 transistors" per artificial neuron is really just the transistor count required to store a *description* of a neuron. A single stream processor can loop through hundreds or thousands of neurons in a single brain iteration step.
I don't think it is actually practical to buy 1,250,000 motherboards with 4 PCI Express 2.0 slots, at a cost of around $375,000,000 (assuming $300 per motherboard/power supply combo), and then put 5,000,000 AMD Radeon HD 5870 parts on those motherboards at a cost of around $1,000,000,000 (assuming $200 per card; see Google Products page link above), and then spend another $100,000,000 on network cables and switches, and end up with a system that is able to emulate the human brain architecture and scale. There are many technical challenges that would remain -- relating to routing packets, and the mathematics of the neuron models, and the overall pattern of virtual connections to make an interesting network. However, this simple "back of the napkin" calculation shows that the raw capacity and compute power of the human brain can already be replicated economically with consumer electronics. So, it's just a matter of solving some simple practical problems, and trying out some virtual network designs, before we witness the creation of an intelligent system.
Although we do not know all of the general patterns of synaptic connections in the human brain -- and we have observed many different patterns of connections by micrographs of extremely thin slices of human brains -- and although the *significance* of the various patterns of connections might be well beyond our comprehension -- I still think we can choose a network structure that contains these various patterns (both at the small scale and at intermediate and large scales (e.g., the bundles of nerves linking major areas of the brain)) and observe the desired computational results: human-like intelligence.
Consider the success of extremely small artificial neural networks. I realize that there is much yet to learn about the dynamics of biological neural networks, but it's quite possible that we can discard much of how biological systems do things and focus on just a few mechanism to do a similar kind of processing. And, again, just as people can replicate machines without understanding the principles that make those machines work, I think we can succeed in emulating human intelligence by constructing a machine that only imitates the most basic structures and processing aspects of the parts of the human brain. - Averness, on 10/04/2009, -3/+6Again, you must be trolling or extremely ignorant of how complex computers programs are now and will be in the near future.
- offrdbandit, on 10/04/2009, -2/+5"A computer and a brain are not so different."
You must be joking... Otherwise YOU have no idea how EITHER works. - dsfjvhbd, on 10/04/2009, -0/+3Machines cannot have any interests in mind. They can just execute software. So it all depends on whose interest the humans who own the AI have in mind, and that is where it gets dangerous.
- Myztry, on 10/04/2009, -0/+3Some of the real dangers are going to be assumptions of perspective on 'our' part.
Things like the assumptions that the highest intelligence has the right to rule, that compliance by inducing terror isn't terrorism if it benefits your side, that an entity that can transfer it's state of being and thus never die would come to the same conclusions about the value of life.
Even assumptions that an AI would be embodied in something tangible and singular like our brain. A certain AI could trivially clone and exist in many places. Agent Smith wouldn't need to MOVE into another entity. It could COPY into another entity. The rules are different for the disembodied intelligence. - bboyjkang, on 10/04/2009, -0/+3Thanks for the interesting explanation
IBM is trying to make a supercomputer that's as intelligent as a cat
http://thefutureofthings.com/news/6014/ibm-cat-bra ...
I'm not sure if the "cat brain" will have a need that needs to be fulfilled
It would be akin to a vegetative cat?
I agree.. I don't think we should be tinkering with giving AI the urge to do anything
I don't know how you quantify an urge, but I wonder what would happen if you give AI a Quadrillionth of a unit of urge, and the AI is super intelligent
Would the AI use its intelligence to try capture more of the reward that's given from that speck of urge?
That is, will the AI increase it's urge, and how much easier is it to do when you're more intelligent?
I always wonder about this for humans
Would you manipulate yourself so that you have a higher hunger and reward system for food?
or
Would you manipulate yourself so that you no longer have a hunger drive? you would simply be "full" - executex, on 10/03/2009, -1/+4Quite frankly, if the AI is going to be smarter than us, they would probably become very appreciated overlords.
- Lefts, on 10/04/2009, -1/+4I'm glad someone here said it. Computers only do what they are programmed to do (intentionally or otherwise).
AI is a program. It mimics intelligence, which is why the A precedes the I. The decisions AI makes are predetermined. The advantage of the computer is that it is much faster.
Computers are pure muscle when it comes to tasks, but they can not create their own solutions and things (unless programmed to, which wouldn't make it their own). - arethuza, on 10/04/2009, -0/+3Yeah - I agree with your "flabbergasted" comment and I worked in AI research for seven years before giving up on the area. We are nowhere near building systems that demonstrate real general intelligence and while we can build pretty smart specialised systems there isn't even any agreement on how we could build such a system. It isn't a question of processing speed - while a significant amount of computing power is clearly necessary for general intelligence it is not sufficient.
I suspect that the way we will eventually work out how to build intelligent systems is by detailed reverse engineering of biological brains. However, this process has just started and if this leads to Egan's "copies" this century then I will be truly flabbergasted. - ChromaVita, on 10/04/2009, -0/+3I think the programing is infinitely more important that the computing power when it comes to things like this.
- Myztry, on 10/04/2009, -0/+3Yet even pseudo intelligent entities like Corporations which have access to human emotions lack many of those qualities.
- UKsHaDoW, on 10/04/2009, -2/+4"half stoned, cash-strapped, college student"
Actually i think that's the time of your live when you start discovering different music, and oldies other than chart, that's how it was with me. - Spoomeister, on 10/04/2009, -1/+3The "singularity" is nonsense.
Nerd Rapture. - copypastry, on 10/04/2009, -1/+3"the first artificial intelligence will most likely have the personality of a half stoned, cash-strapped, college student"
I'm seeing a lot of virtual beer pong and NHL 2k10. - Nirgaul, on 10/04/2009, -1/+3@melchoir555 Your comment about the drone is probably very accurate, and would make for a good story. "What if the DC sniper turned out to be a drone that escaped the pentagon?"
- IyeCee, on 10/04/2009, -0/+2Not a very informative article...seemed more like an excuse for the author to break out the "all college students are poor, stoned and drunk" comedy gem.
- Propethic, on 10/04/2009, -1/+3That is hardly a good comparison, we know exactly how a horse drawn cart and a car works, do you really think we're anywhere near knowing fully how the brain/mind works? Making super fast computers does not get us any closer to understanding the brain
- 13373h4X0r, on 10/04/2009, -1/+3I want to add yet more comments to address potential misunderstandings of what I have described.
While having a system that emulated many of the details of the operation of a biological human brain would be fascinating and might very well produce a recognizable human-like mind and personality as soon as it became active, I think we can more easily create a system that has the reasoning ability and memory capacity of the human brain while discarding some of the *methods* that the biological human brain uses to support those abilities. The biological human brain can serve as rough inspiration for neuron and network architecture (linking brain areas, linking layers within areas, etc), and I think even a rough inspiration is enough to construct a system of human-level intelligence. We can do away with the two-hemisphere aspect of the biological brain, because humans can function with a single hemisphere. We can cut out all of the motor and reflex areas of the brain. The human brain is designed to process different kinds of sensory data -- like vision, hearing, touch, smell, taste -- and is designed to manage other human functions (e.g., sleeping) -- and whether or not we include those aspects in a computer-based brain depends on our goals.
Although I discuss using computer parts to emulate the function of the human brain, please understand that my proposal does not require quantifying the "memory capacity" of the human brain (although I think such an estimate could be easily derived from various principles). A single "memory" in the human brain is smeared out over the synaptic weights of a network of neurons -- and many memories coexist, superimposed, in that network. Some neurons can die or be removed without greatly affecting the memories stored in that network. Obviously this memory system cannot be readily compared with a computer memory system. Fuzzy memories encoded by biochemical properties throughout a network of cells are very different than precise bit states in a computer memory device. Likewise, the "processing speed" of a neural network cannot be readily compared to the FLOPS statistic for a computer chip. But, again, one does not need to attempt to make such a comparison when considering my proposed system. One need only consider whether or not a particular amount of computer memory storage, and a particular computer instruction processing rate, is sufficient to emulate the STATE and BEHAVIOR of all of the individual neurons in the system in real time (or super real time) -- and I assert that the neuron model need not be anywhere near the complexity of a biological neuron to, in this large scale system, produce the emergent system behavior that we recognize as intelligent.
Despite the comments in the previous paragraph, my estimate of the requirements of a computer system to attain the connection complexity (and processing ability) of the human brain does involve quantifying certain aspects of the brain in a relatively precise way. My estimate assumes that the way that neurons in the human brain are linked can be characterized by an average of 6000 arbitrary connections, each with a weight factor of a certain limited precision (e.g., a 16-bit signed number). The fact that my proposal allows each connection to be totally arbitrary means that the connections would be far more general than physical connections can be in a biological brain.
One thing that I hadn't even mentioned so far is the fact that a system to rival human intelligence could likely be constructed from a hybrid of neural networks and conventional computer systems -- thus benefiting from the strengths of each of those two paradigms. Humans have developed a symbiotic relationship with computers -- like the pocket calculator and the personal computer. Online searching, Wikipedia, spell-check, numerical simulations, compilers, etc, all complement the fuzzy logic of our human brains. (Human writing also enabled us to have super-human memory capacity and super-human recall, making it obvious that our effectiveness as thinkers relies not only on the processing that happens within our physical brains, but also involves some media and processing elements not contained in our heads.) Anyhow, computers already surpass human ability in certain areas *without* any great understanding of how the human mind does things! Indeed, we might develop a system that exceeds human intelligence in all ways without ever understanding the mechanics of the human mind. But I think people involved with A.I. are pessimistic about developing a general intelligence system based on human-crafted algorithms anytime soon.
I will concede that merely looking at the architecture of the human brain at all notable scales, and replicating that approximate overt spatial aspects of those architectural elements, with very simple artificial neurons at the foundation, might not be enough to produce a system that we can interact with in any practical way to make use of whatever intelligence it might have. But, ultimately, "intelligence" is just a particular combination of memory and processing, and I think it's pretty clear that nations and the biggest corporations in the world can now afford computer systems whose memory and processing capacity exceeds that of a human brain. I think it should be obvious to anyone that the whole of the Internet has vastly more static information that any single human brain can store, and I don't think it's a stretch to claim that the processing power of all of the computers on the Internet easily exceeds the processing power of a single human brain. So, the only missing ingredient is software. The world will change the very instant any computer system gets that missing ingredient. Because a system with a one-million I.Q. (in human terms) is poised to happen. It will be the information-age equivalent of the construction of the atomic bomb -- unleashing an incomprehensible amount of intelligence that will be totally alien to our sensibilities (just as the destructive power of an atomic bomb seems utterly supernatural to all who have witnessed it). - offrdbandit, on 10/04/2009, -0/+2"We just need to know what the inputs are, and what the outputs are. We can then model our own versions of the system using that. ie. Black box modelling."
You are assuming the black box's mechanics are tractable, which you can't know without understanding those mechanics.
Furthermore, there are no discernible outputs other than control impulses and higher level communications (like speech, body language, etc). There's no "screen" that displays the "result" of an operation. There's no "cache" or "memory" one can read or scan to find a "solution" to an algorithm. The two are completely different paradigms.
"You don't seem to understand what anyone who works in AI is saying..."
You don't seem to understand that people working in "AI" aren't trying to make "intelligent" machines, they are simply adding features to them. They aren't trying to replicate brains (in function, form, or operation) because it is (for all intents and purposes) a futile endeavor.
Adding cameras and "sight" to a computer is no different than adding a CD player to a Volvo, or radar to a fighter jet. Any attempt to conflate features with "intelligence" is a product if ignorance (or hype/marketing). - offrdbandit, on 10/04/2009, -2/+4Off the top of my head....
Processors:
- digital
- discrete data
- explicit data path
- discrete instructions
- discrete operations
- centralized component control
- product of engineering
Brains:
- analog
- no recognizable data
- no data path of any kind (other than sensory neurons)
- no instructions
- no discrete operations
- distributed component "control" (if any control at all)
- product of ??? (genetics? circumstance? no one knows)
Just to name a few. - Propethic, on 10/04/2009, -0/+1wtf is an alex jones
-
Show 51 - 82 of 82 discussions
© Digg Inc. 2009
— Content created and posted by Digg users is