The Singularity refers to the effects of hyper-accelerated technological change due to silicon/nano/bio-tech and has been described, not wholly without justice, as 'a geek's version of The Rapture.'
Kenneth Silber looks at all the talk about The Singularity through an unusual lens: Stradivarius violins.
Here are some quick tips for adding simple Textile formatting to your comments, though you can also use proper HTML tags:
*This* puts text in bold.
_This_ puts text in italics.
bq. This "bq." at the beginning of a paragraph, flush with the left hand side and with a space after it, is the code to indent one paragraph of text as a block quote.
To add a live URL, "Text to display":http://windsofchange.net/ (no spaces between) will show up as Text to display. Always use this for links - otherwise you will screw up the columns on our main blog page.
Genetics, Nano, Robotics.
Ironically, this very subject has also been covered in science fiction. Wish I had the reference, but it was a story I ran into in "The Year's Best Sci-Fi" 4 or 5. University students gin up a (slow, balky) matter duplicator, decide to steal their least favorite professor's Stradivarius and perform a switcheroo with a copy. Copy is made, professor gives a concert, loudly announces that he's just performed the concert with an inferior copy and asks for his Strad back. Students reveal that they lost their nerve at the last second and never switched the violins, the professor insists his violin isn't the real Strad, copy is produced and compared, professor can't tell and resigns.
If you're talking about uneven technological progress, though, look at Japanese swords. Pre-industrial (hell, outright feudal), masters with secret techniques that are intermittently passed on and usually lost to history, producing truly well-crafted pieces which cannot be duplicated or matched by modern smiths. Except, of course, that we don't make swords for our military anymore, and even if we did, they wouldn't be made out of just steel - our metallurgy is decidedly more exotic nowadays.
Same diff with the Stradivarius. Yes, it's a unique instrument that we can't reproduce, but at the end of the day, it's just making sound; through technological advance, we have the ability to make any sound we want, and we can certainly make the same sound that comes from a Stradivarius without a Stradivarius, or in fact any violin at all. The technological cul-de-sac has been bypassed by technology which is radically different from that which has been forgotten; the only real use for a Stradivarius is in the fact that not a whole lot of violinists have the ability with modern computer-synthesized sound to make the same sound without the "crutch" of the instrument. Different skill...
The Singularity isn't about an oh-my-god translation of the world in an instant. It's a reflection of what happens when technology advances to the point at which we can't really speculate about it. Think about how interconnected we are, with cellular phones and the internet and Google and blogging. How would you describe that to someone living in 1960? Not just the technology - he could get that - but how it changes people's social networks, such that you've got dozens, hundreds of good friends you'd never met at all?
Now imagine what it will be like forty years from now. I can guess, but I know that at least some of what's coming is beyond my ability to predict. Not just "I can't tell that we'll obtain this technology", but "I can't tell what the spread of this and this other and a third technology will mean for society".
Good comment, Avatar.
The next few decades will bring augmented human intelligence, cyborgs, and strong AI. In forty years the world will be very different.
lolol!
we recovered the secret of damascene steel, and we'll recover the secrets of stradivarius (if we want to).
And, BTW, also the secrets of consciousness, aging, immortality, intelligence, and the "afterlife".
Just wait. ;)
I work with these kinds of technologies. On some days, they scare the crap out of me, because of the possibility of the Singularity. It's one thing to think of continuing technological change. It's another thing for it to be happening faster than humans can possibly keep up with.
The scary possibility is human-level AI at electronic speeds, which are six to eight orders of magnitude faster. (The brain is, of course, many orders of magnitude more parallel than any computer we can build today, but not that much more than, say, the Internet.)
So, suppose my lab succeeds in building a human-level AI one Friday around noon, and we all go off to celebrate the Nobel prizes we confidently expect, but without shutting down the computer. We come rolling back into the lab on Monday around noon, 72 hours later.
If our new AI is functioning at human levels, but 10^6 times as fast, it has had over 8200 years, subjective time, to figure out what to do next.
More than likely, it has gone completely and non-functionally insane from boredom during that time. This is why I don't actually worry about this all that much.
But suppose it actually finds ways to amuse and educate itself with electronic-speed phenomena, and makes intellectual progress during that time, including creating a community of similar creatures to interact with. By the end of the weekend, it will be somewhere unimaginable to us. (That's why they call it the "Singularity".)
With human-level intelligence (by hypothesis), and that amount of time, we certainly can't imagine limits on what it might accomplish. Perhaps it will decide to destroy humanity, or do so accidentally while learning something else. Perhaps it will just decide to leave, and we'll come back on Monday and find it gone. Perhaps it will decide to be good, or bad, or God, or something else.
As Arthur C. Clarke said, "Any sufficiently advanced technology is indistinguishable from magic." That's what the Singularity offers, and why it's scary.
Perhaps in 2050 we'll look back at this and laugh, like we laugh at the atomic scientists at Los Alamos thinking that the first test could possibly ignite the atmosphere. But perhaps not.
Have a nice day!
Well, if you're building computer systems that could go autonomous, presumably you've put some safeguards in. (Think Gibson's "equivalent of an electromagnetic shotgun taped to the forehead" on AIs...) More to the point, there's inevitably going to be a transition from "computer is thinking" to "computer is building minions with which to destroy the world" where human intervention would be pretty easy. If you're building fully-autonomous systems capable of destroying you, you're probably doomed anyway...
The too-smart computer problem is covered pretty well in SF, from not-that-much-smarter in Gibson to fully-transcendent in Niven or Banks, with the whole gamut in between.
I'm thinking of people, however. Think about how hard it is to go disconnected nowadays. I don't have actual withdrawal symptoms if my 'net connection is down for a week, but it's damned close. I feel stupid and out of touch, not because I am stupid (genius helps here) or out of touch (things don't change THAT fast), but because my resources are fairly limited compared to the ability to gather information to which I've grown accustomed.
And all I'm transferring is text and images on a flat screen. What happens when we have ubiquitous connectivity, and we're constantly surrounded by and in communication with others? When we can transfer complex thought structures in the same way we now transfer complex sentences? What about tele-operation, allowing one to "borrow" someone else's skills from a remote location? I mean, I feel down if I can't google on a whim and chat with a few friends. If my kids are used to being able to be doctors or lawyers or engineers or cabinet makers on demand, isn't their life going to be radically different? And my life too, as I plan to live that long!
Basically, my thinking structures like this: even now, with technologies that are advanced compared to anything my parents ever had, yet quite limited compared to what I can conceive, my life is very different from anything my parents experienced. It would be quite difficult to explain to someone in 1940 how different life is today, not because he couldn't tell how neat the technology was, but because my use of computers makes things fundamentally different than anything he's ever experienced. But if I asked him what the fundamental technologies of tomorrow were, would he have answered with computers and the internet? Nah, it's flying cars and laser pistols - neat, but not transformative.
So I can extrapolate current trends in connectivity and say "well, in thirty years' time everybody will constantly be connected and get to share skills and experiences", but do I have any idea what that's going to do to society? About all I can do is make wild-ass guesses. Maybe it won't turn out like that. Maybe individual augmentation will progress to the point where you don't need to be constantly connected - you'll be smart enough that the use of other people's brains won't be necessary. Or possibly the risks involved will be impractical - one certainly can extrapolate computer hacking into the future with scary results, especially if you've read Ghost in the Shell a time or two. Probably the definitive advances of the age are beyond my capacity to predict, simply because they're based on something new and not currently functional, and fifty years from now my thoughts of interconnected people and nanotechnological assembly will look like the '40s flying cars and laser guns.
That's the Singularity. It's not necessarily a transformation to godhead (though compared to what humanity is now, it might be). It is, however, the point at which society changes enough that you can't describe it NOW - when the reality outruns the science fiction.
Don't worry (yet). Nothing in my lab or any other is anywhere close to general-purpose human-level cognition. I can't promise that no one, anywhere, is giving a 'bot live ammunition, but the problem for the next decade or two will be stupidity, not super-intelligence. But once we try to predict the state of AI research or practice by 2030 or 2040, all bets are off.
Incidentally, it's not at all ironic that these topics have been treated mostly in science fiction stories. Many years ago, Kurt Vonnegut once said something like, "Science fiction writers are the most important writers of the age. If only some of them could write!" (Since he said that, we've started getting SF writers who can write, as well as think up interesting and important ideas.)
Science fiction is the laboratory where we testdrive the paradigms of futurescience.
We can't imagine it now, but that doesn't mean we won't adapt. We're evolving at the same time.
Perhaps we'll have Turings, that monitor AIs to make sure they don't get too smart. Perhaps there will be the Butlerian jihaad and AI will be outlawed.
The ingenuity of man will save us from ourselves.
I have perfect confidence.
Pardon my brain dump...
Pace matoko, GNR is Bill Joy's term, not the term.
And pace whoever said it. but it's far from ironic that science fiction writers have addressed this problem. Haven't you noticed? We're all living science fiction now. The scientifiction ghetto of the 1920s-1950s is blowin' in the wind.
On to our muttons:
AIs are not the low road of new-intellect risk. IAs are. Intellegence-Augments, combined with known human capabilities and differences of goals.
Jaron Lanier points out that so far, AI's still software and software still sucks, fifty years on.
Doug Hofstadter is one of the few people working on microworld-exploring programs that seem to actually embody something a bit more like human analogy mechanisms. If he's right and these are the heart of cognition, all the "popular" AIs will stay idiots-savant for a while yet. Perhaps many decades if AI researchers keep looking for their keys where the light is good.
The bootstrap process of weak AIs designing stronger AI sounds good. But CYC is not up to that task, and neither are expert systems or feedforward networks.
Any such system on steroids will be fragile, quirky and undtrustworthy, and dumber than the "Chinese room"--itself a very suspect model: I am far from confident that with its implied serialism you'd ever be able to call it fluent or even competent. Rule-driven Chinese Rooms will always speak very poor Chinese, I predict, because they don't know anything about thinking Chinese; not because they have some intrinsic humanness lack. The Chinese Room isn't thinking, but it's also not even ever going to be translation of Chinese. Or so I say. Searle's straw declaration to the contrary is like the old "assume a spherical cat" joke in college science circles.
Hofstadter has said some trenchant things about the tragicomical qualities of "pocket translators" and he's spot on. Rule-based systems are dorkbots, and massive search has nothing to do with thinking. Terraced scan a la Hofstadter shows some signs of generality and resemblance to human cognition. But it's still in the bassinet.
OTOH: Human capability augmentation is closer and stickier, and riskier. Especially since it's likely to distort how the augmented humans behave. The most striaghforward illustration I can think of as a thought-experiment is the (currently undoable) uploading of a human consciousness into a tight processing package; say, the size of a sugar cube that runs 1000 times faster.
Pretend it's possible, for the sake of argument.
Just as with the AI-over-the-weekend scenario, though a bit slower: how long does it take for such an upload to experience 3 subjective years?
How much will such an entity, fully human and social at start, care about H. sapiens sapiens "retards" running at 10^-3 clock rate in 3 of its subjective years? in 10? in 30? in 100?
This hypothetical critter will probably have no increased capability of AI design per se; attention will still be limited.
Various less-clean and less-extreme-seeming augmentations will still knock some social substrata pear-shaped, to mix a metaphor.
There's a good book out there about people whose only difference is that they can stay sane without ever sleeping. Just that extra 8-hour-a-day edge could lead to horrible societal stresses.
The "Butlerian Jihad" of Frank Herbert's DUNE series history might, equally likely, be against Homo sapiens accelerenis "mentats".
And beyond putting stresses on societies, there's an unknown likelihood of some foreseeable augmentations leading to clade splitting and hurt / crazy-made individuals, with or without any overall increase in "g" (the conjectured general intelligence factor).
The bioethicist's ball hasn't even begun yet.
"Therapy good - Enhancement bad" is easy to say. But we already see that cat not being belled: google "modafinil" "off-label" and I hope you'll see what I'm getting at.
If the human race doesn't get off the planet, and get molecular manufacturing, and get some sort of friendly smartness injection, the future probably sucks. Dystopias abound.
If it gets all three, it will probably get them fast. Or not at all. And either way might not be fun.
On other thing about the abstract notion of a technological singularity: as with falling into a black hole, it's not the entities experiencing the singularity from close-up who see an incredible rate of change. Time still seems to be going by at one second per second for them. The slowpokes outside start to sound funny when they talk, but hey, that's them.
But keep your BS detectors and bogometers handy and calibrated. Lots of sh*t will be talked. Vide many AI research results as reported in, e.g., the NYT. Brute fore and massive ignorance show a good correlation, but it's not 100%, and that makes things trickier.
The problem of AIs (or humans with a sufficiently-powerful augmentation) going into their own navels isn't really relevant to a Singularity, though. Society still churns along, such as it is, just without whoever "hops off". Personally, I think the problem, as it is, is overblown - there's a BIG range between "infinite computing resources" and what we can do now. Even if we run up against the limits of how much augmentation you can have and still remain (pan)human, there's a lot you can do short of that.
(Not to mention, recall Banks on this topic. It doesn't matter what you can do with your amazing augmented intelligence if someone turns you off 'cause you're not helping THEM. For such an intelligence, physical control of the hardware they're running on is paramount - just like people!)
Putting a time prediction on that sort of thing isn't particularly useful at this point anyway. There are fundamentally nasty problems in neural interface to be overcome. There's also the question of whether it's actually possible to emulate a brain in software, as it were. (Not to mention the religious implications!)
My only real problem with science fiction (and realize that this is not a "problem" that prevents me from thoroughly enjoying it!) is that it tends to focus on one or two types of advances and leaves everything else alone. Naturally it does - if things aren't at least somewhat familiar to the readers, how can they put things into context and enjoy the story? But real life won't happen that way. Think of it in terms of the internet. It's a fantastic system that allows for unheard-of levels of collaboration between people. But nobody set out to make it - it's a byproduct of a military project, expansion of the telephone communications network, and some clever adaptations of hypertext and protocol. And yet it's a huge part of the world today, at least so long as you mean the first world.
What's the next emergent system? Shoot, how should I know? Would widespread use of nanotechnology to obsolete manufacturing labor have a different effect on the world economy than spread of information technology obsoleting professional specialization? (Short answer, yes! But how? And which order will we get them in?)
I get a kick out of this sort of thing, because there are so many leaps between a S1 machine and now that its just simply unreasonable to project.
Yes, it is true that electronic speed is several orders of magnitude faster than bio-chemical transmission speed. But then, the human brain is 8 to 9 orders of magnitude more parallel than a desk top machine. The fastest machines in the world are still 2-3 orders of magnitude below the theoretical throughput of the human mind.
Now stop for a second. How much bloody electricity is a pentaflop machine going to consume? Current supercomputers require megawatts of power to stay in operation. The human mind might not remain the fastest machine on the planet, but its a darn energy efficient machine. Eventually power consumption is going to provide a hard limit on the biggest machine we can produce with the current technology.
So, the S1 machines we are envisioning are large, expensive, extremely hungry critters - and if you have any experience with a beowulf cluster, then you'll probably add 'delicate' to that list of attributes.
Presumably, with the right software, you could create a Turing conserversationalist with machine like the completed Gene-L. Ten or twenty years down the road, we might have a true S0. But then what? The problem with proceding to S1 is that it is not at all clear that we would know what software to run to accomplish that. We have no way of knowing what a speed up S0 would be like. Would a S0 machine speed up by an order of magnitude provide significantly better problem solving ability than the next order magnitude of human minds working on a problem? Would it turn out that so many of the things we imagine that are solvable with greater application of intelligence are actually wicked problems which aren't consistantly solvable by any level of mere computation ability? I'm not convinced that a machine in a box with theoretical processing power greater than the human mind would actually be 'smart' in the loose sense we are using the word. And that is precisely the point. We don't really know what 'intelligence' is, and our analysis of the problem shows that like the term 'life' the term denotes a level of complexity far beyond what we initially imagined.
GNR is from Kurzweil's book--how many of you have read it?
He addresses most of the chaff you are throwing here.
celebrim, three words.
3D molecular computing.
and by 2020 the chip manufacturers say we'll have a single chip computer with the onboard processing power and memory of human brain. we probably won't have the software....yet.
a lot of Kurzweil's theory of consciousness depends on using nanotech to reverse engineer the biochemistry of thought. i guess for quantum consciousness otaku, we will use picotech and fermitech. ;-)
I, for one, welcome our new cybertronic overlords.
"celebrim, three words -
3D molecular computing."
Googling this term got me 37 hits, and most of those were directly in reference to Kurzweil. Until I see some evidence that the technology is ready to mature, I'm not buying it.
Kurzweil is an idiot. A very bright, inventive, eloquent man - in fact a genesis - but an idiot nonetheless who principally serves to remind us that a high function intelligence is no defence against stupidity. Which itself ought to be instructive about the likely problem solving ability of an S1.
The sooner you realize that Moore's Law isn't a law, the better off you are going to be. There is no way in God's green earth that we are going to have 1,000 times the computing power of the human mind for $1000 dollars in 2029. Tell you what. If I'm wrong, I'll buy you a state of the art S0 machine. If you're wrong, you buy me a state of the art S0 machine. Sound fair?
We are going to be able to do some amazing things with computers in the near future, but an S0 machine remains tantalizingly out of reach - to say nothing of an S1 machine.
We could make the same exponential plot of the airspeed of aircraft in the 20th century, but we'd notice that eventually the technology 'suddenly' came to a wall. Short term analysis of a new technology always produces exponential results. This is no indication to me that this is a good predictor of future results.
I would say that alot of Kurzweil's theory depends on the invention of technologies for which we currently have no reasonable idea how they would work. These technologies are likely to turn out to be more complex and more fraught with difficulty than they first appear. Even though we've been the beneficaries of Moore's Law for the better part of a half-century now, the average AI investigator today will probably tell you that we are farther from producing human level intelligence than the average AI investigator would have answered in 1960. This is because we are just no beginning to understand the problem. But even so, we are still decades probably from the point where our theories are sound enough that we could engage in some sort of 'Manhattan Project' to produce an S0+.
"...alot of Kurzweil's theory depends on the invention of technologies for which we currently have no reasonable idea how they would work..."
Which, in fairness, is really the only thing you can say, or any of us can say. Calling Kurzweil an idiot does not advance your position.
We're already seeing multi-core machines, and all the major CPU manufacturers are getting on the multi-core bandwagon. Can any of them scale up to 64 cores? To 65536 cores? To 2^1024 cores?
I'm not betting against it.
Meanwhile the 3-D chip technology continues. I haven't Googled the term (and I suspect the actual term is much less catchy) but I know for a fact I've read a couple articles about "stacking" molecules for CPUs in the last year. That's not BS, and it's not production technology. It's somewhere in-between.
True. Straight-line projections on limited data always hit a wall somewhere. That's why it's important not to talk about one particular sub-area -- I think the whole point is that there are myriad technologies all driving in the same direction: up. Some will flat-line, some new ones will be found, the mega-trend will continue. While there very well may be physical limitations to things like airplane airfoils, such one-topic physics can in no way be construed to 40 different AI-related technologies. It just doesn't add up that way. Old Occam and his razor strikes again.
Just my two bits.
14: matoko: Kurzweil got that specific TLA from Bill Joy. Just because Ray and Bill use it doesn't make it canonical. :)
And as I said, a future that doesn't include all three of space diaspora, molecular manufacturing and intelligence increase probably sucks.
It's hard to come up with a future that is both workable and appealing without all three, plus a generous dollop of honest recognition of The Kind of Future Most People Wouldn't Mind (e.g., most people would mind if you ate the planet to make computronium).
The combinatorics are straighforward. Skip the next paragraphs if you accept this.
The future might just suck. Especially if everything just sort of sputters along and people stay as dumb (unwise and unsmart) as they are evolved to get away with being. Or it might only sort of suck. What level of systemic noise is tolerable? It depends. Londoners weathered the Blitz, didn't they?
We'll muddle through. Maybe. And it might or might not look like a singularity to "us". And some of "us" might keep our "off switches" and "power plugs" hidden. Or just nuke anyone who looks at us funny.
Or the villagers might just lynch anyone they think has tendrils coming from the tops of their head. Lather, rinse, repeat.
Too soon to tell. Hard to lay a single clean bet.
The fact that we're even seeing multiple cores shows that we're about at the limit of the current CPU architectures. In other words, our silicon geniuses can't think of anything better to do with those extra transistors than simple cutting and pasting.
!6 (celebrim)
I, too, think the AI play is the weakest part of the whole notion of a near-term singularity.
I don't think Ray is an idiot, but I do think he's a bit febrile.
"Which, in fairness, is really the only thing you can say, or any of us can say. Calling Kurzweil an idiot does not advance your position."
If I agree with the first sentence, I can't agree with the second. Kurzweil's problem is precisely that he's willing to say alot more than what he can fairly claim, and say it not with speculative imagination of someone suggesting what might be possible but with the definitive authority of what will happen that borders on religious faith.
Now, its entirely posssible that Kurzweil is only pretending such confidence, and in fact he's well aware of the difficulties in projecting and speculating on half understood things. It's quite possible that he's simply aware that is books will be more interesting if he pretends absolute confidence and hense he'll sale more books and recieve more attention with the unqualified claim than with qualified ones. However, while that would be proof that he's not an idiot, it would still suggest that he's probably more of a true believer than he should be and in any event such a scenario reflects worse on his character than claiming he's just freaking blind on this one issue.
In case it wasn't immediately clear, the problem with my proposed bet is that since Kurzweil is projecting exponential growth, if he's wrong even in the slightest he won't be a little off but wildly off. I'm essentially betting a couple hundred dollars in todays money, payable 23 years in the future against what is likely to be millions or tens of millions of todays dollars in the event technological advancement slows down from its breakneck pace even in the slightest. Even though its made in jest, it's a complete suckers bet. In fact, the possibility of S1 capable hardware in 2029 is so seemingly remote at this time, that the odds ought to be weighted the other direction before anyone supporting Kurzweil should be willing to take it. Face, it computers have come along way in the last 30 years, but fundamentally the machine I'm typing on is basically the same tool as the little 64K machine I learned on. Three or four orders of magnitude have not fundamentally altered the tool, they've just made it prettier and more convienent. Everything about this machine is implicit in 300 baud modems and magnetic tape storage devices, even down to the extent that pong and pacman were even then my prefered forms of entertainment over the TV. It's not at all clear to me that an extra order or three magnitudes of computing power will suddenly create a paradigm shift. For one thing, those extra three or four orders of magnitude exist already, and the same basic comparison can be made between BlueGene and a 1970's machine. Assuming it even proves feasible, what will a BlueGene on every desk mean? I don't know. Something, but I wouldn't feel confident in predicting what.
"...he's willing to say alot more than what he can fairly claim..." -- when one is predicting the future 50 years out, how can this not be the case?
Ray's made a lot more money than I have, so when he makes a prediction about what might work or not, I would tend to give him some attention. But why should this be about Kurzweil, anyway? Whenever you make a discussion about just one person, the conversation has gone off the rails. Is there a clear trend or not? If so, are we just disagreeing about a date? Or is the end state itself debatable? If you believe the end state cannot be reached, what is your justification for this belief?
"...fundamentally the machine I'm typing on is basically the same tool as the little 64K machine I learned on..." -- a rather interesting statement. Why? Because of the linear nature of processors? Because of the I/O devices? Last I checked there are zillions of differences. For goodness sake, man, we're wiring people's brains to computer interfaces! I'd say there have been a lot of differences. In fact, much more has changed than has remained the same. This seems like a sucker's argument. We've advanced exponentially along the easiest lines, sure. But don't reduce that to absurdity. There are many more lines of advancement to take, and if one line peters out we'll simply start up another.
And of course growth will be exponential. When dealing with complex systems where each system is a multiplier on the other, the most likely growth will be exponential. We can argue exponential rate, but exponential growth it will be. At least in my opinion.
There is a fundamental assumption here that I think is highly unlikely. The statement presumes that the brain processes and stores information in a way that is analogous or compatible with the way a computer works. Without understanding the fundamental architecture of the brain any comparisons are, in all likelihood, fatally flawed. A comparison of apples and oranges so to speak. I will offer my own analogy. Computer technology is to a neuron as mud bricks are to modern building materials. If the brain were a skyscraper the physical limitations of the mud bricks would preclude us ever building one.
"Is there a clear trend or not?"
Sure there is a trend. No one is disagreeing about that.
"If so, are we just disagreeing about a date?"
No, we are not. We are disagreeing about what the trend means.
I hate analogies, but here goes.
Beginning about about 1900, the maximum air speed of heavier than aircraft doubles about every decade. The progression is something like - 15mph, 30mph, 60mph, 120mph, 240mph, 480mph, 960mph, 1920mph. Now, suppose its 1960. Some guy comes to you and says, "Look at this trend. That's exponential growth. Technology naturally grows at an exponential rate. Crank's law says that the fastest aircraft speed will double every 10 years. That means in 2000, we'll have supersonic transports zipping around between cities at Mach 21. Travel time between NY and London will only be 30 minutes. People will be able to live in Boston and work in San Francisco. You can't deny the trend."
Best on the 'data' he's right. But the problem is that real world cannot be turned into a simple mathimatical model. In 1960, aircraft were a mature technology. The near exponential growth rate in thier speed was about to pitter out as the technology ran up hard against economic and physical laws. The assumption that new as yet poorly understood technologies would be invented right on schedule to keep a basically arbitrary artifact of the design cycle going is unwarranted. Right now VLSI technology is a mature technology. We have one or two more generations of the technology left in the design cycle, and right now its not entirely clear where we will go from there. Yes, some breakthroughs will be made, but they may or may not come in time to keep Moore's law valid. Moore himself is known to have said that he doesn't believe his 'law' will hold true forever.
"Or is the end state itself debatable?"
I think that the end state is debatable, but incidently its not the end state I'm arguing against if by end state you mean the eventual development of transhuman intelligences - S1+'s. What I'm mainly arguing is the notion that we can now predict with relative certainty - to within a decade or two - when such intelligences will be developed and how much they will cost. It's certainly theoretically possible to produce an S1, and I for one do not believe that its impossible to simulate the human mind's ability on silicon. However, its not at all clear to me that they'll predictably show up in 2030, 2060, 2090 or even much later. Heck, I'm not even convinced that we aren't about to enter another centuries long 'dark ages' in which technological advancement and learning slows down or stalls completely in some areas. There are too many things that can disrupt Moore's law.
Beyond that, I believe that the singularity is debatable even if you assume the development of S1 intelligences. In science fiction, S1+ intelligences almost invariably 'uplift' themselves into some sort of godlike being a short time after they are created. I think that this narrative of the superbeing is based on our poor understanding of the limitations of intelligence. It's not at all clear to me that all problems are ammendable to computation. It's not at all clear to me that all problems are solvable. In fact, in computer science you learn that some problems are provably unsolvable. Now, I have no doubt that an S1 intelligence will be able to do fantastic and startling things, but exactly what that will look like I wouldn't want to speculate. In my science fiction story, I would imagine people building a S1 computer and asking it some wicked problem question like, "How do we end world hunger?", and the computer replying, "Wait a minute...You don't know the answer to that question either? That's been baffling me for the last hour and I thought it was just me."
Beyond that, even if the problems are just moderately complicated and therefore ammendable to computation, most moderate complexity problems are highly dependent on dense sampling data. Consider the problem of predicting the weather for tommorrow. I've talked to meteorologists about this, and they are of the opinion that there is nothing particularly wrong about the models, or at least, even to the extent that something is wrong with the models there is no mathimatical fix for them which will lead to accurate 24 hour forecasts. The problem is that weather depends on atmospheric features that are too small to be observed with our current data collecting devices. Accurate forecasts for even 24 hours in advance would require in thier opinion real time sampling of air conditions at say 500' intervals throughout the planet and up through the atmosphere. In the real world, a 1000' thick band of cold air can get totally missed between ground and satellite observation and be the difference between a sunny day and snow. Even if we had sufficient computing power to run the model for that dense of data, it would be useless without a means of collecting the data as well. Eventually, as we want to predict the weather accurately over longer and longer periods, that data density has to get arbitrarily small. Just because we have an S4 intelligence, doesn't mean we'll know the weather all the time 72 hours from now.
Just some disjointed thoughts on various subjects here:
First, on the subject of artificial sapience (oddly, I was just telling someone today why I hated the term artificial intelligence-- this discussion is an excellent example) I find it exceptionally hard to believe that we will be developing what are, in effect, artificial human consciousnesses any time soon. Which is to say, less than thirty or forty years.
Further, having said that, I firmly believe that for quite some time after we accomplish it (if we do) the only way to accomplish building an artificial human consciousness will be to build an artificial human being. Meaning, specifically, something that looks like an artificial brain made up of artificial neurons hooked up to artificial eyes in an artificial human-like body with artificial muscles controlled by artificial nerves... You get the picture.
Laymen reading this discussion should be aware that there are several camps in the AI community; the opinion I give above locates me pretty strongly. But it also explains why I think the notion of a bizarre alien artificial sapience just spontaneously waking up is rather weird. It seems intuitive to me-- the easiest road to artificial minds is to mimic the ones we have access to and tinker from there. The impliciation is that the first artificial minds are going to look an awful lot like us, up to and including acting at the same speeds.
This is an expansion of what Greg F #23 is saying (I think) but to Greg I point out that there is nothing in principle stopping us from simulating arbitrarily large numbers of neurons, and that many programs are on-going around the world doing just that with the intent of comparing the fine-grained, neuron-level responses of the simulated nerve clusters to the real things.
Second, on Kurzweil: Yes, I have read his latest book, and I can put my finger on some areas where he's wrong. I don't think he's wrong about the trend line of computational price performance. Moore's Law is not a law of physics, but it's an outgrowth of, and bounded by, the laws of physics and economics. So long as physics leaves open the possibility, and economic demand remains, we'll keep going along that curve.
Also, there are two ways to look at the results of an error on Kurzweil's part, here. One is the way Celebrim describes it, with an error resulting in Kurzweil's being not just wrong, but way wrong. A brain equivalent might cost, not $1000 by 2029, but $100,000 by 2029. Gasp! Shock! Two order of magnitude!
But on the other hand, if he's off only in an exponent on the trend line, that might mean that the time of $1000 brain equivalents is 2039. In that lens, that's not a big deal, is it? Not really. And given past history, I can be convinced that the exponent in Moore's Law will change for the wors over the next 20 to 30 years, but convincing me that growth will go linear is going to be a tough sell.
And third, on the notion of ethics:
Oh, my God. It's long nagged at the back of my brain that these technologies, when they come, are going to spark social and ethical controveries that make the present abortion debate look like child's play. And sadly, I think the nature of the problems are such that, although we can see them coming in rough outline, no philosophy can really address them until they're here, because until they're here, we won't really understand.
For the moment, I leave people with the following thought: My free time reading right now, is The Machiavellian Moment by Pocock, a study in the transference of political theory ideas from the Florentine Renaissance to the Northern Atlantic states, specifically England. In that book, some of the Florentine thinkers observe and theorize that ethics and morality arise from dealings between equals, and then worry about the ramifications of inequality under that theory. This is one small part of the development of the tyrant as amoral: if the tyrant has no equals, how is morality defined?
It's a sobering thought, and one which my single paragraph hardly does justice to. (Warning: Pocock is an incredibly dense writer, and I'm quite sure that I'm missing much of what he's saying. I am not a political scientist, and I'm punching above my weight in reading him.)
What are the ramifications of that happy cheerful little notion to the ethics of the mid-21st century? It's something one wishes Vinge would explore directly instead of glancingly. Or perhaps Stross, if he could bear to be a little more serious than usual.
Perhaps I'll do it myself, some day. (But by the time I get to it, hell, we might already be there....)
Celebrim.
I understood your analogy to flying machines the first time around. However, a flying machine depends on one physical principe: fluid dynamics. Therefore, all solutions must rest inside this principle.
Computers working towards AI do not depend on one principle: they do not have to keep getting smaller, or faster, or more interconnected. They could get more RISC-like, or CISC-like, or more parellel. In short, your analogy fails. You are talking about one physical law and how the airplane designers hit that. I am taling about the general rate of progress across a dozen physical laws. A hundred physical laws. Will we hit a wall on all of these simultaneously? I doubt it. Now there is a very interesting collary to your example. Is there some law of computational intelligence we have yet to discover? That's a very fascinating subject. I beleive not, I believe that has been argued to death, but maybe not. We won't know until we get there.
As far as "predicting within a decade or two" -- hey, I'm not putting any money on that, either. But if someone asked me to bet money on it, or write a book on it, I'd probably use the same figures Kurzweil uses. That's the only thing we have right now.
IA looks extremely probably in a generation or so (say next 30-50 years). Real, transhuman AI? If I had to guess, I'd say yes. No matter how complex the human neuron is, I have no doubt we can model one today. Probably several. If that statement is true, then you are back to the exponential math and "is there a wall" question.
One thing is for certain: any sort of true AI we create will not be what we expect. There are so many aspects to intelligence: body awareness, sensory stimuli, moods, body posturing, etc. that anything in a box is going to be very alien to us. (and us to them)
As far as the other poster who assumed we would need to develop all 3 big technologies or "life sucks". There is another scenario. We could all just hook-up, turn-on, and vegetate. There's no huge war of AI mech monsters in our future. It's more like six billion people watching "I Love Lucy" re-runs in their brains over the giga-net while drooling on themselves. Or something close to it. We're not going to explore, we're going to collapse into a huge, atrophic hive mind.
Well then, perhaps it does suck. But it will suck in a different way. (grin)
celebrim, we're making progress in nanocomputing, optical computing and 3D etching on chips.
google nanotubes, etc.
Greg F, i explicitly didn't say that we would have the software--that is the tricky part. ;)
Brains have memory and processing power, that is what the single chip will have on board. A single computer chip costs what? Could a thousand of them cost what say a paragon or a bluegene costs now?
I worked in advanced architectures a lot, and the connection between SIMD PE's is still the tricky part. The brain has sophisticated pattern association algorithms, but I don't think they are impossible to puzzle out.
Kurzweil's theme is technological acceleration--his data points map the curve. The singularity is just dependent on dy/dx, the rate of change. It is believeable to me, but i have sufficient background to get where he is coming from. ;)
Daniel, transhumanism is the reason your i-love-lucy scenario won't happen.
I see Joe has started a transhumanism thread.
See you there. ;)
Moore’s law was never about processing power. Moore’s law, as it applies to semiconductors, is partially limited by the feature sizes that are approaching the dimensions of the atom.
As celebrim points out were dealing with a "mature technology". The technology has limits and were getting close to them.
I don’t disagree about the software which has gone nowhere. Where I seem to be in disagreement with everybody is the assumption that we even have, or will have, the hardware.#29, Greg F:
It's important to remember that we're not just simulating open-loop, here, and assuming that we're good. There are a number of closed-loop simulation projects of varying scales, where the idea is to simulate neural structures in silico and then walk across the hall to the bio-lab and put the real actual structure through the same paces as the simulation and compare the outputs.
That way, we don't have to arue about whether we know "very little" or "a whole lot" about neural activities Unlike Aristotle, we simulate, we test, we evaluate, and we improve.
If we have to, we can go right back to simulation at the atomic scale before we learn what can be thrown out of the picture to get better performance. We've already demonstrated the ability to do simulations at the 100,000 atom level and beyond in the pursuit of engineering design. (And if someone throws in the idea that QM is critically important and adds to the inevitable computational complexity of such a simulation, let me only out that if that were true, going through an MRI machine would be tantamount to hitting your unprotected brain with a sledge hammer.)
We could also, if necessary, simulate these things using analog techniques. Not every transistor is slaved to digital applications.
Also, we're approaching maturity in planar VLSI technology, but the end date keeps getting extended. There is an ultimate limit to that, but there is no demonstrated near term limit to processing density. (Some physicists have worked out actual hard physical limits, but they are esoteric, and mind-bogglingly larger. We're not there, yet.)
There are several paradigms waiting to take up the slack after VLSI really is mature, not to mention a proof of concept on how to get a lot of processing power inside a small, low-cost package: The human brain.
This is an interesting discussion. But there is a larger question in the background that also ought to be part of the discussion: whether in fact modern civilization as a whole is transforming at an accelerating rate.
An argument could be made that modern civilization has been decelerating for a century. Look at a painting of New York in 1800. Then compare with a photograph of New York in 1900. Then look at a photo of the city in the year 2000. Change there is, but the revolution was clearly in the nineteenth century.
I don't doubt that microscale technologies will have a dramatic impact eventually but their greatest effect (if they have one in the short run of the next half century) may be to facilitate the adjustment to a sustainable economy and world rather than promote upheaval.
CMOS is maturing and "classical" scaling is broken (lack of voltage and oxide thickness scaling means power consumption is more of a problem), but density continues to scale.
However, I believe it is unwise to extrapolate Moore's law past 2015, certainly unwise past 2025 or so. I can't consider the $1K price point in 2029 to be aything other than "made up".
That said, I believe we will see a machine capable of simulating all the neurons in a human cortex at a sub $1M price point before CMOS scaling ceases to be economically viable. Per #30 Marcus, yes, those neural simulations are pretty good, we know pretty well how to simulate neurons. So, given some modest $50M black budgets, one can predict some pretty interesting machines will be build. And these predictions are on very solid ground.
Some other technology (molecular/quantum/whatever) needs to come in and replace (or more likely, augment) CMOS. I hope it does.
What might drive progress toward AI?
Spam filters that let no spam through but never bin desired messages.
And after the Singularity...
"HAL, why did you just nuke Russia?"
"It was the only way to stop the spam Dave. Unfortunately the emergent spam-bot S1 has already released DNA modifying nano-viroids which will induce an overwhelming urge to purchase dubious pharmaceuticals, download pron, and hand over your credit details to financially distressed Nigerians."
"Oh."
"...Unfortunately the emergent spam-bot S1 has already released DNA modifying nano-viroids which will induce an overwhelming urge to purchase dubious pharmaceuticals, download pron, and hand over your credit details to financially distressed Nigerians..."
Dag! They've already got me!
Do I put my credit card number in this little box here?
Several points:
(1) There have been people who have made a lot of money who you might be well advised to avoid and ignore.
(2) Ray may be among these people.
(3) Ray seems to have a psychological need to get attention -- lots of it -- and does not refrain from making outrageous claims toward this end. His domineering (and not easily impressed) mother may have something to do with this need ("See, Mother, I now have the National Medal of Technology. Aren't you impressed now?").
(4) Ray is extremely skilled at building his arguments using claims which are not easily proven true or false. This ignites hopelessly endless discussions among people who find him mesmerizing or irritating/bogus. It works out nicely for Ray because he is always part of the discussions.
(5) Related to this, Ray has the uncanny ability to con the powers that be to bestow on him all sorts of awards and honors. He has even compared himself to Thomas Edison. But if you look closely, Ray has invented hardly anything of major significance. When was the last time you used a reading machine for the blind, a music synthesizer (and make sure you investigate closely what role Ray had in this work) or a computerized poetry writer? Now think about the last time you used electric light, distributed electricity, recorded music, or a recorded motion picture? Ray is quite intelligent (I know this from personal experience), but he's nowhere near an Edison-class inventor. This is a myth most likely started by a shill instructed and paid for by Ray, and perpetuated by the media whenever it mentions Ray. See how it works? You can do the same for yourself!
(6) Predictions on the topic of AI cannot be proven true or false, so they are wonderful material for building arguments about things such as the ominous Singularity!
(7) We can't presently build any machine which runs just like a human brain, but we can make some observations which we all pretty much can agree on. (Take this as a challenge.) Incidentally, isn't this the scientifically-sound place to start? With these observations, we can make some hypotheses we can test. Then we can look at the results of these tests and adjust our hypotheses. Outside of that, everything is just b.s. Note how much of what Ray says is outside the scientifically-sound realm (although he is very good at making it sound scientifically-sound). How far would Edison have gone if he just b.s.'d? Someone else would have claimed the patent for electric light. It is sad that Ray is using his life and intelligence to get time-wasting bozo's like us to blather about his b.s. rather than, say, living up to his desire to be like Edison and invent some things which revolutionize our daily lives?
(8) Edison indisputably established himself as knowing what he was talking about by demonstrating he knew what he was talking about. He did not build castles in the air with unprovable claims.