I think a good model for "the singularity" is a long straight highway. Stand in the middle and look down its length, and it appears to converge to a singularity in the distance. But when you actually drive there, you see it's just more highway.
Basically I believe it's an issue of cultural time perspective. We can't imagine the tools that will be in use in 2112. But, they won't seem so strange to the people of 2111.
I like your highway analogy. Mainly I think the singularity argument suffers from an acute case of perception bias in favour of recent advancements; just as picking any arbitrary point in time and asking a scientist of that era what they thought the most important scientific advancements were.
> We can't imagine the tools that will be in use in 2112.
Pretty much just the hall-filling computers of the Priests of the Temples of Syrinx. Guitar-smashing jerks.
Well, I mean when you look at some of the incredibly broken tools that we use today, it is a bit ridiculous:
1). Emacs/Vim are the popular hardcore/moddable editors, despite the severe clunkiness/outdated architecture
2). C/C++ has a huge mindshare despite the fact that it severely reduces the impact of the individual programmer
3). Proprietary tools are falling away, but their entrenched closed gardens are remaining for us to take care of
4). OS design has almost stalled
I think big paradigm shifts are coming at much faster rates than anyone wants to admit. If you are looking at the past to predict what comes tomorrow, you are an idiot, at least in the field of computing. I feel like we are standing on the precipice of a hole so scary no one really wants to look down at it.
OT Emacs clunkiness/outdated architecture...give me a break, minimal
core with lisp interpreter on top giving programmable and extensible
platform next to none is still what...30 years ahead of anything called
"modern".
I have yet to find anything vim can't do that I'd like it to do. Like most critics you only make vague statements. Anyway, the source is out there, if there really was a need for new features somebody would implement them. Hint, hint!
This is a reasonably plausible argument for progress within fields and paradigm shifts; but consider between fields, and the rate of new field creation. To use the given examples, Newtonian physics, quantum physics and Information Theory began in 1687, 1930 and 1948. That seems quite accelerating (of course you'd need more complete stats to demonstrate this properly).
BTW: I disagree that intelligence and imagination are separate. I think intelligence necessarily includes imagination, so AI would include artificial "imagination" too.
I'd like progress to continue rather than to stagnate; and since progress has historically continued, that seems the way to bet. Those pundits predicting stagnation have historically been wrong.
However, I find it curious that Kurzweil does not address the availability of innovations/discoveries (and the difficulty in finding them), in contrast to the compelling case he makes for historical progress. Are potential discoveries infinite? Is their density uniform, or do they become less likely as you go along because they are limited in supply (the assumption of the article)?
How would one sensibly model the supply of potential discoveries, since it is intrinsically unknown? Perhaps a mathematical model of computation might help, of computable functions and Kolomogorov/Solomonoff program length to compute them. Using this model, the search space of functions increases exponentially with program length - but at what rate would the number of "interesting" computable functions increase? And with what density (with respect to the search space, i.e. how hard they are to find)?
While I agree with you that the rate of paradigm shift increases over time, I argue that their average impact/scope decreases. Newtonian physics opened more room for technological progress than quantum mechanics, etc. Earliest inventions have the most impact (the invention of the scientific method having the most impact of all?).
So my point is that if you look at paradigm shift rate not in volume but in impact, you'd get a linear rate.
For this reason, an AI given imagination and exploding resources would still only make linear progress.
I didn't notice your point about impact. Impact is hard to measure purely objectively, but I agree Newtonian physics, being more fundamental, has had more impact. But it's also had 325 years to do so, whereas Information Theory has had only 64 (to stay with these examples). Given an exponentially expanding search space, there's always more to see (but this is an assumption, not a conclusion; I think the nature of the search space is the essential question at issue here).
How much impact did Newton have on the world by 1751? How much impact will Shannon have had by 2337?
Intelligence and Imagination could not be more separate.
I'm defining intelligence as sheer IQ, analogous to a computers' ability to process information. Imagination is so essentially human in nature that there is no way to describe current AI as "imaginative". For example, the person with the highest IQ in the world is not more creative than the best musician. Some of the smartest minds in the world cannot "think outside of the box" effectively, because their minds are not wired that way. Intelligence is not imagination.
Can anyone please explain how an intelligence-self-improving computer will shift from being incredibly smart to wildly imaginative?
That's the one hole in singularity that I can't figure out.
You started off defining intelligence as something limited and not independently useful, and then you concluded that its presence was limited and not independently useful.
The answer is that it's not actually obvious that "intelligence and imagination could not be more separate" or that "imagination is...essentially human in nature," so when most people talk about AI, they are not excluding those abilities.
Imagination isn't any more human specific, than inteligence, or walking upright. That we can't do machines that are imaginative now, doesn't meant it's impossible to do.
Besides - what if imagination is just heuristic to search solution space quicker? Maybe machine with orders of magnitude more processing power can give better solutions without using this heuristic? That would mean it's us who thinks inside the box (because we use this heuristic), and computers can search solution space exhautively, so they will come up with solutions that escape our heuristic - we wouldn't even consider such solutions.
Yes, it doesn't really matter how intelligence is defined, just that in aiming at creating an artificial mind, we'd try to include whatever was useful, including imagination. My reasoning is, if one machine can do it (us), it's possible for another machine to do it. (Though there's a question of whether we are smart enough to create that machine.)
Exponential search space is a killer. Some form of heuristic seems necessary. Perhaps we will one day be able to simulate the heuristics we ourselves use - and then improve on them (e.g. apply them to more data than we can; use more complex heuristics; etc along the lines you say). Given variation between intellectual abilities of human beings, it is does not have an obvious fixed limit, and we may be able to extend the range of that variability. Though I expect increasing the associations an "imagination" can find suffers from exponential explosion.
> Perhaps a mathematical model of computation might help, of computable functions and Kolomogorov/Solomonoff program length to compute them. Using this model, the search space of functions increases exponentially with program length - but at what rate would the number of "interesting" computable functions increase? And with what density (with respect to the search space, i.e. how hard they are to find)?
"To use the given examples, Newtonian physics, quantum physics and Information Theory began in 1687, 1930 and 1948."
Some things that we consider significant today won't seem significant from the perspective of centuries. Similarly, someone in 1750 would be able to list significant new fields developed recently that we would omit. For example, Newton also did work in optics, which we pretty much forget about today.
While I would consider Newtionian physics to be much more fundamental, I would challenge the notion that it had the most impact. Compare the ~1900 feeling that almost everything had already been invended. Newtonian physics impact to technology was mainly mechanics. Then came thermodynamics which gave us (steam &) combustion engines and more. Nuclear & quantum physics lead to bombs, reactors and finally seminconductors and so on. I would say todays thnological relevance of thermodynamics and quantum physics is at least as high as that of Newtonian physics. Also: most of todays physicists do not look for Bosons, they do things like advancing semiconductor masking technology, liquid crystal physics and the like - i.e. they are not "fundamental" but "applied" physicists.
Also - with respect to the original article - isn't the much faster exponential growth of the resources invested in science since around 1900 mostly an indicator of expected technological returns by society at large? Science is a huge economy success story since about a century ;-) The question then would be how long can we afford to sustain the current exponent?
If the Singularity is defined as intelligence explosion following the creation of strong AI, and if intelligence explosion is impossible, then the Singularity is not coming, even in the far future. Progress will just carry on linearly.
Saying something is impossible, even in the far future, is just stupid. If you think about it, humans are only familiar with an infinitely small sliver of what's out there in the universe. We're taking the little knowledge we have, and saying AI intelligence explosion is impossible. That's like a baby saying he's stood up before, and walking is just impossible.
Also, Ray Kurzweil is not known just for his imagination. Some of his research has enabled groundbreaking technologies. For example, he is one of the creators of speech recognition technology (Nuance, which spun out of Kurzweil's company, allegedly powers Siri). I had heard of Kurzweil long before I knew anything about the singularity.
Indeed. However, a number of researchers have had arguably more impact that he did on NN research (such as LeCun), and these people are completely unheard of outside of the NN community. I can't imagine Kurzweil becoming the media star that he is without his singularity books.
Thanks for this article. I've always been fascinated by singularity, and even though I don't know too much about it, I'm excited to see the people who fully support your position to create some great conversation on this topic.
You are not using a valid mathematical method because paradigm shifts include emotions and social behavior. What you obviously left out or just don't understand is that AI is not bound to pain or instincts of self-preservation. When the time has come for an AI to supersede the human intelligence the singularity will happen faster than the writing of your non-scientific patchwork article.
Arrogance is unjustified because your arguments are without any solid background and only show someone in research at the beginning of his career. If you are young you should better listen than speak out loud.
Saying scientific progress is linear is a pathetic explanation of someone who obviously compares scaling to a paradigm shift. Of course being able to write 10GB on 1cm² of a hard-disk was done by researchers, but they took a hard-disk with 10MB on 1cm² and tried to make it smaller. That's scaling and not scientific progress which could result in a paradigm shift or even a singularity.
My only advice for you, as I am myself am too arrogant to write more, is that you take the works of John von Neumann about AI. Then you compare the maximum different states of the human brain to the actual capabilities of super computers, find the break-even-point with Moore's law, then add 10 years of progress to improve the software and you have 2040.
And yet, what are ants to us? Limits of an AI may be incomprehensible to us, and the point of singularity is not that intelligence might increase without bound, but that it will reach our level and appear be increasing at an exponentially beyond us. (Any increase in intelligence beyond our level may be obvious but unquantifiable, so the question of exponential vs simply "beyond us" may be unanswerable.)
Logistics curves are a favorite concept on HN relating to resource usage, so imagine that intelligence follows a logistic curve, and humanity is way above ants, but still very close to the bottom. Once an AI gains the ability to iteratively improve its own mind, first software, and then hardware, it could accelerate up the curve and the separation between our intelligence and AI, if measured, would be approximately exponential until it's so far ahead of us it doesn't matter whether there's a hard limit (logistic curve) or not (true exponential curve).
Sorry, I think my point wasn't properly explained.
My point is that intelligence isn't everything.
An AI may become more "intelligent" than a human being, but it doesn't mean any singularity will happen. It does not mean it can self-improve. It does not mean it can find solutions we did not think of.
It's a bit hyperbolic to say "we don't know what greater intelligence can do" and conclude "it will therefore do incredible things".
And to bounce on your answer, are we more intelligent than ants? Does it even make sense to say "we're more intelligent than ants"? Our differences with ants go far beyond intelligence.
An AI would be able to self-improve its software (unless artificially limited, and it's not clear to me whether that would be effective), and with enough initial robotic and sensory capabilities it would be able to self-improve its hardware, both computational aspects and robotic aspects.
The only reasons it wouldn't are if it didn't put effort into it (too busy watching Lost), or if intelligence cannot really be improved.
Humans might do the equivalent of rewriting software through thinking. But it's a more limited and slow process than could be accomplished with full read/write access to software running an AI that's completely malleable and rewritable in moments.
There's plenty of evidence that intelligence can be improved. Working memory is important. We have very limited working memory. We are frequently distracted by emotions. If an AI can operate at roughly equivalent speeds with greater working memory, and if it's not distracted by emotions, it may miss some of the "finer" points of life, but I can't imagine it not leapfrogging ahead of meager human progress.
Maybe it's not possible. Like the thing at the beginning of Fire Upon the Deep, if it is possible, we won't really know if it is until it exists. If it's got the temperment of a class 2 perversion or the Blight, we're toast. If it's a Power that finds us amusing or irrelevant, we might be okay.
Even if you disagree with the ideas of what general artificial intelligence could do to the human existence on Earth and elsewhere, it is justified to spend some effort thinking about it. The potential consequences to human existence are all-encompassing. So when something is deemed a potential existential risk, it's worth thinking about.
If the current development in computing continues, I believe that superhuman artificial intelligence is inevitable. And that if we don't prepare for it properly, we might screw it up in a historical way.
This seems like arguing with a door. The whole concept of the singularity is absurd. Taking your time to construct arguments of why such absurdity would happen slow instead of fast, if it happened. Is just like trying to come up with scientific explanations why unicorn poop couldn't be technically as magical as rainbows. Reminds me of those nerdy arguments we all had, debating whether superman's clothes would burn upon re-entry on the atmosphere. Fun times, but useless.
Proponents of singularity have a broken understanding of "intelligence". We already have computers that perform plenty of intellectual tasks better than humans. Even learning itself, or coding and reprogramming itself. But this is not "intelligent" yet according to them. They'll only be satisfied when "intelligent" means mimicking a human being for no other reason other than mimicking a human being's sake. They don't want an intelligent computer, nor actual practical solutions to real problems. What they really want is a really complex fart app.
That is like claiming the universe is more intelligent than we are because we could never hope to calculate one millisecond of what the universe does if we all had IQs of 200 and 15 billion years with pencil and paper.
What computers do is not intellectual. Intelligence is probably just a stochastic random process of clawing one's way up a hill, but even with enormous computational power, it took the solar system 4.5 billion years of not-well-directed effort to produce us... (and to produce whatever creature that blotch was on the curiosity camera :). Only once we existed could the universe, through us, produce the works of art, literature, culture, and scientific knowledge that we've accumulated very rapidly.
That's obviously subject to whatever your definition of "intelligence" is. The only reason why you can even disagree, is because we don't have an objective definition. So we could argue forever and not get anywhere.
But what matters, is that at the end of the day. For almost any practical and objective definition of "human intelligence" that you might write down. I can write a computer program that will follow those specs and execute it better than a human being. Today. Fact is, humans are nothing but machines and our computers are already better than us in most tasks. Talking about the point which "computers overtake human intelligence" is dismissing the hundreds of different sub-fields in AI research which already got way past that point. Not only with better processing (as you imply), but also with better algorithms than us.
Or we might stick with our current subjective and useless definitions of the word. And keep dreaming about how would it be when we get to somewhere undefined.
> For almost any practical and objective definition of "human intelligence" that you might write down. I can write a computer program that will follow those specs and execute it better than a human being. Today.
If that were true, how come humans are still superior in many (intelligence based) tasks? From games such as Go and Arimaa to more practical tasks such as laying out electronic parts on circuit boards (autorouting).
Mix in self-learning and generality, and I doubt that a better computer program could be programmed today.
I don't get your argument at all. As in, I'm not sure what you're trying to argue. If you agree that computers will soon be capable of everything a human brain does, and that the brain is just a machine that could eventually be copied and improved, how could there _not_ be tremendous and unpredictable changes to the world when it happens?
If humans create artificial minds that are able to operate on their own, what we decide are the important problems to work on could end up becoming a moot point. You can't really negotiate with someone who's a lot smarter than you (not more than a chimpanzee can negotiate with humans).
If you're criticizing philosophers and researchers for "dreaming" about these ideas, you might want to have a look at some of the thoughts around what would happen if we created self-improving, human-level artificial intelligence. The consequences are potentially catastrophic to human existence, and hence justifies some effort even if you think it is completely unlikely. And whether it is unlikely or not is subject to strong debate. Have a look at
The information revolution must be accelerating the speed of scientific discoveries. In the past scientists would make paradigm shifting discoveries that would be forgotten or that would be confined to a specific region. It would take other scientists centuries to rediscover them and build on top.
Many discoveries come about by improving on and combining previous discoveries, so what needs to be improved to speed up science is better access, categorization and distribution of discoveries. And this is where I think a human-level AI would have an advantage over a human. Even if the AI is unable to retain more things in memory than a human it could clone itself and have an army of scientist AIs in charge of analyzing each new discovery and seeing how it might fit in with all other discoveries. Depending how resource-hungry the AI is we might be able have millions of brains working on every potential far-fetched hypothesis to see if it leads anywhere. It would be an industrialization of science.
That seems like it would have to lead to faster discoveries.
But I'm under the feeling that reality is much more complex: access to more information can as well inhibit creativity as it can inspire new discoveries.
I can't say I fully the relationship between the creative scientific process and information exchange/consumption, but I'm pretty sure that past a certain point, the more papers you read the less likely you are to come up with something new. At a macro scale, communication is definitely a driver of progress, but at a micro scale?...
And I'm pretty sure that the fewer papers your read the more likely you are to come up with something you think is new, but isn't.
Of course, depending on the context whether something is "new" or not might not matter, it certainly does in academic research, but if you are making stuff work it's less of a factor.
The singularity may not be coming, however, there are still huge questions and challenges looming ahead. Individual humans don't seem to be able to resist the thirst for power, and society as a whole seems too passive to stop them... We just watch as it unfolds, hoping we'll get cool gimmicks in the process. That's insane.
"Whenever a man chooses his purpose and his commitment in all clearness and in all sincerity, whatever that purpose may be, it is impossible for him to prefer another. It is true in the sense that we do not believe in progress. Progress implies amelioration; but man is always the same, facing a situation which is always changing, and choice remains always a choice in the situation. The moral problem has not changed since the time when it was a choice between slavery and anti-slavery."
-- Jean Paul Sartre
It's all just a bunch of tools! We could use them to be good to each other, or to just keep doing what we did; moving from scarcity of resources and horsepower and intellect, to artificial scarcity based on greed and control. Technology just amplifies the impact of our choices on us, others and our environment; but it doesn't make them for us.
I think the advocates of the singularity see it as coming from an explosion of the possibilities of technology rather than an explosion of science in general. Kurzweil himself mentions how to the average research scientist, science appears slow and linear (with jumps most coming when a scientist gets a different set of tools to work with).
> Let’s have a hypothetical AI that researches artificial intelligence, and that constantly rewrites its own code to incorporate its finding into its own intelligence.
No,
A hypothetical General AI would double the amount of circuits involved in it's processes to double its intelligence and go from Monkey-brain-sized to human brain sized. Repeat as necessary. The intelligence algorithm running on those circuits doesn't have to improve (though we'd hope it would at least a little).
Of course, you would need an initial intelligent parallel algorithm. That might indeed be hard or impossible. But his argument in itself doesn't proving anything about this.
Sure, even with flexible general AI, some problems would remain hard but a lot of things aren't the Knap-sack problem. In fact there's no evidence the massive success of human intelligence has had much to do with tackling NP-complete problems. Instead, flexibility and pattern finding set human apart from both other animals and the most powerful artifact we can so-far create.
Just consider, Moore's Law continued at least for a fair period driven by humans whose innate intelligence didn't increase at all.
You don't understand intelligence. Doubling the circuits doubles the speed of the intelligence, but it has no effect on it's quality.
You are assuming the existence of not just an intelligent parallel algorithm but an infinitely scalable one (i.e. one that could actually take advantage of more circuits)! Because without that, you don't need to double the circuits - you could just give them more time.
Concepts that are beyond a particular intelligence will still be beyond it, even if the speed is increased.
People have a really hard time admitting that some people are truly fundamentally smarter than others. It's not just that they can think faster - they can think things that other people simply can not, even if they worked really hard at it.
I'm sure you've heard the saying "I didn't think of that". That's exactly it, even if it thinks faster, it will still never "think of that".
Given that intelligence is probabilistic and approximate, doubling the circuits would increase accuracy. I think it's pretty plausible that intelligence does work this way, since humans make mistakes all the time.
This is actually under debate. There is disagreement among neuroscientists whether intelligence is a result of brain modularity (specific regions of the brains optimized for certain tasks), or due to brain size alone. All that said, I'm skeptical of the singularity since it assumes that a recursive process can continually improve intelligence without significant diminishing returns. The problem with this thinking is that it fails to take into account evolution. All this research into AI is based on the assumption that "intelligence" should be like human intelligence. But human intelligence has had several billion years to evolve and is highly optimized for our environment (actually our environment from several tens of thousands of years ago when we were hunter gatherers). It seems naive to me to assume that we are not nearing a local optimum in what is possible with human intelligence. I don't believe a singularity is possible because by recursively improving "intelligence" you will near the local optimum of that form of intelligence but that does not mean you can continue improving that intelligence indefinitely.
You don't understand intelligence. Doubling the circuits doubles the speed of the intelligence, but it has no effect on it's quality.
Uh the singularity may be a crack-pot idea in the end but it seems like brings out a whole of slew crackpot refutations.
No, I don't understand flexible human intelligence or even animal intelligence fully and neither do you or other human. That's why it's not been programmed.
I do know that doubling circuit size doesn't double speed.
People have a really hard time admitting that some people are truly fundamentally smarter than others.
And you'd have a hard coming up with any evidence for this "fundamental" rather than relative intelligence difference.
The singularity, crackpot or not, is based on the hypothesis that there is a single "neocortical algorithm". It may indeed be dubious theory but it's cobbled together from some real evidence. If you're going to have counter-arguments to this, why not do the same?
The author completely misses the point of these arguments.
This is not about science increasing the rate of scientific progress. It's about computation improving the rate of computational progress. Particularly, the "Singularity" is about that happening in a world where strong AI exists.
And I'm not even going to harp on the fact that the number of computations per second that we perform on this planet has been undoubtedly exponential at least for a while, because that's not the real point.
The real point is that like the author says, human creativity is still the driving force in computational advances, and human creativity can't be sped up by periodically doubling clock rates or core counts.
If it could? Those doublings would happen even quicker, and the exponential growth we've seen so far would be nothing.
This does, of course, require that you believe that strong AI is possible at all, and that the entire process of creativity could eventually be put into software. If you don't believe in that in the first place, then there's no point having the argument at all - you don't argue the details of a nuclear chain reaction with someone that doesn't believe in neutrons.
How is throwing more (truly intelligent) computers at the problem any different than throwing more human brains at it? It's not just computing power that had been exploding, it is also the number of people trying to crack the nut of AI (for instance).
Because the human brains can only communicate and acquire knowledge at a limited rate. This causes the "exponential decrease of discovery impact of each succeeding researcher" that the original post refers to, but he misses the point completely. We can't 'stand on the shoulders of giants' until we've climbed up there, slowly and painfully. An AI can just download the lot. And share any new discoveries with the hive mind.
it's sci fi, but we haven't ruled out that human brains can't acquire knowledge at rates orders of magnitudes faster than current learning methods. Big advances in neuroscience and linguistics could make this possible.
Right now, the limits on our bandwidth is language and our sensory inputs. It's not completely ridiculous to think we could encode certain kinds of knowledge in more compact forms, and transmit them via an artificial sensory process that has higher higher bandwidth that audio/visual language based learning...
It's about computation improving the rate of computational progress
But computation doesn't improve the rate of computational progress. That's achieved by improvements in manufacturing. Computation does help a little, giving us better tools for chip design. But mainly it's manufacturing.
The only way for computational progress to exponentiate, past the limit of Moore's Law, is to use more physical resources - computers get literally bigger.
And so expensive that only the five richest crowned heads of Europe can afford them!
But computation doesn't improve the rate of computational progress.
As long as it's people developing computing devices, that holds. But when you have a machine that is capable of improving its own software, you get a self reinforcing loop where computation improves the rate of computational progress.
Intelligence doesn't even have to feed on itself. Imagine a human-level hardware designing AI. As long as Moore's law applies, it can double its computational power every 18 subjective months. From our point of view, that's 18 months, then 9, then 4.5…
If Moore's law where unlimited, such an AI would have infinite computational power in 36 months, with absolutely no increase in actual intelligence. But of course Moore's law is limited. Still, we could expect quite a bit of progress from a mere AI hardware designer.
When hardware is twice as fast, the AI works twice as fast. In 9 months, it will do 18 months worth of work. Now keep the doubling up, and it will have done infinite work after 36 months.
This assertion is groundless. Why would Moore's law accelerate? In fact, Moore's law is currently decelerating and gains are increasingly harder to find. Further, Moore's law says nothing about computational speed, but only the density of transistors. These two are related but not identical.
Edit: also, from my understanding, the gains in transistor density largely depend on new discoveries in physics. Before a computer can aid in accelerating Moore's law it would need to be sufficiently advanced to generate new discoveries in physics, but at this point you'd have a computer smart enough that it Moore's law wouldn't matter much. Seems like putting the cart before the horse to me.
Transistor count and computational are so tightly correlated that's nitpick.
Now I did not assume "Moore's law" would accelerate. I assumed it would stay constant. The key point comes from the fact the AI (or AI hive) would trivially benefit from hardware speed-ups.
And yes, Moore's law won't really count, compared to the rest we will be able to do. I was just trying to be as conservative as possible. (Though Moore's law still holding until strong AI is quite wild).
Further, if we build a human-like intelligence on a chip, this would plausibly be a highly parallel algorithm. By that token, you wouldn't need more transistors-per-inch, you could use chips that snap-together with high bandwidth connections and perhaps some longer-distance traffic control connections.
The kind of design isn't an NP-complete problem but it certainly could be helped by a powerful AI.
Also, Moore's Law is breaking for speed but not for transistors-per-inch. That's why we're seeing multicode chip now.
are you stating that computer manufactures don't rely on computers to design the next generation manufacturing techniques? Or that faster/better software isn't going to increase the rate of improvement in computer manufacturing?
Material scientists , physicists, chemists, chip designers are all using software for modeling and exploration. And bigger/faster/better computers are helping drive innovations back into computer manufacturing. The process is feeding itself, and it has been for awhile.
Absolutely right, as I pointed out below, the very young author obviously compares scaling to scientific progress. As we currently have no model in AI which could create a cognitive skill, his writing becomes obsolete and will be laughed at when we do.
Actually he misses a second point too ... the reason you don't see exponential growth in human-led research is that we each start out with "empty" brains. To become truly knowledgeable in a field may require a life-time.
I'm not arguing that the singularity is near either, but machines would have the advantage of creating offspring that are already as knowledgeable/intelligent as themselves. Something I couldn't do for my children.
Linear increases in science can drive exponential increases in what's possible.
I remember computers of 30 and 40 years ago. Today everyone in the world has free access to supercomputers that translate languages and search petabytes of data (aka Google). We didn't need to revolutionize basic physics to get here from punched cards.
Expect a similar revolution in the next 30 to 40 years.
He just downgraded "intelligence" and replaced it with "imagination" and his arguments are valid. However, an AI with imagionation, now thats something and still opens up possibility of singularity.
I'm not that much worried about whether or not there's an asymptote of diminishing returns for self-improving intelligence. I'm quite a bit worried about just how far away a genuinely self-improving intelligent system will move from the level of human capacity before it starts hitting the wall.
I think Kurzweil is mostly the one who keeps going on about literal unending exponential growth. Vinge's original singularity idea was more about things just becoming irreversibly incomprehensible for regular humans, since ultimately limited or not, self-improving intelligences are going to end operating on a level very distant from human thought.
Okay, here's the thing. It's certainly possible that strong AI could exist. Pretty much the only way of finding out is by attempting to build it. If it's possible, we'll be able to build it eventually. If not, we won't. If the magical singularity happens, it'll flow naturally from the strong AI. And if it doesn't, it doesn't. And, well, if we get strong AI but no singularity, great, we still get strong AI and we can use it to do some very clever shit that makes the sort of things the AI community are doing now seem primitive.
Given this, I'm not sure why singularity promoters give that much of a shit whether people believe it. I'm certainly skeptical of it. Given I pretty much use a 40 year old text editor to write in a language that's about 20 years old and is just about approaching the discoveries that the Lispies made back in the 50s, I'm not sure what progress we're talking about. I write code that basically instructs a computer with pretty much the same level of semantic complexity as one does when giving instructions to a particularly stupid child.
But perhaps I've got this wrong. Perhaps the singularity will happen. I don't understand why the singularity promoters are so angry about people who are skeptical of their claims. While I was doing the rounds on Wikipedia a while back, I found this image...
Apparently, if someone tells me that a super-intelligent AI will create a brain many hundreds of times better than our own and then allow us to live forever and my reaction isn't an immediate "oh my god, you are so, so right!" then I'm a 'singularity denier', equivalent in status to a holocaust denier. If the singularity idea is pretty much bound to happen as a result of emergent technological progress or whatever, why is my doxastic consent such an important requirement for it coming about?
I freely admit I'm probably not as smart a guy as Ray Kurzweil. But let's not forget: the history of bizarre cults shows us that smart people can believe some very stupid things.
> Given this, I'm not sure why singularity promoters give that much of a shit whether people believe it.
In my case that would be because the default scenario for a technological singularity is Eternal Doom. By default, if AI researchers do not pay extreme attention, the first smarter-than-human AI will have different goals than we do, and recognize us as a threat to those goals and eliminate us all. Or squash us like bugs without even paying much attention. Or use our constituent atoms to maximize its goals. Or keep us drugged and lobotomised in soft cages, so that we're safe and happy. Or something.
And of course, we won't be able to stop it, because it will outsmart us at every turn (we can do the same to chimps).
This is why this singularity thing is at least worth looking into. Being able to nuke away our civilization is bad enough, but this could be even worse. Or it could turn our world into a paradise.
And of course, with more believers, there will be more donors to the organizations working on this, more researchers working on it… So belief does count, to some extent. (On the other hand, becoming angry at non-believers like a crusader is unlikely to spur positive reactions. By cleaning up this image description, you actually helped our cause.)
It's not the first smarter than human AI that is the threat. It's the first self-improving AI that people have to worry about. It can start out at a much lower than human intelligence (whatever that means) but as long it can keep improving it could turn into the nightmare scenario.
The safeguards need to be built into the self improving AI, no matter how 'dumb' it starts out as.
Well, any path that lead to superhuman intelligence is dangerous. Self-improving AI may be the likeliest, but it's not the only one. Think Upload, brain-computer interfaces, or computing overhang, for instance.
The Problem with his model is, that he assumes that he ignores the "shoulder of giants" effect in his model, since he assumes that a later researcher has the same number of points as the earlier, but only hard problems to solve. ( Perhaps I will extend his model a bit later.)
Rather use your imagination, the one thing that makes you a beautiful unique snowflake. Intelligence and hard work should be merely at the service of our imagination. Think outside of the box. Break out. Shake the axioms.
Honest question: can't this process of "thinking outside the box" be reduced to a logical process? Wouldn't a system which has ability to identify axioms, also have the ability to vary those axioms and try out alternatives?
In other words, why shouldn't better logical ability lead to better imagination?
83 comments
[ 4.8 ms ] story [ 123 ms ] threadBasically I believe it's an issue of cultural time perspective. We can't imagine the tools that will be in use in 2112. But, they won't seem so strange to the people of 2111.
> We can't imagine the tools that will be in use in 2112.
Pretty much just the hall-filling computers of the Priests of the Temples of Syrinx. Guitar-smashing jerks.
http://www.energybulletin.net/stories/2012-06-27/ruthless-ex...
I think big paradigm shifts are coming at much faster rates than anyone wants to admit. If you are looking at the past to predict what comes tomorrow, you are an idiot, at least in the field of computing. I feel like we are standing on the precipice of a hole so scary no one really wants to look down at it.
BTW: I disagree that intelligence and imagination are separate. I think intelligence necessarily includes imagination, so AI would include artificial "imagination" too.
I'd like progress to continue rather than to stagnate; and since progress has historically continued, that seems the way to bet. Those pundits predicting stagnation have historically been wrong. However, I find it curious that Kurzweil does not address the availability of innovations/discoveries (and the difficulty in finding them), in contrast to the compelling case he makes for historical progress. Are potential discoveries infinite? Is their density uniform, or do they become less likely as you go along because they are limited in supply (the assumption of the article)?
How would one sensibly model the supply of potential discoveries, since it is intrinsically unknown? Perhaps a mathematical model of computation might help, of computable functions and Kolomogorov/Solomonoff program length to compute them. Using this model, the search space of functions increases exponentially with program length - but at what rate would the number of "interesting" computable functions increase? And with what density (with respect to the search space, i.e. how hard they are to find)?
So my point is that if you look at paradigm shift rate not in volume but in impact, you'd get a linear rate.
For this reason, an AI given imagination and exploding resources would still only make linear progress.
How much impact did Newton have on the world by 1751? How much impact will Shannon have had by 2337?
I'm defining intelligence as sheer IQ, analogous to a computers' ability to process information. Imagination is so essentially human in nature that there is no way to describe current AI as "imaginative". For example, the person with the highest IQ in the world is not more creative than the best musician. Some of the smartest minds in the world cannot "think outside of the box" effectively, because their minds are not wired that way. Intelligence is not imagination.
Can anyone please explain how an intelligence-self-improving computer will shift from being incredibly smart to wildly imaginative?
That's the one hole in singularity that I can't figure out.
The answer is that it's not actually obvious that "intelligence and imagination could not be more separate" or that "imagination is...essentially human in nature," so when most people talk about AI, they are not excluding those abilities.
Besides - what if imagination is just heuristic to search solution space quicker? Maybe machine with orders of magnitude more processing power can give better solutions without using this heuristic? That would mean it's us who thinks inside the box (because we use this heuristic), and computers can search solution space exhautively, so they will come up with solutions that escape our heuristic - we wouldn't even consider such solutions.
Exponential search space is a killer. Some form of heuristic seems necessary. Perhaps we will one day be able to simulate the heuristics we ourselves use - and then improve on them (e.g. apply them to more data than we can; use more complex heuristics; etc along the lines you say). Given variation between intellectual abilities of human beings, it is does not have an obvious fixed limit, and we may be able to extend the range of that variability. Though I expect increasing the associations an "imagination" can find suffers from exponential explosion.
Like AIXI?
http://www.hutter1.net/ai/aixigentle.htm
Some things that we consider significant today won't seem significant from the perspective of centuries. Similarly, someone in 1750 would be able to list significant new fields developed recently that we would omit. For example, Newton also did work in optics, which we pretty much forget about today.
Just seems silly.
But I suppose you did not read past the title? Sorry for hurting your faith in the Singularity...
Arrogance is unjustified because your arguments are without any solid background and only show someone in research at the beginning of his career. If you are young you should better listen than speak out loud.
Saying scientific progress is linear is a pathetic explanation of someone who obviously compares scaling to a paradigm shift. Of course being able to write 10GB on 1cm² of a hard-disk was done by researchers, but they took a hard-disk with 10MB on 1cm² and tried to make it smaller. That's scaling and not scientific progress which could result in a paradigm shift or even a singularity.
My only advice for you, as I am myself am too arrogant to write more, is that you take the works of John von Neumann about AI. Then you compare the maximum different states of the human brain to the actual capabilities of super computers, find the break-even-point with Moore's law, then add 10 years of progress to improve the software and you have 2040.
I've always found the concept of singularity naive.
Greatest discoveries haven't been done by the most intelligent persons.
Science isn't just a question of intelligence or will, it's also a question of timing, luck and duration.
Everything has a limit. Ants have limits. We have limits. And the supposed "all mighty AI" will have limits if it ever exists.
Logistics curves are a favorite concept on HN relating to resource usage, so imagine that intelligence follows a logistic curve, and humanity is way above ants, but still very close to the bottom. Once an AI gains the ability to iteratively improve its own mind, first software, and then hardware, it could accelerate up the curve and the separation between our intelligence and AI, if measured, would be approximately exponential until it's so far ahead of us it doesn't matter whether there's a hard limit (logistic curve) or not (true exponential curve).
My point is that intelligence isn't everything.
An AI may become more "intelligent" than a human being, but it doesn't mean any singularity will happen. It does not mean it can self-improve. It does not mean it can find solutions we did not think of.
It's a bit hyperbolic to say "we don't know what greater intelligence can do" and conclude "it will therefore do incredible things".
And to bounce on your answer, are we more intelligent than ants? Does it even make sense to say "we're more intelligent than ants"? Our differences with ants go far beyond intelligence.
The only reasons it wouldn't are if it didn't put effort into it (too busy watching Lost), or if intelligence cannot really be improved.
Humans might do the equivalent of rewriting software through thinking. But it's a more limited and slow process than could be accomplished with full read/write access to software running an AI that's completely malleable and rewritable in moments.
There's plenty of evidence that intelligence can be improved. Working memory is important. We have very limited working memory. We are frequently distracted by emotions. If an AI can operate at roughly equivalent speeds with greater working memory, and if it's not distracted by emotions, it may miss some of the "finer" points of life, but I can't imagine it not leapfrogging ahead of meager human progress.
Maybe it's not possible. Like the thing at the beginning of Fire Upon the Deep, if it is possible, we won't really know if it is until it exists. If it's got the temperment of a class 2 perversion or the Blight, we're toast. If it's a Power that finds us amusing or irrelevant, we might be okay.
http://singularity.org/research/
Even if you disagree with the ideas of what general artificial intelligence could do to the human existence on Earth and elsewhere, it is justified to spend some effort thinking about it. The potential consequences to human existence are all-encompassing. So when something is deemed a potential existential risk, it's worth thinking about.
If the current development in computing continues, I believe that superhuman artificial intelligence is inevitable. And that if we don't prepare for it properly, we might screw it up in a historical way.
Proponents of singularity have a broken understanding of "intelligence". We already have computers that perform plenty of intellectual tasks better than humans. Even learning itself, or coding and reprogramming itself. But this is not "intelligent" yet according to them. They'll only be satisfied when "intelligent" means mimicking a human being for no other reason other than mimicking a human being's sake. They don't want an intelligent computer, nor actual practical solutions to real problems. What they really want is a really complex fart app.
What computers do is not intellectual. Intelligence is probably just a stochastic random process of clawing one's way up a hill, but even with enormous computational power, it took the solar system 4.5 billion years of not-well-directed effort to produce us... (and to produce whatever creature that blotch was on the curiosity camera :). Only once we existed could the universe, through us, produce the works of art, literature, culture, and scientific knowledge that we've accumulated very rapidly.
But what matters, is that at the end of the day. For almost any practical and objective definition of "human intelligence" that you might write down. I can write a computer program that will follow those specs and execute it better than a human being. Today. Fact is, humans are nothing but machines and our computers are already better than us in most tasks. Talking about the point which "computers overtake human intelligence" is dismissing the hundreds of different sub-fields in AI research which already got way past that point. Not only with better processing (as you imply), but also with better algorithms than us.
Or we might stick with our current subjective and useless definitions of the word. And keep dreaming about how would it be when we get to somewhere undefined.
If that were true, how come humans are still superior in many (intelligence based) tasks? From games such as Go and Arimaa to more practical tasks such as laying out electronic parts on circuit boards (autorouting).
Mix in self-learning and generality, and I doubt that a better computer program could be programmed today.
If humans create artificial minds that are able to operate on their own, what we decide are the important problems to work on could end up becoming a moot point. You can't really negotiate with someone who's a lot smarter than you (not more than a chimpanzee can negotiate with humans).
If you're criticizing philosophers and researchers for "dreaming" about these ideas, you might want to have a look at some of the thoughts around what would happen if we created self-improving, human-level artificial intelligence. The consequences are potentially catastrophic to human existence, and hence justifies some effort even if you think it is completely unlikely. And whether it is unlikely or not is subject to strong debate. Have a look at
http://singularity.org/research/
Many discoveries come about by improving on and combining previous discoveries, so what needs to be improved to speed up science is better access, categorization and distribution of discoveries. And this is where I think a human-level AI would have an advantage over a human. Even if the AI is unable to retain more things in memory than a human it could clone itself and have an army of scientist AIs in charge of analyzing each new discovery and seeing how it might fit in with all other discoveries. Depending how resource-hungry the AI is we might be able have millions of brains working on every potential far-fetched hypothesis to see if it leads anywhere. It would be an industrialization of science.
That seems like it would have to lead to faster discoveries.
But I'm under the feeling that reality is much more complex: access to more information can as well inhibit creativity as it can inspire new discoveries.
I can't say I fully the relationship between the creative scientific process and information exchange/consumption, but I'm pretty sure that past a certain point, the more papers you read the less likely you are to come up with something new. At a macro scale, communication is definitely a driver of progress, but at a micro scale?...
Do you have any evidence this is a property of optimizing systems in general, as opposed to simply a limitation of human capabilities?
Of course, depending on the context whether something is "new" or not might not matter, it certainly does in academic research, but if you are making stuff work it's less of a factor.
"Whenever a man chooses his purpose and his commitment in all clearness and in all sincerity, whatever that purpose may be, it is impossible for him to prefer another. It is true in the sense that we do not believe in progress. Progress implies amelioration; but man is always the same, facing a situation which is always changing, and choice remains always a choice in the situation. The moral problem has not changed since the time when it was a choice between slavery and anti-slavery."
-- Jean Paul Sartre
It's all just a bunch of tools! We could use them to be good to each other, or to just keep doing what we did; moving from scarcity of resources and horsepower and intellect, to artificial scarcity based on greed and control. Technology just amplifies the impact of our choices on us, others and our environment; but it doesn't make them for us.
> Let’s have a hypothetical AI that researches artificial intelligence, and that constantly rewrites its own code to incorporate its finding into its own intelligence.
No,
A hypothetical General AI would double the amount of circuits involved in it's processes to double its intelligence and go from Monkey-brain-sized to human brain sized. Repeat as necessary. The intelligence algorithm running on those circuits doesn't have to improve (though we'd hope it would at least a little).
Of course, you would need an initial intelligent parallel algorithm. That might indeed be hard or impossible. But his argument in itself doesn't proving anything about this.
Sure, even with flexible general AI, some problems would remain hard but a lot of things aren't the Knap-sack problem. In fact there's no evidence the massive success of human intelligence has had much to do with tackling NP-complete problems. Instead, flexibility and pattern finding set human apart from both other animals and the most powerful artifact we can so-far create.
Just consider, Moore's Law continued at least for a fair period driven by humans whose innate intelligence didn't increase at all.
You are assuming the existence of not just an intelligent parallel algorithm but an infinitely scalable one (i.e. one that could actually take advantage of more circuits)! Because without that, you don't need to double the circuits - you could just give them more time.
Concepts that are beyond a particular intelligence will still be beyond it, even if the speed is increased.
People have a really hard time admitting that some people are truly fundamentally smarter than others. It's not just that they can think faster - they can think things that other people simply can not, even if they worked really hard at it.
I'm sure you've heard the saying "I didn't think of that". That's exactly it, even if it thinks faster, it will still never "think of that".
Uh the singularity may be a crack-pot idea in the end but it seems like brings out a whole of slew crackpot refutations.
No, I don't understand flexible human intelligence or even animal intelligence fully and neither do you or other human. That's why it's not been programmed.
I do know that doubling circuit size doesn't double speed.
People have a really hard time admitting that some people are truly fundamentally smarter than others.
And you'd have a hard coming up with any evidence for this "fundamental" rather than relative intelligence difference.
The singularity, crackpot or not, is based on the hypothesis that there is a single "neocortical algorithm". It may indeed be dubious theory but it's cobbled together from some real evidence. If you're going to have counter-arguments to this, why not do the same?
This is not about science increasing the rate of scientific progress. It's about computation improving the rate of computational progress. Particularly, the "Singularity" is about that happening in a world where strong AI exists.
And I'm not even going to harp on the fact that the number of computations per second that we perform on this planet has been undoubtedly exponential at least for a while, because that's not the real point.
The real point is that like the author says, human creativity is still the driving force in computational advances, and human creativity can't be sped up by periodically doubling clock rates or core counts.
If it could? Those doublings would happen even quicker, and the exponential growth we've seen so far would be nothing.
This does, of course, require that you believe that strong AI is possible at all, and that the entire process of creativity could eventually be put into software. If you don't believe in that in the first place, then there's no point having the argument at all - you don't argue the details of a nuclear chain reaction with someone that doesn't believe in neutrons.
Right now, the limits on our bandwidth is language and our sensory inputs. It's not completely ridiculous to think we could encode certain kinds of knowledge in more compact forms, and transmit them via an artificial sensory process that has higher higher bandwidth that audio/visual language based learning...
But computation doesn't improve the rate of computational progress. That's achieved by improvements in manufacturing. Computation does help a little, giving us better tools for chip design. But mainly it's manufacturing.
The only way for computational progress to exponentiate, past the limit of Moore's Law, is to use more physical resources - computers get literally bigger.
And so expensive that only the five richest crowned heads of Europe can afford them!
As long as it's people developing computing devices, that holds. But when you have a machine that is capable of improving its own software, you get a self reinforcing loop where computation improves the rate of computational progress.
If Moore's law where unlimited, such an AI would have infinite computational power in 36 months, with absolutely no increase in actual intelligence. But of course Moore's law is limited. Still, we could expect quite a bit of progress from a mere AI hardware designer.
Perhaps I just mistook your meaning - what do 'subjective months' mean? Why do they mean that the doubling time starts halving?
When hardware is twice as fast, the AI works twice as fast. In 9 months, it will do 18 months worth of work. Now keep the doubling up, and it will have done infinite work after 36 months.
Edit: also, from my understanding, the gains in transistor density largely depend on new discoveries in physics. Before a computer can aid in accelerating Moore's law it would need to be sufficiently advanced to generate new discoveries in physics, but at this point you'd have a computer smart enough that it Moore's law wouldn't matter much. Seems like putting the cart before the horse to me.
Now I did not assume "Moore's law" would accelerate. I assumed it would stay constant. The key point comes from the fact the AI (or AI hive) would trivially benefit from hardware speed-ups.
And yes, Moore's law won't really count, compared to the rest we will be able to do. I was just trying to be as conservative as possible. (Though Moore's law still holding until strong AI is quite wild).
The kind of design isn't an NP-complete problem but it certainly could be helped by a powerful AI.
Also, Moore's Law is breaking for speed but not for transistors-per-inch. That's why we're seeing multicode chip now.
If the rate of improvement were constant, the progress would be linear.
In short, something is improving the rate (but it probably includes more designers and better methods, not just more complexity management).
Material scientists , physicists, chemists, chip designers are all using software for modeling and exploration. And bigger/faster/better computers are helping drive innovations back into computer manufacturing. The process is feeding itself, and it has been for awhile.
I'm not arguing that the singularity is near either, but machines would have the advantage of creating offspring that are already as knowledgeable/intelligent as themselves. Something I couldn't do for my children.
I remember computers of 30 and 40 years ago. Today everyone in the world has free access to supercomputers that translate languages and search petabytes of data (aka Google). We didn't need to revolutionize basic physics to get here from punched cards.
Expect a similar revolution in the next 30 to 40 years.
I think Kurzweil is mostly the one who keeps going on about literal unending exponential growth. Vinge's original singularity idea was more about things just becoming irreversibly incomprehensible for regular humans, since ultimately limited or not, self-improving intelligences are going to end operating on a level very distant from human thought.
Given this, I'm not sure why singularity promoters give that much of a shit whether people believe it. I'm certainly skeptical of it. Given I pretty much use a 40 year old text editor to write in a language that's about 20 years old and is just about approaching the discoveries that the Lispies made back in the 50s, I'm not sure what progress we're talking about. I write code that basically instructs a computer with pretty much the same level of semantic complexity as one does when giving instructions to a particularly stupid child.
But perhaps I've got this wrong. Perhaps the singularity will happen. I don't understand why the singularity promoters are so angry about people who are skeptical of their claims. While I was doing the rounds on Wikipedia a while back, I found this image...
https://commons.wikimedia.org/wiki/File:Singularity_Deniers_...
Have a look at the image description before I cleaned it up:
https://commons.wikimedia.org/w/index.php?title=File:Singula...
Apparently, if someone tells me that a super-intelligent AI will create a brain many hundreds of times better than our own and then allow us to live forever and my reaction isn't an immediate "oh my god, you are so, so right!" then I'm a 'singularity denier', equivalent in status to a holocaust denier. If the singularity idea is pretty much bound to happen as a result of emergent technological progress or whatever, why is my doxastic consent such an important requirement for it coming about?
I freely admit I'm probably not as smart a guy as Ray Kurzweil. But let's not forget: the history of bizarre cults shows us that smart people can believe some very stupid things.
In my case that would be because the default scenario for a technological singularity is Eternal Doom. By default, if AI researchers do not pay extreme attention, the first smarter-than-human AI will have different goals than we do, and recognize us as a threat to those goals and eliminate us all. Or squash us like bugs without even paying much attention. Or use our constituent atoms to maximize its goals. Or keep us drugged and lobotomised in soft cages, so that we're safe and happy. Or something.
And of course, we won't be able to stop it, because it will outsmart us at every turn (we can do the same to chimps).
This is why this singularity thing is at least worth looking into. Being able to nuke away our civilization is bad enough, but this could be even worse. Or it could turn our world into a paradise.
And of course, with more believers, there will be more donors to the organizations working on this, more researchers working on it… So belief does count, to some extent. (On the other hand, becoming angry at non-believers like a crusader is unlikely to spur positive reactions. By cleaning up this image description, you actually helped our cause.)
The safeguards need to be built into the self improving AI, no matter how 'dumb' it starts out as.
Humans came from microbes, yet we couldn't care less about them. A super intelligent being would see us the same way.
Honest question: can't this process of "thinking outside the box" be reduced to a logical process? Wouldn't a system which has ability to identify axioms, also have the ability to vary those axioms and try out alternatives?
In other words, why shouldn't better logical ability lead to better imagination?