Superintelligence: the Very Idea Wednesday, Oct 3 2007
singularity and superintelligence 6:08 am
I’m interested in finding people to go on record, or who have already gone on record, as saying that superintelligence (smarter than human in every field) is impossible. I’m sure there are quite a few such people out there. We can use Nick Bostrom’s definition for a superintelligence:
“By a “superintelligence” we mean an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.”
Based on a scholarly analysis of the complexity of the human brain and improvements in computing hardware and our understanding of intelligence, Bostrom came to the conclusion “we will have superhuman artificial intelligence within the first third of the next century”. Meaning, by 2033, Nick Bostrom believes there will likely be AIs smarter than human brains in practically every field, including:
- common sense
- humor
- artistic creativity
- imagination
- scientific knowledge and capability
- deriving solutions to political problems that please everyone
- inventing better political systems no human would never think of
- finding ways to get clean water to third world countries
- developing cures to diseases such as AIDS, cancer, etc.
- ability to visualize complex shapes
- ability to visualize complex systems
- ability to invent new shields and weapons
- ability to shock and surprise others with its intelligence
- ability to put human mathematicians, dancers, and philosophers to shame
- invent gadgets to help you get past daily annoyances
- etc…
Because such an AI would presumably run on a supercomputer of the time, it could then be copied, probably at least a few thousand times, into additional supercomputers. Then the superintelligence could develop new methods of computing for spinning off alternate versions of itself, building new superintelligences from scratch, or changing human beings into superintelligences. By developing more powerful computers, mass-producing them, shielding and cooling them, putting them in custom-built bodies, distributed systems, or whatever, presumably you could have millions or billions of such superintelligences running around the planet a decade or less after the first one is created.
Is this “the Singularity”? No, because the word Singularity has lost all meaning and is all but useless. It’s a word you drop to sound cool around science fiction nerds. People who are serious about communicating refer to specific scenarios and assumptions, so that others can have some glimmer of an idea of what they’re talking about. I only use the word “Singularity” so people will find these posts when they Google it, and just maybe be convinced to stop using it.
Anyway, what I’m interested in finding are people that not only believe superintelligence won’t be possible by 2033, as Nick Bostrom has argued, but that it won’t be possible in 100 years, or 1000 years, or ever, because humans are fundamentally the smartest creatures that can physically exist. And I hate to always bring Christianity into this, but I think the Church is a key reason why people believe superintelligence is impossible. Most churches teach that man was made in God’s image, presumably meaning man is about as awesome as any being can get. But what about…
- human brains engineered to have faster metabolisms?
- human brains augmented with cultured cortical tissue?
- human brains interfacing with neuroelectronics?
- artificial intelligences based on the human brain?
- artificial intelligences based on Bayesian learning algorithms?
- extraterrestial intelligences anywhere in the universe?
Naysayers of superintelligence would argue (or would they?) that there is a glass ceiling of intelligence that presumably halts at around the human level. Or, they would argue that smarter-than-human intelligence is possible, but it would only excel in fields traditionally associated with “braininess”, like playing chess or doing mathematics, but never artistic creativity, spiritual insight, charisma, imagination, wisdom, etc. This artificial separation between the former and the latter is quite odd to me, as it seems like all these qualities originate in the brain. But again, more Christian theology arrives on the scene to muddy things up: priests often say that these latter qualities originate in the soul, not the brain. But then why does blood circulation in certain brain areas light up when we paint a painting, or talk to people at a party? If the nerve impulses that direct our arms to engage in dance do not originate with the brain, then where do they come from?
And by the way, I myself am not particularly attached to the 2033 date, although I do consider it reasonable. I’m more interested in the debate over whether superintelligence will be possible ever… not whether or not it will be possible in our lifetimes, for instance.
Do all transhumanists believe superintelligence is possible? If so, why is there so much focus in transhumanism on bodily augmentation, life extension, nanotechnology, and biotechnology, but very little on superintelligence specifically? When transhumanists think of “superintelligence”, are they mainly thinking “Einstein”? Or do they think that intelligence enhancement advances will happen slowly, allowing us to raise only one IQ point per year rather than 1000 at once, and therefore it’s really not the juicy fruit that some think it is?
I’m not necessarily a superintelligence naysayer, though I am a bit skeptical. I think there is a distinction that should be made between “superintelligence” and “faster intelligence.” I think the latter is certainly likely (after all, it already applies to things like basic math), and maybe that’s enough to qualify as the former to some people.
But, for example, I don’t think there will ever be such a thing as “superhumor,” even if there are many varieties of humor. I don’t think one can have “more sense of humor;” either you find things funny or you don’t, and simply finding more things funny doesn’t make you superhumorous — in fact, it might mean you’re simply a childish dolt with unsophisticated tastes.
So I guess I might fall into the “traditional braininess” category. If we upload people and run them at 1000x speed, then to the physical folks left behind their research and development may well seem “super,” even if the uploads themselves are subjectively just doing the same thing they always have. But I’m skeptical of the feasibility of making something qualitatively “super” rather than quantitatively.
Some Christians, such as I, interpret the idea that humans were created in the image of God to mean that humans should seek to become as we understand God to be, in intelligence, capability and compassion. Furthermore, Christians with thoughts similar to mine already recognize our own intelligence as proof-of-concept. Christian faith is not necessarily incompatible with faith in artificial general intelligence — but, of course, some Christians want them to be incompatible.
“I don’t think one can have “more sense of humor;” either you find things funny or you don’t”
Error: The word “you” indicates a human, and superintelligences will not usually share human qualities. You use “you” in the generic sense, to include all possible subjects (see http://en.wikipedia.org/wiki/Generic_you); this will not work, as it includes both humans and AGIs, which are more different from each other than a giraffe is from a potato.
Even if you couldn’t have a superhuman “passive sense of humor” you could certainly have a superhuman “active sense of humor”, meaning simply that you could easily make up stuff people found extremely funny.
“priests often say that these latter qualities originate in the soul, not the brain”
Do they?
I think superhuman intelligence is certainly possible. All you have to do is find out how the human brain works and recreate that structure, but *better*. There’s the rub though… how DOES the human brain actually work on a fundamental level? You can tell me about axons and dendrites and inhibitors all day but we know essentially nothing about the actual underlying (mathematical?) mechanisms going on. A lot of this is because the brain is an evolved piece of machinery, and evolved systems tend to be ridiculously complex and chaotic. It’s not like a computer processor where things are laid out as simple as can be, simplicity isn’t a concern at all, so you get really really weird things going on. Will we ever be able to know (let alone recreate) precisely how the brain works? I don’t know. We may be able to create a system that works in a similar way, though I’d say our best bet is to use true evolution to design our circuitry. Many engineers will be resistant to this because it undermines the very idea of engineering, but I firmly believe its our best bet. The goal of Artificial Evolution is far more important than Artificial Intelligence.
Modern computers are very nice when playing chess, but they work on a very fundamentally different system than the brain, even when doing logical exercises like solving equations or playing chess, and especially when doing artsy fartsy stuff. TRUE artificial intelligence, I believe, must be more than a normal computer, but faster.
Do all transhumanists believe superintelligence is possible?
I doubt most people calling themselves “transhumanists” are even transhumanists.
Modern computers are very nice when playing chess, but they work on a very fundamentally different system than the brain, even when doing logical exercises like solving equations or playing chess, and especially when doing artsy fartsy stuff. TRUE artificial intelligence, I believe, must be more than a normal computer, but faster.
Hey. 1985 called. They want their rebuttal back.
I think that the idea of super intelligence has some merit, but I think that the timeline as stated is incredibly optimistic. I don’t believe that SI is impossible, I believe that it is inevitable – it’s going to be created. However, the arrival date isn’t going to be any time soon. I hear a lot about when SI (and many other things) are going to be invented but I don’t hear the reasoning or evidence to support these assertions. Anyone can pull a figure out of the air but getting a prediction of the future correct is notoriously difficult.
SI (when it arrives) won’t be limited by biology the way that organic intelligence is (since intelligence has a reasonable correlation to brain volume), so this factor alone leads to the conclusion that SI will definitely exceed what we know as genius today. Whilst a brain that is twice the size may not necessarily be twice as good, it is a reasonable assumption that removing the physical constraints will result in a better system.
When SI comes, I don’t think that it will be made as much as evolved. We simply do not understand what consciousness is, and whilst we make many strides forward in technology our understanding of how our brains work remains primitive. However, we can definitely make an environment where SI can evolve, we live in one and we understand to a large extent how it works. We may never be able to assemble a working intelligence but we can culture one in exactly the same way we culture bacteria and breed animals – we can steer the process in a direction of our choosing.
I am a big fan of biologically based sociology, I tend to believe that behaviours evolve to fit niches. Humans are smart because that helps us to be successful and reproduce, thereby passing on our genes and continuing the cycle. But what pressures will drive the SI mind to grow? Why does it need to be smart or creative or humorous? I think that there is a real possibility that SI is likely to be very alien to us. It doesn’t take a lot of imagination to see how different such minds could be. At the heart of thought is the understanding of concepts, would we understand the concept of individuality if we could be copied exactly, would we understand gender if we had none, would we understand illness if we could never be sick, would we understand death if we never grew old? These questions are both fascinating and disturbing.
An SI’s level of friendliness is going to be wildly unpredictable, we won’t know if we can trust it until it would be too late to stop it from causing great harm. It, by it’s very nature, would be smart enough to fool us. Not enough of the negative possibilities of SI are being discussed in my opinion. What about an SI that is angry, or careless, one that gets bored, or lies, what about one that is a fundamentalist that believes that God’s on it’s side? What about an SI that can kill? If we can do it, then so can they. How many of us have burnt ants with a magnifying glass as children? Did you worry about they felt? Of course not. An SI has the potential to be an extinction level event for us.
Despite my reservations, I think that ultimately SI will exist, and I think that the best insurance we have against the potential risks is to ensure that we can integrate SI into ourselves (or is that us being integrated into SI?). If us and them becomes just us then we will be only one species and have no reason to be in conflict with the other.
> Do all transhumanists believe superintelligence is possible?
I do, and not in the mundane Einstein sense.
> If so, why is there so much focus in transhumanism on bodily augmentation, life extension, nanotechnology, and biotechnology, but very little on superintelligence specifically?
Because SI is probably quite a bit further away than the above current and near-term technologies.
I don’t think superintelligence is possible.
Why? Because we just don’t have the hardware for it. Clock speeds have shown no significant increase since 2002, and the computer industry seems more interested in developing chips that need less power and run cooler than in developing chips that run faster.
Also, we’re running into fundamental barriers. You can’t make information travel faster than the speed of light, nor can you make transistors smaller than an atom. Whatever few years left of moore’s law will not be used to make faster processors, it’ll be used to make those chips run cooler and consume less power.
Sure, single threaded performance will still see some minor improvements due to more RAM and cache size, but it won’t be anything significant.
For the record, I don’t belong to any organized religion. Wake me up when clock speeds go above 5GHz, and I’ll be more likely to believe you.
Stuart, very interesting point.
I believe that Super Intelligence will be achieved and I am OK on the 2033 time-frame.
But before we deal with SI we will have to deal with near-SI. What would that be like?
- It could be an AI as good as humans in some aspects but not all, e.g. “common sense + creativity + Turing test” but zero social skills. Imagine the kind of conversation you could have with this one!
- I think most of all it would be DIFFERENT from our intelligence. I mean, there is no chance we would achieve human-like intelligence straight away. So before we get there we will have different-from-human intelligence. The differences between male/female thinking or kids/adults or occidental/oriental thinking are nothing compared to the difference between us and our first achieved strong AIs.
So I believe that the 2020-2030 period is going to be very strange, full of hopes and fears, of uncanny experiences.
And this will emerge not only in labs, but everywhere on the Internet, with companies struggling to get their newly designed strange AI out there…
Naysayers of superintelligence would argue (or would they?) that there is a glass ceiling of intelligence that presumably halts at around the human level.
I would argue that “human” is a type, not a level.
In order to determine whether an artificial intellect was truly “smarter” than a human intellect, would you have to make sure this intellect pursued numerous different fields of study and became equally proficient in all of them? How would this be tested? And who would judge?
Hi Mike, see this:
http://en.wikipedia.org/wiki/Megahertz_myth
Anne, I have confidence that one day a quantitative, inter-species measure of pattern-recognizing horsepower will be developed, although this will miss specializations that intelligences may have in any given area. It will have to be based on how efficiently the intelligence is able to extract predictively useful patterns from perceptual data. But the science and theory is not there yet.
Looking at human intelligence as a “level” is useful to drive it home to people that we aren’t the most capable type of intelligence that can exist. I.e., a superintelligence could conceivably kick all our asses. People are uncomfortable about visualizing this possibility and therefore will come up with any way they can to slither out of recognizing it. I prefer to use “level” to discourage them from doing so.
We should distinguish between moral worth and pragmatic capability. If one intelligence has greater pragmatic capability than another, then I’m inclined to call it “higher”, but this does not mean I consider it morally better. Sort of like how an atom bomb has much greater firepower than a conventional bomb but this doesn’t make it morally better at all.
In the real world, unfortunately, the entity with greater power and intelligence gets to set the rules. Our objective should be to guide the creation of superintelligence such that it can have power without abusing it.
“But then why does blood circulation in certain brain areas light up when we paint a painting, or talk to people at a party?”
A priest or christian would argue that the soul make use of these brain centra in the process i.e the soul relies on the brain to function.
Michael: “inventing better political systems no human would never think of” is a double-negative. Take the “n” off “never”… ;)
Stuart is right in the sense that it is reasonable to conjecture, based on current trends, that AI and IA will converge and coalesce, probably around the “usual suspect(s)” timeframe, i.e. 2018-2030. This if handled correctly, will tend to insure “Friendliness”. But, again, as I’ve emphasized before, this is NOT to say that we (which is, practically, to say Ben, Eli, and a few others) shouldn’t still strive mightily in terms of coming up with a viable theory of Friendliness (and a means to implement it), and then proceeding to impelement it in a seed AGI. That is to say, we should still definitely proceed along the lines of creating an independent (independent of co-evolution of IA) AGI that does indeed embody (instantiate) Friendliness.
Ciao…