On Artificial Intelligence
aeon.co
aeon.co
Obviously, understanding intelligence better would promote more effective directed research toward artificial intelligence. But if we can identify it (which the Turing Test is about), then it is quite possible that we can develop it -- and know that we have -- without understanding it. (And it may only be through developing it that we end up understanding how it works.)
> The argument that artificial intelligence requires understanding how intelligence works is an argument that natural intelligence requires Intelligent Design.
I think that statement makes sense when phrased that way, so it is an attractive idea. However, I don't think it's true. From an evolutionary standpoint, biological intelligence developed naturally because biological components are natural. Furthermore, machines do not develop when left in isolation, while biological organisms do. If you leave a large population of simple machines running in an environment, this is overwhelmingly not likely to result in a machine intelligence millions of years later. Machines were developed by humans and do not develop in the same way; comparing the two as in your Intelligent Design argument doesn't make much sense.
I do not think artificial intelligence can arise naturally because its components are not natural. This discussion could also foster a discussion on two other interesting questions:
1. Does intelligence require organic components that operate in a deterministic way (i.e. the brain) to transcend it from merely a "machine"?
2. If intelligence requires biology, where do you draw the line between creating an intelligence through natural human reproduction and creating an artificial intelligence another way?
Personally, I believe artificial intelligence does not require biological components, and I believe that under certain circumstances, it could develop unintentionally from a relatively advanced computer, but that is not the same as naturally.
Biological organisms are machines.
Entities which meet the necessary requirements for Darwinian evolution (which are, approximately, that they can pass on their traits from generation to generation, have a source for variability, and have selective pressure in the environment) change over successive generations, and all kinds of machines change individually over time in response to interactions with the environment (in fact, pretty much all physical objects do.) Biological organisms obviously meet the requirements for Darwinian evolution (since that's where it was first observed and described), but there's no magic "biological sauce" required.
> I do not think artificial intelligence can arise naturally because its components are not natural.
The whole natural/artificial divide is unsound, since humans, and hence all of their products, are part of and products of nature. Nothing not-natural exists.
> Personally, I believe artificial intelligence does not require biological components, and I believe that under certain circumstances, it could develop unintentionally from a relatively advanced computer, but that is not the same as naturally.
The whole thing about artificial intelligence arising naturally is a strawman you've constructed. You've just agreed with and illustrated my argument -- which is that artificial intelligence does not require understanding intelligence.
No, I think we're arguing two different things here. I don't mean machines in the sense of a deterministic mechanism, I mean machines in the sense of a non-biological computer.
Machines do not have generations. Machines are not alive. I can simplify this: we do not yet have computers which are alive, thus they fundamentally do not operate the same way humans do in terms of evolution. This circles back to what I said about leaving machines alone to develop in an evolutionary sense - they won't, they'll die.
Machines do not yet improve themselves after a certain "age" of maturation.
The natural/artificial divide is not unsound, because artificial means something which is created by a human, and natural means something which is not. There are definitions of these terms and philosophical schools of thought that make this divide unsound, but colloquially, I don't mean those in this context.
Here is the basic point I'm trying to make - human beings arose naturally with no apparent intelligence to guide them into existence. Machines did not do so. You can argue they did because "everything is natural", but that's not my point here. They are fundamentally different from human beings and only exist because human beings existed first.
You can't compare the relationship between an Intelligent Design and human species and the relationship between machines and humans for this reason because they are the opposite. Humans apparently have no guiding force that opted to create them or guide their existence, whereas machines do - us.
That is what I mean by natural. I think an intelligence can arise naturally, but only if it follows the natural conduits that would form conventional intelligence, which is via biological organisms.
In order to replicate what happened through evolution using a completely different "container" if you will, you'd need to be able to understand intelligence completely, or at least enough to implement it.
Ultimately, intelligence never arises from non-organic components (at least not on Earth). To make it do so where it would not ordinarily happen is what I mean by artificial versus natural. And to do that is to essentially reverse-engineer intelligence itself, which would require understanding it.
I apologize if I'm still not being clear, but does this make my point any better?
We do, in fact, now have machines -- in the software sense -- that have all features necessary for Darwinian evolution to operate. We don't yet have that for hardware devices, but we're fairly close, and we certainly don't need to understand intelligence do build such self-replicating hardware/software systems (though the universality of computation suggests we don't need hardware/software systems, since anything they can exhibit, pure software systems can as well.)
The problem isn't that you haven't explained yourself well, it's that your argument rests on a fundamental distinction between natural biological organisms and all other machines which does not exist.
Of course, first we need to define "alive", and then we should ask: can we build something that will be "alive"?
Do you consider the entities in Avida simulation to be alive? How about biological vs computer viruses? Is there a fundamental difference between them?
If we simulate a biological organism on a atomic level, together with its immediate environment, so that it behaves exactly like its real-world counterpart would, do we call it "alive"?
No. They're not.
First rule of dumb people not understanding evolution:
You can be lucky, you can be crappy, you can have some terrible design flaws (hello mammal eyes), but it's enough as long as you survive. Also, evolution works over POPULATIONS, which means that if 10% are ninjas, and 60% are the bog standard model and 30% are the crappy models - if your environment is 'fitted', YOU ALL SURVIVE.
Yep, even the 30% crappy ones!
Hands up who doesn't understand the difference between ADaption and ABaption!
Selection only occurs on a massive time-line or when environmental changes occur or there's a predator/prey war going on. And even then, the crappy crappy 10% can get lucky and still breed. That's why ecology / evolution is fitted over POPULATIONS, not individuals.
AI doesn't work like that, at all (once you get into the real stuff). All AI works on limited models (at the moment).
>Hint: this guy is a muppet, as shown by his attempts to "downvote" counter-arguments.
HN: land of the muppets.
>>On a more serious note, anyone actually working in the field of AI development knows this already. HN proving that... it doesn't. #Dragonswhoarenotdragons
You don't have to understand how intelligence works to create a machine with intelligence if you can create a machine with the same or similar evolutionary forces that worked to create intelligence in humanity and want to wait long enough for it to happen. This is a totally viable approach to creating an intelligent machine, as you can certainly create a machine that can mutate extremely rapidly.
The problem is, that doesn't help us figure out what intelligence is, unless the process of watching this happen somehow gives us insight in the matter, or we manage to create a cooperative machine intelligence superior to our own that's more up to the task.
You're applying evolutionary forces to a simplistic binary plateau.
This isn't intelligence, it's a subset of maximal solutions to bounded problems.
Hint: your dog bounds you as much as you bound your dog. Your dog has some input into your modern homo sapien sapien mind. Same goes for your atmosphere, your gut fauna, your planet's iron core and so on.
Your models cannot even touch this at the moment.
I think intelligence may be composed of several sub-modules, the combination of which, working in concert, produces the sought-after effects people often talk about when they talk about the Turing test.
I do think tests which propose to pin AI on human characteristics is not as useful as an investigative avenue as it could be. I think the first real test of AI is the identification of causal factors in a phenomena. I think another major point of intelligence would be if AI could perform arbitrary analogical mapping. I think all of math is arbitrary analogical mapping, where you start with a set of capricious but useful building blocks, build a big structure and then... analogically map the math onto a phenomena.
I think these two ingredients makes for the kind of mental abilities that people have been craving for in AI, abilities like a computer developing its own software to use hardware, or a computer which models and reasons about phenomena.
Essentially, the idea is that "creativity" is the act of sampling from a distribution over the kind of thing you are trying to create. Learning algorithms like Boltzmann Machines will learn a distribution over the inputs they see. One thing you can do with such a distribution is checking the probability of a given input under it, which may be good to classify them. Another thing you might want to do is generate representative samples, i.e. generate an example E with probability X iff P(E) = X under the model. The latter is what I would call "creativity".
Under this definition, creativity depends on both the learned distribution (which should only assign high probability to meaningful data) and of course on the sampling algorithm. As it turns out, it is very hard to write good sampling algorithms for non-trivial distributions (naive MCMC will often get stuck). So creativity is hard, but so are a lot of other tasks, so I don't think it's fair to single it out.
The prevailing misconception is that by assuming that ‘the future will be like the past’, it can ‘derive’ (or ‘extrapolate’ or ‘generalise’) theories from repeated experiences by an alleged process called ‘induction’. But that is impossible.
He is looking for solutions outside of the distribution that has been previously observed.
Theories that have relevance or potential obey a certain distribution, and it is this distribution that you are trying to sample from. Sure, theories may not be directly derived from the extrapolation of sense data, but they are nonetheless derived from the extrapolation of theories in general. So what you're looking for is not a fundamentally new paradigm, it's more like an additional level of indirection. But there's no point to experiment with multiple induction levels if we can't even make a single one work well enough.
Also, generative or inductive mechanisms have a wider scope of application than prediction. They can be used to inspect your own belief systems and pinpoint inconsistencies: the easiest way to know if your model of the world is inconsistent is to generate ideas and examples that fit the model but trigger contradictions.
But, as the author explains, epistemology doesn't work this way, and is certainly not inductive (maybe high-order inductive -- whatever that means). We don't treat physics as simply "something that works", but as knowledge based on assumptions (codified in symmetry laws) which are not inductive by any means. The laws of symmetry (assumptions, really) are a justification for induction, but can't be a result of induction alone. In fact, all of mathematics is a set of justifications for inductions that humans have developed.
But you don't need to prove them to use them. Many people go around believing unproven theories. In the sciences the preference is for verified theories(and theorems in math) and we prove them by deduction.
But it seems that you know that. I don't understand why you don't like induction in a general algorithm. We don't have to restrict ourselves to a single type of reasoning. Induction, deduction, abduction are all valid and used by humans for generating new knowledge.
I might be misunderstanding something in which case please correct me.
The article says that current research focuses on achieving intelligence by means of induction alone, but induction cannot explain all of intelligence, because we reason in ways that contradict induction (although maybe they're a result of higher-order induction).
> The Skynet misconception likewise informs the hope that AGI is merely an emergent property of complexity, or that increased computer power will bring it forth (as if someone had already written an AGI program but it takes a year to utter each sentence). It is behind the notion that the unique abilities of the brain are due to its ‘massive parallelism’ or to its neuronal architecture, two ideas that violate computational universality.
But I don't think he's done a very good job of supporting his assertion that thinking (in the AGI sense) is a computational process. The closest he comes is:
> But that’s not a metaphor: the universality of computation follows from the known laws of physics.
That's it? Because physics?
The center of gravity far enough should make them sattelites, if they were made from something that can withstand the forces involved?
That's very much applicable to AI. As an illustration, when we started to use GPUs for training neural nets with the exact same algorithms as a decade earlier but with an order of magnitude more parameters, we got dramatically better results (see Ciresan et al. 2010 [1]).
Building the same dumb skyscrapers but making them 100 times taller might in fact get them to fly.
So that's a space elevator. But can a space elevator be said to be "flying"? It stays above a single spot the whole time.
First of all before a (hypothetical) "skyscraper" could fly it would have to have its top beyond geostationary orbit. How so? Because a skyscraper is geostationary and if its center of mass is below geostationary orbit then it's simply not orbiting. It doesn't have enough kinetic energy.
Secondly, there's no law that an object revolving around earth must do so at "orbit" speed/altitude. You and I are perfect examples of that: we're going far too slow. Only consequence of that is that earths gravitational pull must constantly be counteracted by a supporting force. If an object were to revolve around earth at a speed faster than "orbit" that's also fine, as long as there's a force keeping it down.
So, no, this "skyscraper" could have its top well below geostationary orbit as long as it rests firmly on mother earth. Once its center of mass is beyond geostationary orbit though it would need to be held down.
How many times do we have to hear these arguments to realize they are just hot air? Either computers are capable of intelligence, or they are not. The answer to that question depends entirely on how you define intelligence. If you define it as the set of things humans are capable of and computers are not, then the answer will always be no. But just like humans, computers learn and are taught new things every day, and as time goes on the set of things human's are uniquely capable of grows smaller and smaller.
> Despite this long record of failure, AGI must be possible. And that is because of a deep property of the laws of physics, namely the universality of computation.
> [Turing] concluded that a computer program whose repertoire included all the distinctive attributes of the human brain — feelings, free will, consciousness and all — could be written. This astounding claim split the intellectual world into two camps, one insisting that AGI was none the less impossible, and the other that it was imminent. Both were mistaken.
etc etc etc. He literally spends more than 60% of the (very long) essay arguing against your "summary".
Anyway, a real tl;dr is right at the top of the page: "Expecting to create an AGI without first understanding in detail how it works is like expecting skyscrapers to learn to fly if we build them tall enough." and then the last sentence: "it is plausible that just a single idea stands between us and the breakthrough. But it will have to be one of the best ideas ever."
In other words, AGI is provably possible, but the author believes that we're going about it all wrong (behaviorist-inspired neural nets running on training sets, etc.) and need a philosophical (specifically: epistemological) breakthrough to move forward.
Philosophers can not seem to comprehend how a mass of meat can create consciousness and take leaps to try to rule out the mass of meat, regardless of the fact that they all know that when bad things happen to said mass of meat bad things also happen to said consciousness.
The Turing Test could not have been a more wrong direction to define AGI. Looking like something does not make you that thing, and it did not at all address capability, just the appearance of capability.
Insight into intelligence will come slowly, and it will come from examining, categorizing, and eventually understanding the internal workings of the human brain.
> Philosophers can not seem to comprehend how a mass of meat can create consciousness and take leaps to try to rule out the mass of meat, regardless of the fact that they all know that when bad things happen to said mass of meat bad things also happen to said consciousness.
You seem to be referring to dualism, a philosophical idea that is largely discredited among philosophers since the early 1900's. Functionalism, emergentism, materialism, and all other leading serious philosophies of mind that implicate the brain were first formalized and studied by philosophers, not neuroscientists.
> The Turing Test could not have been a more wrong direction to define AGI. Looking like something does not make you that thing, and it did not at all address capability, just the appearance of capability.
And philosophers are among the most vocal critics of the Turing Test. The Turing Test was made by Turing, who was first and foremost a mathematician & logician, not a philosopher of mind.
> Insight into intelligence will come slowly, and it will come from examining, categorizing, and eventually understanding the internal workings of the human brain.
Which, of course, is what modern philosophers of mind do.
Then what did philosophers do? Well they applied the Chinese Room argument to all of conciseness, a blunder even worse than the Turing Test.
Neither can anyone else, so far.
> Insight into intelligence will come slowly, and it will come from examining, categorizing, and eventually understanding the internal workings of the human brain.
Insight -- yes. AI -- not so certain. Here's why: it's plausible that intelligence in the human mind emerges as a consequence of an intractable process. We might not be able to understand that intractable process. But suppose that instead of understanding it, we mimic it: we build a replica of the human brain, and it works! But this is not quite artificial intelligence but a replica of natural intelligence. This is not just a semantic difference: even if this mimicry works we may not be able to direct it in any way. For example, we won't necessarily be able to give this brain replica super-human intelligence, without, say, giving it a mental illness at the same time.
But most AIG researchers don't throw their hands up and say, "Oh damn, it's impossible (intractable), it's magic because we don't understand." Forgive me if I feel this a less-than-useful sentiment.
As far as the possibility of intelligence being intractable, what are the chances of this. Is it likely? If we don't even know what AIG entails, how would we possibly put odds on this? If we don't have the odds, what use is there is thinking about it? Should we stop researching AIG because of the possibility that it might be too much for us to handle?
Neither does the author of the article.
> As far as the possibility of intelligence being intractable, what are the chances of this.
I think the odds are very good. Every known complex system -- the weather, for example, or an ant colony -- is intractable, let alone the human brain.
> Should we stop researching AIG because of the possibility that it might be too much for us to handle?
Intractability doesn't mean it's too much for us to handle. We can simulate intractable processes, and we do it all the time (that's how the weatherman knows if it's going to rain tomorrow), so of course it can be studied. The problem with intractable processes is not that they can't be artificially created, but they can't be artificially controlled (or predicted) beyond very short time frame. So it's very possible we could build an AGI, but won't know how to make it any smarter than us.
Make it as smart as the smartest of us and then it'll figure out how to become smarter on its own. Then ask it to explain to us how we work.
Tongue in cheek but once almost-as-intelligent-as-a-human level is reached the next step is not that far fetched - quantity, multiple times faster than humans learning and we have progression on a different scale. Or we can always try mutating the thing to see what comes out - crude but worked at least once.
But this doesn't necessarily follow. The experience of a mind working at a higher speed will be a slowing down of time -- imagine what you'd do if time slowed down. The machine will not necessarily learn more (we might have a limit of how much we can learn), and will probably experience boredom and frustration (as everything around it will be slow) that might drive it crazy.
tl;dr: AGI is possible, but we're going about it all wrong and won't create it until we philosophically understand it.
No, he doesn't say that. On the contrary he names the principle ('Universality of Computation'; see paragraph 4) which guarantees that computers are capable of true intelligence, since, if programmed correctly, they can simulate the behaviour of any physical object, including human brains.
>there is some divine truth and wisdom that only humans possess.
He also explicitly repudiates supernatural explanations (see para immediately before the one mentioning John Searle).
Why is it so hard to remove our self from that equation? Why should intelligence be a skill only humans acquire? It is not only about machines: There is a constant stream of papers revealing animals are not that dumb as thought either. We survived recognizing we are not the center of the universe, may be we are also not alone on the top of the intelligence pyramid.
Latest edge question "What do you think about machines that think?" provides a good and very broad overview of many aspects. I go with George Church: "What do you care what other machines think?"
http://edge.org/responses/what-do-you-think-about-machines-t...
I personally prefer (although it's certainly less measurable) Ben Goertzel's definition - the ability to achieve complex goals in complex environments. It's still wishy-washy enough for people to write off any given achievement of AI though, I guess.
I think Deutsch's argument is that there is also creativity (and creative reasoning) that is omitted from today's approaches to AI.
This is where I gave up. Deutsch dismisses thousands of years of thought in a sentence. Not to mention that "justified true belief" is a phrase you find much more often in a textbook or an encyclopedia article than in a real work of philosophy.
The example of years that started with 20s seemed quite odd. Given the easiest way to understand numbers at least for me is inductively. Of course if you lop off a bunch of information such as the digits after 20 or 19 it sounds like an impossible problem.