All this talk of 'intelligence' gets confusing - it seems like the core distinction people are talking about most of the time with 'general' or 'strong' AI is really something more like Artificial Consciousness.
This has issues too - if consciousness is an emergent property of a neural net with the right feedback mechanisms and training material then even if you have the right feedback mechanisms you could still create an artificial consciousness that's stupid.
You see this problems in humans - depending on a lifetime of 'test data' exposure (parents, peers, environment) and the underlying brain neural net 'hardware' you can get people that believe a lot of stupid things.
Maybe consciousness doesn't have to work that way and we're just dealing with a local maxima of evolution (or some reproduction/sex drive constraint), but we might end up stumbling on the ability to create a neural net with the ability to emerge a consciousness before we can craft the type of consciousness we'd want.
Actually understanding how the system works is harder.
Intelligence exists in human culture, not in individual humans. A high IQ has limited usefulness without a culture it can learn from and operate in. (Ask any octopus.)
Human culture persists outside of individuals, so the occasional genius can change culture and culture stays changed because collective memory is altered. This doesn't happen with individual animal intelligence.
So I don't think we'll ever see a true AGI. What will happen instead is that collective cultural memory will improve, and connections between human brains will become faster and deeper - just as they have done for millennia.
If there's ever a singularity it will be created by human brains operating together in a new and vastly accelerated shared state, not by silicon brains that have been programmed to appear sentient.
Your first point is what I was trying to get at - basically the culture is the 'test data' and IQ would be the capability of the 'hardware' or genetic make up of an individual brain/person in this case.
I think your point about intelligence not existing in an individual is too narrow, but ultimately that comes down to how you decide to define intelligence.
Not sure I agree that AGI won't happen - in fact I suspect the computing power we have now is good enough to do it, but we're just missing some insight with how.
stupidity is typically correlated with low ability. If hardware insured that ability is high, then we wouldn't get a lot of stupidity from emergent consciousness
Multiplying 3842 by 543 is very hard for humans. They're slow and inaccurate. Computers do it perfectly at incredible speeds.
Inventing multiplication is something no computer (as we currently understand them) would ever be able to do.
For another analogy, moving ten tons of rock a thousand feet is something humans can do, and have done for millennia. It's very slow and difficult. A bulldozer can do the same thing in minutes. But a bulldozer would never, ever have a reason to move ten tons of rock.
A lot of mathematics doesn't correspond to anything happening in the known universe. Mathematics describes infinities, infinitesimals and uncomputable numbers that are only hinted at using symbols which are not to be confused with those numbers themselves.
That's what Lenat himself argues in his follow-up paper. But I generally agree, there's nothing preventing computers "as we know it" to exhibit human level general intelligence. Humans weren't just smart enough so far to program it yet.
> Inventing multiplication is something no computer (as we currently understand them) would ever be able to do.
AM, the discovery learning program developed by Doug Lenat, discovered multiplication, beginning with a few basic facts about set theory, and using about 200 heuristics.
I don't consider this an example of discovering multiplication. The set theory rules were given to the computer. Can a computer be built to start only with sensory input and then discover the rules of set theory? Would it ever be able recognize why the patterns and rules for operating with sets are important and worth further study? I'm skeptical.
Would a program flag the discovery of multiplication as important because it understood that this discovery is interesting or would it flag its discovery as important because the programmer encoded it to have this understanding? That is can a program discover something new and alert to humans to this discovery and for it to be truly interesting and new?
Absolutely. Look at Pat Langley's BACON program back in the 80's. He didn't work on multiplication per se, but given nothing but "sensory input" it discovered laws of planetary motion, rules of physical chemistry, conservation laws, etc.
I looked at a summary of his work. I've never heard of him or of the BACON program so my knowledge is essentially zero in this. Looking at summary of his work it doesn't seem to me to qualify either.
I'm not sure how to convey the idea in my head to English in such a way to make it clear why I think BACON doesn't qualify as discovering new ideas. I'll try though.
From the write up, by Langely, of the program he wrote the program to detect regularity in data and if it found such regularity it would "leap out" to the program so that "it could take appropriate" action. As I understand the program it picks up on any sort of regularity. The program doesn't understand which forms of regularity are important and which are not. The programming system and data it looks at are geared to finding regularity that are meaningful. If the data inputs are vastly widened then it seems to me it would find regularity that wouldn't be considered meaningful. I can imagine the program fixating on a regularity pattern that is meaningless. I hope what I'm writing makes sense.
The work he did is interesting and I wonder what advances have been made since then.
Let me ask this question. Would a program that looks for regularity discover that in finite probability spaces probability zero events never occur? I think this is something a computer could discover. Could it end up concluding that probability zero events can occur if one looks at uncountable probability measure spaces? It seems to mean that programs will never be able to think or discover things at a meta level. Would a program be able to discover Godel's incompleteness theorem? Of course I don't know the answer. I just suspect that the answer is no. However, humans took thousands of years to discover these things. So perhaps an AI that ran for that long could as well.
Thank you for pointing me to the BACON program. I'd never heard of it before.
I'm not the one giving you downvotes. I believe you are incorrect, but you have taken the effort to explain your thinking and have the courage to admit that you could be wrong, so I wouldn't down-vote that.
In any case, let me flip the question around at you: how did Goedel come up with his incompleteness theorem? Surely it didn't just magically appear in his head? Most likely he followed some standard procedures of mathematical practice: manipulate symbols according to learned rules, look for 'surprising' features in the result, build a sample demonstration of a surprising result, and use intuition regarding its outcome to drive further thinking, etc.
You can look at any intellectual behavior and find processes underlying it, unless you want to argue that we don't live in a materialist universe and people have souls. Unless you go down that route, there must be mechanisms to thought, and indeed applied psychology agrees. So if there are mechanism to thought, why can't those run on a computer instead of biology?
Eurisko, AM's successor, did that:
AM and Eurisko used a measure of interestingness to direct their searches. Here is one of Eurisko's inventions:
From EURISKO: A program that learns new heuristics and domain concepts: "Almost immediately, symmetry heuristics produced a very powerful yet simple device, one which simultaneously computes NAND and OR, using only two small metal regions, two n-doped regions, two p-doped regions, and one intrinsic channel region."
Non sequitur. In fact, what appears to be your stated premise actually argues against the conclusion, not for it. Besides, the mechanistic characterization of mind (you say brain, but for all intents and purposes here it makes no difference) is highly problematic.
Small nit, but as hinted at by the other responses, humans didn't "invent" multiplication, they discovered it. In the same way we didn't "invent" the fact that the earth revolves around the sun, we discovered it.
Generally invention applies to the first discovers. I would suggest in the entire universe both observable or not we are unlikely to be the first species to discover multiplication. Languages like English, French etc on the other hand have much better odds even if language as a concept is really old.
PS: I have even seen 'rediscovery' used for the same situations when a math idea was lost etc.
that's debatable, once you appreciate the point about having a reason to do something.
calculation itself really only exists with purpose. it's a thing to do if you have a reason to do it. of course, these kinds of things too seem more "discovered" than "invented", but it's a very different kind of discovery than "the earth revolves around the sun". the earth revolves around the son, as far as we can figure these days, whether any agent/subjectivity/etc intends anything or not.
however, no full-blooded calculation will ever occur anywhere in the universe unless someone needs to figure something out. in other words, intention is implicit with calculation (and truth-functional representation, by the by). as such, you could make a case for an argument for "inventing" strategies in mathematics. and of course, a lot depends on your ontology of mathematics in the first place. if you're a wittgensteinian about numbers, and you think any of what i said above is at all compelling, you may be inclined to believe that mathematics itself is an invention. you may admit of some more nuanced way that entities like "number" still exist without mathematics as a tool, you may not, but you'd probably see all mathematics existing as active work, not pre-existing objective Truth.
of course, most people think of math as an always-there abstract object (roughly). maybe they're right. even so, there must be some force in the "i use it as a tool" and "it just revolves around the sun no matter what anyone--god or man or beast--wants" distinction.
Calculation can be done purely for pleasure. I have math nerd friends who sit and do math for fun. It's related to other challenging, creative pursuits, like playing music or writing poetry.
I actually don't think this is true. Multiplication isn't some natural phenomenon that humans just stumbled upon, like gravity or nuclear energy, even if it does help to explain some observable phenomena. Multiplication is entirely a human idea -- it only exists in human brains. Multiplication was only "discovered" in the same way love, or music, or any of the liberal arts were "discovered" -- which is to say, not at all.
I guess the straightforward counterargument here is that if multiplication only exists in humans' minds how can computers do it, but computers have absolutely zero autonomy -- a computer can't compute a product, or even "discover" multiplication without receiving explicit intructions from a human to do so -- in my mind a computer can't be deemed the one discovering multiplication if that discovery relied 100% on the design of human beings.
Does that really matter though? It seems a bit like not giving humans credit for their accomplishments since, after all, sexual reproduction evolved a billion years before humans even existed.
The average human couldn't invent multiplication either - it took humanity as a whole thousands of years to get to that point. Even writing was only independently invented a handful of times in history.
How many people work jobs today that are akin in nature / level of fulfillment to inventing multiplication? How many people work jobs that are best done by a machine? How much energy does it take for a human intelligence to do these jobs vs a machine? How many people are furthermore stunted in their growth due to forced submission to their middle managers, a power differential that hardly need exist but feels great for the manager (king in their little castle)?
I'm pretty sure computers, as we currently understand them, are entirely capable of inventing multiplication, provided sufficient source code and input. Evolution is unique in being able to "create" a thing that can formalize multiplication? Human intelligence is not magic.
Human intelligence is not magic, but it is categorically different from machine intelligence.
What's amazing isn't that Deep Blue could beat Kasparov, given the depth of analysis it could do. What's amazing is that Kasparov could beat Deep Blue, despite not being able to analyze thousands of possible moves a second in his head.
I think Alpha Go is more in the category of human intelligence.
My probably oversimplified understanding is that the neural network training on a massive amount of games develops something like intuition. It can't explain why a certain move it makes is the best, but only that it is the most likely to win.
It is no truism. It is merely dismissive and does not constitute a legitimate argument. There are numerous philosophers who have made arguments that challenge the notion of AI in the strong sense and who have no such interest in "preserving" feelings of human specialness that are allegedly being threatened. To lump all arguments against strong AI and computationalism in with "AI effect" is ignorant.
He makes a flawed argument that computers are smarter than humans: but the flaw is obvious in that a computer couldn't even make that argument. It takes a human.
> the argument goes both ways here as well: take an arbitrary human (such as yourself, if you happen to be human) and try placing this human in the cockpit of a landing jet plane, in a semiconductor factory, in the oval office of the White House, in the kitchen of a gourmet restaurant, on a horseback in Siberia, or equipped with only a spear in the middle of the Amazonas jungle. There are humans that have been programmed to do well in each of these situations, but it is very unlikely that the human you were thinking of (perhaps yourself) would know what to do in more than at most one of these situations.
This part seems wrong. I think most humans would make a decent go of most of those situations. Not as good as an expert by any means, but they'd be capable of doing something, unlike a computer program.
Agreed. He also subtly introduces a double-standard here, in that a person with no training or experience might fail in these situations, thus we're inferior to machines...but without the right algorithms and training data a computer would fail worse!
While your point is true, I'm not sure it is relevant to measuring intelligence. The human has two advantages in this situation: First off, all humans have agency: they can take action to do things. If you want the comparison to be remotely fair you need to compare a program that has agency to see the same stimuli as the human would and have access to the same range of actions. Not a particularly difficult programming task.
If you put an untrained system and an average person in the crashing plane situation, the average person is probably going to do better, but it's because your average person /has/ some training in what to do in this situation. They've been exposed to enough culture to know that the joystick is the main way of controlling the plane and that planes moving towards the ground at steep angles are likely to be bad, and probably know generally what action to take that might help the situation.
Fair comparisons would be say, an untrained program vs a baby, or a system trained with some general knowledge of vehicle control vs an average person. I'm not really sure that humans win out in either of those comparisons.
> First off, all humans have agency: they can take action to do things. If you want the comparison to be remotely fair you need to compare a program that has agency to see the same stimuli as the human would and have access to the same range of actions. Not a particularly difficult programming task.
Huh? I think the whole point is that that's a very difficult task.
> Fair comparisons would be say, an untrained program vs a baby, or a system trained with some general knowledge of vehicle control vs an average person. I'm not really sure that humans win out in either of those comparisons.
The human got very little direct training. They mostly got exposed to the environment and picked it up. You can give the computer program the same number of years and the same environment as the human got, and it won't do much good.
The article is informal about intelligence and then comes up with a couple of ad-hoc examples where computers beat humans. It's unclear that they have much to do with intelligence. The following definition of the term has been proposed:
Intelligence measures an agent’s ability to
achieve goals in a wide range of environments.
Togelius even addresses this a little by pointing out that humans have to be trained to be a pilot of a president. But
it is unclear at this point to what extent computers can be intelligent in this sense. Alpha-Go's reinforcement learner, probably the most astonishing part of Alpha-Go, was not (to the best of my knowledge) Go-specific, instead, it was a general-purpose reinforcement learner. I doubt it can learn much more complicated forms of interaction without a simple reward function (such as games).
Nevertheless, I'm quite optimistic, but it's far from the foregone conclusion that the author implies it it.
> But the argument goes both ways here as well: take an arbitrary human (such as yourself, if you happen to be human) and try placing this human in the cockpit of a landing jet plane, in a semiconductor factory, in the oval office of the White House, in the kitchen of a gourmet restaurant, on a horseback in Siberia, or equipped with only a spear in the middle of the Amazonas jungle. There are humans that have been programmed to do well in each of these situations, but it is very unlikely that the human you were thinking of (perhaps yourself) would know what to do in more than at most one of these situations.
At least the random human will have a chance at doing something and if the situation isn't life or death like the landing plane or the Amazon jungle could with time actually learn to operate in the new environment even without interacting with other people who could teach them. That's what's missing from AI the flexibility to operate in a chaotic environment. Until relatively recently the slightest thing going wrong in even moving through an area or performing a simple task like moving a box would completely break.
> (It's hard to understand why anyone would want to be in a plane flown by a human, now that there are alternatives.)
Detecting and ignoring spurious inputs and extreme edge cases are one of the main reasons that I feel better if there's a person alert and at the controls of a plane. Extreme cases like Quantas 32 where a huge number of systems are absolutely destroyed would be a huge challenge for modern autopilots which are great but aren't tested or designed for emergencies. [1]
That's partly because nobody has tried very hard to write one piece of software that's good at landing planes and being the president. Why not just write two pieces of software? What business model would a combined piece of software enable?
Yes. Because a specialized pilot or president will always be better than a generalist at that particular task.
Part of the argument given for the virtues of generalization is about availability. Like, if you find yourself the only available person in a plane over the Amazon, you can improvise a landing. But as the network connects more things in more places, this becomes less and less important.
I'm not arguing that writing 2 different software isn't a solution or against software specialization. I'm saying (in part) that comparing a domain specific program(s) to a generalist human and calling the strictly domain specific 'intelligence' better misses some very critical points in capability. Even a combined program that's able to fly a plane and run a factory (changed from being president because doing that job well would probably require a full strong AI which could do any other task) is still just as constrained except now to just 2 fields instead of 1.
To me it's not about the exact number of tasks that a particular bit of software can perform but that to call it's intelligence better than human it has to be able to increase it's own capability over time in novel tasks and environments. I'm also extremely resistant to equating simple calculating and data look up speed with intelligence as the author does with the multiplication and random SSN look up examples.
But what we might want is a "code base" that has some percentage of those two activities lying latent in them so we can exploit them in less time. Then again, maybe not - it may be that from-scratch works better.
We can't really tell until we do it. And the hard part is the analysis of the business model, anyway.
> Detecting and ignoring spurious inputs and extreme edge cases are one of the main reasons that I feel better if there's a person alert and at the controls of a plane
But what is more likely: a human pilot detecting and resolving problems with the computer systems, or the human accidentally crashing the plane?
I think a human has a better chance of determining which inputs are false and which are reliable. I'm also not talking about a random person there. A random person off the street would be more likely to crash but that's true even if the plan is doing just fine judging from what I've seen of youtubers in flight sims.
> It would be very easy to invent games that were so complicated that only computers could play them; computers could even invent such games automatically.
Of the many errors in this article this is one of the most flagrant. Computers have yet to invent any novel and challenging games. That would take ingenuity, something computers have failed to demonstrate. However, I suspect the author is trolling a bit.
Indeed. I would also posit that novel and challenging are relatively easy. Games are not just a set of rules but also rewarding to the players.
Games exist that are variants of "click the arrow on the coloured pixel 100 times". And people pay to play them. Good luck finding one with a Deep Conv.
i believe i've read about automatically --(i believe automatically, and it's a crucial distinction, but i'm not looking it up right now, so take it with a grain of salt)-- generated proofs so long they can't be reasonably understood by a human, but it "convinces" a computer. such a proof isn't a proof in the sense that it doesn't convince a human being, which is the point of the proof. but, since it checks out, maybe a better way to think of it is it's a mathematical proof but for an audience other than human beings.
i wonder if a game is much different from a proof. i mean, we can implement games as programs, so there is a connection at that level.
My suspicion is that Chomsky's "language instinct" is actually a derangement of our ability to reason about probabilities that leads us to make the same mistakes consistently, thus understand each other.
> Now let's take another activity that humans should be good at: game-playing.
It should be just the opposite. There was no evolutionary pressure on humans to make them play Chess well. The games are interesting to us partly because they're challenging and make us think differently.
If you want a fair comparison, you need to look at how successful we've been at making machines to do things animals were facing evolutionary pressure to do successfully.
Imagine trying to build an ant. A machine with a tiny, power efficient brain, that coordinates with its fellows to gather resources and build enormous hives. Could we make computers do that? Maybe eventually, but we're nowhere close today. Just making a machine walk with grace is at the edge of what we can do.
Comments seem to have gone off on a tangent about multiplication.
What stands out for me is the irony, that ultimately the purpose of AI is not to be especially good at multiplication but rather to replicate the tenuous, fragile and indefinable properties of Human Intelligence that can only come about through some process more sophisticated than logical binary determinism.
It seems that any discussion about artificial intelligence eventually devolves into arguments essentially about whether human brains are magical or deterministic.
> Ask a human to raise 3425 to the power of 542 and watch them sit there for hours trying to work it out.
You can't compare the speed of a computer to the speed of a human. Only because a human is slow doesn't mean that the human is stupid, nor computer are smarter. The electric current in a integrated circuit acts near the speed of light. Light itself is also very fast but not very intelligent.
> [...] the world Chess champion has been a computer.
Again, this is a matter of speed. If you give a human being the same amount of time in relation what a computer had taken to calculate the same steps of a chess game it would be a more true comparsion of intelligence.
> Humans have almost no memory
That's because the human brain isn't trained and not used. We use almost 10% of our brain. So this comparsion also hinks.
> The face recognition software that Facebook uses can tell the faces of millions of people apart.
And again, this is only a matter of training and not a matter of intelligence.
No offense to OP or Mr. Togelius but this argument is terrible in almost every way imaginable and completely unconvincing at that. It makes its case entirely by subtly introducing straw men and double standards.
Leaving aside that its primary point is made by a redefinition of “intelligence” (itself being nebulous without providing two definitions), it completely ignores the fact that computers would be completely unable to do any of these things had smart people not told them how to. You may say the same of people as well, but people can learn things independent of knowledge. Even something as simple as space and time is understood a priori by people but most computers are oblivious to what these things actually mean.
The arguments about memory are terrible because one might as well say a library is smarter than a person if all that matters is the accuracy of recall and the amount of data stored. The computer itself does not know these things in the same way that we do, if you ask it to find them for you it will search them the same way you might but more quickly. Knowledge is contextual and intuitive, and computers currently are not great at context or intuition.
And that’s just knowledge! It’s so easy to demonstrate that human knowledge is more complex than computer “knowledge” that it’s barely even worth discussing.
I personally think one of the strongest indicators of intelligence in people is their intuition and the ease at which they adapt to new things, i.e. how many things come easily to them. Nothing comes easily to a computer. Everything must be specified clearly and carefully to the computer by a person who is better at intuition than the computer is. The computer has no way of knowing whether something is “right” or “wrong” in the sense that those words intersect with both morality and logic. They might understand that something is “incorrect,” but that does not carry the negative connotations that “wrong” does for a human being. Computers have not “computed how to play the game perfectly,” they were told how to do so through increasing levels of abstraction. Computers have not “computed how to play the game perfectly” just as pencils have not learned how to make marks on paper. That a computer does what you tell it to and by definition can’t do what you don’t tell it to is evidence enough of its utility as a tool and not a person.
This is, of course, in the current context of AI. I have no doubt that some day we will create software with the sort of intuition and ability to contextualize information that people have. But until then, there’s not sense deluding ourselves that we’re already there. If that were the case, we might as well have stopped doing computer science research with the Bombe 70 years ago.
Will a robot/AI crossing the street ever realize (without anyone explicitly telling it) that if it is stuck by a car then it will be incapable of moving?
Will an image recognizer ever "truly understand" what it means for something to be in a category?
If I set up a Go board such that the pieces on the board resemble a smiley face or some other pattern, will there ever be a version of AlphaGo that is able to recognize that? Will it be able to stop playing Go and start placing pieces on the board that fit into the pattern?
Will an AI be able to make an original joke that is not based off of any template?
To my knowledge we don't have an AI that can do any of the above, but any person would find those tasks easy.
> "Humans are quite stupid in many ways, compared to computers. Let's start with the most obvious: they can't count. Ask a human to raise 3425 to the power of 542 and watch them sit there for hours trying to work it out. Ridiculous"
I hope this guy is just trolling, but in case he is not, this is a tired argument that should be debunked once and for all.
The brain in the course of seemingly mundane activities (interpreting what we see, for instance) effectively performs a stupendous amount of complex calculations per second [1]
What people confuse is conscious calculations vs effective calculations.[2] The brain does not need to output intermediate results of basic operations because that is not it's computational objective.
I was actually a bit disappointed at the shallowness of the article; from the title I was expecting maybe a discussion of how complex even the very concept of intelligence is, and how speed of calculations does not necessarily equate to intelligence.
If all you have is a hammer, everything looks like a nail...
I'm not familiar with Togelius' other writings so I don't know how fast and loose the man is with his words, but in isolation, these "arguments" are like a compilation of youtube comments. Normally I ignore them, but some days I take the bait.
The question at issue is made out to be "who's more intelligent, computers or human beings?" when the real question is "what is intelligence in the first place?". To merely assume some definition because it suites the author's position is nothing short of question begging.
There also do exist powerful arguments against strong AI and computationalism. Searle's Chinese room argument is perhaps the best known, but by all appearances, often unappreciated or misunderstood. The essential point he makes is that computers are syntactic machines, i.e., machines that transform strings of symbols (which are intrinsically meaningless) according to syntactic rules. However, human minds contain semantics (concepts). Because computers are syntactic machines only, they necessarily do not and cannot possess semantics. They can simulate semantics when a human being formalizes semantics by producing syntactic rules for the simulation, but no amount of syntax ever results in semantics any more than skillfully adding clay to a sculpture can ever produce a human being. Remember, a computer is anything that implements anything equivalent to the Turing machine (a formalization of effective method).
Aristotle, on the other hand, makes a much deeper argument about the nature of the intellect that can reinforce a restricted form of Searle's argument, viz., his arguments can be used to explain why computers lack semantics by showing that matter per se cannot possess "concepts" as such and apart from particular instances. This argument is difficult to appreciate without an understanding of Aristotle's broader metaphysics. However, the outline of the argument is as follows:
1. Matter is particular/concrete (e.g., "this tree/that rose").
2. Concepts are abstract (e.g., "Tree as a class/Redness as such").
3. The intellect, the organ of abstracting concepts from particular instances, holds concepts.
4. Therefore, the intellect is not material. QED.
...adding own minor premise and conclusion...
5. Computers are purely material.
6. Therefore, computers cannot be intelligent.
Note that "intellect" is not a synonym for "mind". Aristotle distinguishes such things as imagination (phantasm) from the intellect, the former of which he argues is material. To better see how concepts are immaterial, consider the word "tree". You may imagine a tree, or even a number of trees, but the image is always particular, it is always an image of a particular tree whether real or not. However, none of these is the concept "tree" which is not particular (if it were particular, then there could only be one particular tree). You can repeat the same reflection with anything: every triangle you imagine will be isosceles, scalene or right-angle and of some particular color, and indeed something triangular and not a triangle as such.
The general problem here can be related to the problem of qualia (and intensionality) and thus the mind-body problem introduced by Descartes' metaphysics and haunting much of philosophical discourse since (even when the mind is dropped and the body endowed with the powers attributed to the mind). Note that Aristotle's immaterial intellect is NOT Descartes' mind.
Others who have argued against computationalist or materialist conceptions of the mind include Kripke and Popper, but there are many in-depth treatments of the subject that address many of the claims and objections raised by the computationalists. That being said, I find "AI" (arguably a misnomer) to be a very interesting field.
>4. Therefore, the intellect is not material. QED.
That is face-palmingly bad logic. Can we pretend that famous ancient sages weren't so blindingly, obviously dumb as to believe one piece of stuff can't do multiple things?
>The general problem here can be related to the problem of qualia (and intensionality) and thus the mind-body problem introduced by Descartes' metaphysics and haunting much of philosophical discourse since (even when the mind is dropped and the body endowed with the powers attributed to the mind).
I don't see how the acquisition of abstract knowledge has anything to do with qualia or the mind-body problem. Abstract knowledge is just hierarchical modeling.
97 comments
[ 3.1 ms ] story [ 171 ms ] threadAvoiding totally spoiling the book, it asks the question: "Is consciousness really a survival trait in the long term?"
This has issues too - if consciousness is an emergent property of a neural net with the right feedback mechanisms and training material then even if you have the right feedback mechanisms you could still create an artificial consciousness that's stupid.
You see this problems in humans - depending on a lifetime of 'test data' exposure (parents, peers, environment) and the underlying brain neural net 'hardware' you can get people that believe a lot of stupid things.
Maybe consciousness doesn't have to work that way and we're just dealing with a local maxima of evolution (or some reproduction/sex drive constraint), but we might end up stumbling on the ability to create a neural net with the ability to emerge a consciousness before we can craft the type of consciousness we'd want.
Actually understanding how the system works is harder.
Human culture persists outside of individuals, so the occasional genius can change culture and culture stays changed because collective memory is altered. This doesn't happen with individual animal intelligence.
So I don't think we'll ever see a true AGI. What will happen instead is that collective cultural memory will improve, and connections between human brains will become faster and deeper - just as they have done for millennia.
If there's ever a singularity it will be created by human brains operating together in a new and vastly accelerated shared state, not by silicon brains that have been programmed to appear sentient.
I think your point about intelligence not existing in an individual is too narrow, but ultimately that comes down to how you decide to define intelligence.
Not sure I agree that AGI won't happen - in fact I suspect the computing power we have now is good enough to do it, but we're just missing some insight with how.
Smart people in the past (and currently) have believed stupid things because of what they were exposed to and what they were optimizing on.
Inventing multiplication is something no computer (as we currently understand them) would ever be able to do.
For another analogy, moving ten tons of rock a thousand feet is something humans can do, and have done for millennia. It's very slow and difficult. A bulldozer can do the same thing in minutes. But a bulldozer would never, ever have a reason to move ten tons of rock.
http://web.onetel.net.uk/~hibou/ai-course/lenat.txt
https://en.wikipedia.org/wiki/Automated_Mathematician
Of course, you could argue that multiplication was implicitly there in the rules and facts it was programmed with.
Basically everything that isn't a science.
I think the OP's point was until humans invent true strong AI there is a qualitative difference between human and computer intelligence.
And even if/when a true "thinking machine" is invented, the intelligence which was able to create said machine would be nothing to sneeze at!
AM, the discovery learning program developed by Doug Lenat, discovered multiplication, beginning with a few basic facts about set theory, and using about 200 heuristics.
Would a program flag the discovery of multiplication as important because it understood that this discovery is interesting or would it flag its discovery as important because the programmer encoded it to have this understanding? That is can a program discover something new and alert to humans to this discovery and for it to be truly interesting and new?
I'm not sure how to convey the idea in my head to English in such a way to make it clear why I think BACON doesn't qualify as discovering new ideas. I'll try though.
From the write up, by Langely, of the program he wrote the program to detect regularity in data and if it found such regularity it would "leap out" to the program so that "it could take appropriate" action. As I understand the program it picks up on any sort of regularity. The program doesn't understand which forms of regularity are important and which are not. The programming system and data it looks at are geared to finding regularity that are meaningful. If the data inputs are vastly widened then it seems to me it would find regularity that wouldn't be considered meaningful. I can imagine the program fixating on a regularity pattern that is meaningless. I hope what I'm writing makes sense.
The work he did is interesting and I wonder what advances have been made since then.
Let me ask this question. Would a program that looks for regularity discover that in finite probability spaces probability zero events never occur? I think this is something a computer could discover. Could it end up concluding that probability zero events can occur if one looks at uncountable probability measure spaces? It seems to mean that programs will never be able to think or discover things at a meta level. Would a program be able to discover Godel's incompleteness theorem? Of course I don't know the answer. I just suspect that the answer is no. However, humans took thousands of years to discover these things. So perhaps an AI that ran for that long could as well.
Thank you for pointing me to the BACON program. I'd never heard of it before.
In any case, let me flip the question around at you: how did Goedel come up with his incompleteness theorem? Surely it didn't just magically appear in his head? Most likely he followed some standard procedures of mathematical practice: manipulate symbols according to learned rules, look for 'surprising' features in the result, build a sample demonstration of a surprising result, and use intuition regarding its outcome to drive further thinking, etc.
You can look at any intellectual behavior and find processes underlying it, unless you want to argue that we don't live in a materialist universe and people have souls. Unless you go down that route, there must be mechanisms to thought, and indeed applied psychology agrees. So if there are mechanism to thought, why can't those run on a computer instead of biology?
From EURISKO: A program that learns new heuristics and domain concepts: "Almost immediately, symmetry heuristics produced a very powerful yet simple device, one which simultaneously computes NAND and OR, using only two small metal regions, two n-doped regions, two p-doped regions, and one intrinsic channel region."
Yes. Abstract knowledge is learnable, and the brain is an inference machine.
PS: I have even seen 'rediscovery' used for the same situations when a math idea was lost etc.
calculation itself really only exists with purpose. it's a thing to do if you have a reason to do it. of course, these kinds of things too seem more "discovered" than "invented", but it's a very different kind of discovery than "the earth revolves around the sun". the earth revolves around the son, as far as we can figure these days, whether any agent/subjectivity/etc intends anything or not.
however, no full-blooded calculation will ever occur anywhere in the universe unless someone needs to figure something out. in other words, intention is implicit with calculation (and truth-functional representation, by the by). as such, you could make a case for an argument for "inventing" strategies in mathematics. and of course, a lot depends on your ontology of mathematics in the first place. if you're a wittgensteinian about numbers, and you think any of what i said above is at all compelling, you may be inclined to believe that mathematics itself is an invention. you may admit of some more nuanced way that entities like "number" still exist without mathematics as a tool, you may not, but you'd probably see all mathematics existing as active work, not pre-existing objective Truth.
of course, most people think of math as an always-there abstract object (roughly). maybe they're right. even so, there must be some force in the "i use it as a tool" and "it just revolves around the sun no matter what anyone--god or man or beast--wants" distinction.
I guess the straightforward counterargument here is that if multiplication only exists in humans' minds how can computers do it, but computers have absolutely zero autonomy -- a computer can't compute a product, or even "discover" multiplication without receiving explicit intructions from a human to do so -- in my mind a computer can't be deemed the one discovering multiplication if that discovery relied 100% on the design of human beings.
I'm pretty sure computers, as we currently understand them, are entirely capable of inventing multiplication, provided sufficient source code and input. Evolution is unique in being able to "create" a thing that can formalize multiplication? Human intelligence is not magic.
What's amazing isn't that Deep Blue could beat Kasparov, given the depth of analysis it could do. What's amazing is that Kasparov could beat Deep Blue, despite not being able to analyze thousands of possible moves a second in his head.
My probably oversimplified understanding is that the neural network training on a massive amount of games develops something like intuition. It can't explain why a certain move it makes is the best, but only that it is the most likely to win.
This seems similar to advanced human play.
Are Kalman filters machine intelligence? It sort of doesn't matter, really.
[1] https://en.wikipedia.org/wiki/AI_effect
This part seems wrong. I think most humans would make a decent go of most of those situations. Not as good as an expert by any means, but they'd be capable of doing something, unlike a computer program.
If you put an untrained system and an average person in the crashing plane situation, the average person is probably going to do better, but it's because your average person /has/ some training in what to do in this situation. They've been exposed to enough culture to know that the joystick is the main way of controlling the plane and that planes moving towards the ground at steep angles are likely to be bad, and probably know generally what action to take that might help the situation.
Fair comparisons would be say, an untrained program vs a baby, or a system trained with some general knowledge of vehicle control vs an average person. I'm not really sure that humans win out in either of those comparisons.
Huh? I think the whole point is that that's a very difficult task.
> Fair comparisons would be say, an untrained program vs a baby, or a system trained with some general knowledge of vehicle control vs an average person. I'm not really sure that humans win out in either of those comparisons.
The human got very little direct training. They mostly got exposed to the environment and picked it up. You can give the computer program the same number of years and the same environment as the human got, and it won't do much good.
Nevertheless, I'm quite optimistic, but it's far from the foregone conclusion that the author implies it it.
Intelligence is knowing what to do when you don't know what to do.
At least the random human will have a chance at doing something and if the situation isn't life or death like the landing plane or the Amazon jungle could with time actually learn to operate in the new environment even without interacting with other people who could teach them. That's what's missing from AI the flexibility to operate in a chaotic environment. Until relatively recently the slightest thing going wrong in even moving through an area or performing a simple task like moving a box would completely break.
> (It's hard to understand why anyone would want to be in a plane flown by a human, now that there are alternatives.)
Detecting and ignoring spurious inputs and extreme edge cases are one of the main reasons that I feel better if there's a person alert and at the controls of a plane. Extreme cases like Quantas 32 where a huge number of systems are absolutely destroyed would be a huge challenge for modern autopilots which are great but aren't tested or designed for emergencies. [1]
[1] http://lifehacker.com/the-power-of-mental-models-how-flight-...
Specialization of software is mostly a virtue.
And generalization of humans is mostly a virtue
Part of the argument given for the virtues of generalization is about availability. Like, if you find yourself the only available person in a plane over the Amazon, you can improvise a landing. But as the network connects more things in more places, this becomes less and less important.
To me it's not about the exact number of tasks that a particular bit of software can perform but that to call it's intelligence better than human it has to be able to increase it's own capability over time in novel tasks and environments. I'm also extremely resistant to equating simple calculating and data look up speed with intelligence as the author does with the multiplication and random SSN look up examples.
We can't really tell until we do it. And the hard part is the analysis of the business model, anyway.
But what is more likely: a human pilot detecting and resolving problems with the computer systems, or the human accidentally crashing the plane?
Of the many errors in this article this is one of the most flagrant. Computers have yet to invent any novel and challenging games. That would take ingenuity, something computers have failed to demonstrate. However, I suspect the author is trolling a bit.
Games exist that are variants of "click the arrow on the coloured pixel 100 times". And people pay to play them. Good luck finding one with a Deep Conv.
i wonder if a game is much different from a proof. i mean, we can implement games as programs, so there is a connection at that level.
It should be just the opposite. There was no evolutionary pressure on humans to make them play Chess well. The games are interesting to us partly because they're challenging and make us think differently.
If you want a fair comparison, you need to look at how successful we've been at making machines to do things animals were facing evolutionary pressure to do successfully.
Imagine trying to build an ant. A machine with a tiny, power efficient brain, that coordinates with its fellows to gather resources and build enormous hives. Could we make computers do that? Maybe eventually, but we're nowhere close today. Just making a machine walk with grace is at the edge of what we can do.
What stands out for me is the irony, that ultimately the purpose of AI is not to be especially good at multiplication but rather to replicate the tenuous, fragile and indefinable properties of Human Intelligence that can only come about through some process more sophisticated than logical binary determinism.
You can't compare the speed of a computer to the speed of a human. Only because a human is slow doesn't mean that the human is stupid, nor computer are smarter. The electric current in a integrated circuit acts near the speed of light. Light itself is also very fast but not very intelligent.
> [...] the world Chess champion has been a computer.
Again, this is a matter of speed. If you give a human being the same amount of time in relation what a computer had taken to calculate the same steps of a chess game it would be a more true comparsion of intelligence.
> Humans have almost no memory
That's because the human brain isn't trained and not used. We use almost 10% of our brain. So this comparsion also hinks.
> The face recognition software that Facebook uses can tell the faces of millions of people apart.
And again, this is only a matter of training and not a matter of intelligence.
Leaving aside that its primary point is made by a redefinition of “intelligence” (itself being nebulous without providing two definitions), it completely ignores the fact that computers would be completely unable to do any of these things had smart people not told them how to. You may say the same of people as well, but people can learn things independent of knowledge. Even something as simple as space and time is understood a priori by people but most computers are oblivious to what these things actually mean.
The arguments about memory are terrible because one might as well say a library is smarter than a person if all that matters is the accuracy of recall and the amount of data stored. The computer itself does not know these things in the same way that we do, if you ask it to find them for you it will search them the same way you might but more quickly. Knowledge is contextual and intuitive, and computers currently are not great at context or intuition.
And that’s just knowledge! It’s so easy to demonstrate that human knowledge is more complex than computer “knowledge” that it’s barely even worth discussing.
I personally think one of the strongest indicators of intelligence in people is their intuition and the ease at which they adapt to new things, i.e. how many things come easily to them. Nothing comes easily to a computer. Everything must be specified clearly and carefully to the computer by a person who is better at intuition than the computer is. The computer has no way of knowing whether something is “right” or “wrong” in the sense that those words intersect with both morality and logic. They might understand that something is “incorrect,” but that does not carry the negative connotations that “wrong” does for a human being. Computers have not “computed how to play the game perfectly,” they were told how to do so through increasing levels of abstraction. Computers have not “computed how to play the game perfectly” just as pencils have not learned how to make marks on paper. That a computer does what you tell it to and by definition can’t do what you don’t tell it to is evidence enough of its utility as a tool and not a person.
This is, of course, in the current context of AI. I have no doubt that some day we will create software with the sort of intuition and ability to contextualize information that people have. But until then, there’s not sense deluding ourselves that we’re already there. If that were the case, we might as well have stopped doing computer science research with the Bombe 70 years ago.
Will an image recognizer ever "truly understand" what it means for something to be in a category?
If I set up a Go board such that the pieces on the board resemble a smiley face or some other pattern, will there ever be a version of AlphaGo that is able to recognize that? Will it be able to stop playing Go and start placing pieces on the board that fit into the pattern?
Will an AI be able to make an original joke that is not based off of any template?
To my knowledge we don't have an AI that can do any of the above, but any person would find those tasks easy.
I hope this guy is just trolling, but in case he is not, this is a tired argument that should be debunked once and for all.
The brain in the course of seemingly mundane activities (interpreting what we see, for instance) effectively performs a stupendous amount of complex calculations per second [1]
What people confuse is conscious calculations vs effective calculations.[2] The brain does not need to output intermediate results of basic operations because that is not it's computational objective.
I was actually a bit disappointed at the shallowness of the article; from the title I was expecting maybe a discussion of how complex even the very concept of intelligence is, and how speed of calculations does not necessarily equate to intelligence.
[1] http://gizmodo.com/an-83-000-processor-supercomputer-only-ma...
[2] http://chrisfwestbury.blogspot.com/2014/06/on-processing-spe...
I'm not familiar with Togelius' other writings so I don't know how fast and loose the man is with his words, but in isolation, these "arguments" are like a compilation of youtube comments. Normally I ignore them, but some days I take the bait.
The question at issue is made out to be "who's more intelligent, computers or human beings?" when the real question is "what is intelligence in the first place?". To merely assume some definition because it suites the author's position is nothing short of question begging.
There also do exist powerful arguments against strong AI and computationalism. Searle's Chinese room argument is perhaps the best known, but by all appearances, often unappreciated or misunderstood. The essential point he makes is that computers are syntactic machines, i.e., machines that transform strings of symbols (which are intrinsically meaningless) according to syntactic rules. However, human minds contain semantics (concepts). Because computers are syntactic machines only, they necessarily do not and cannot possess semantics. They can simulate semantics when a human being formalizes semantics by producing syntactic rules for the simulation, but no amount of syntax ever results in semantics any more than skillfully adding clay to a sculpture can ever produce a human being. Remember, a computer is anything that implements anything equivalent to the Turing machine (a formalization of effective method).
Aristotle, on the other hand, makes a much deeper argument about the nature of the intellect that can reinforce a restricted form of Searle's argument, viz., his arguments can be used to explain why computers lack semantics by showing that matter per se cannot possess "concepts" as such and apart from particular instances. This argument is difficult to appreciate without an understanding of Aristotle's broader metaphysics. However, the outline of the argument is as follows:
1. Matter is particular/concrete (e.g., "this tree/that rose").
2. Concepts are abstract (e.g., "Tree as a class/Redness as such").
3. The intellect, the organ of abstracting concepts from particular instances, holds concepts.
4. Therefore, the intellect is not material. QED.
...adding own minor premise and conclusion...
5. Computers are purely material.
6. Therefore, computers cannot be intelligent.
Note that "intellect" is not a synonym for "mind". Aristotle distinguishes such things as imagination (phantasm) from the intellect, the former of which he argues is material. To better see how concepts are immaterial, consider the word "tree". You may imagine a tree, or even a number of trees, but the image is always particular, it is always an image of a particular tree whether real or not. However, none of these is the concept "tree" which is not particular (if it were particular, then there could only be one particular tree). You can repeat the same reflection with anything: every triangle you imagine will be isosceles, scalene or right-angle and of some particular color, and indeed something triangular and not a triangle as such.
The general problem here can be related to the problem of qualia (and intensionality) and thus the mind-body problem introduced by Descartes' metaphysics and haunting much of philosophical discourse since (even when the mind is dropped and the body endowed with the powers attributed to the mind). Note that Aristotle's immaterial intellect is NOT Descartes' mind.
Others who have argued against computationalist or materialist conceptions of the mind include Kripke and Popper, but there are many in-depth treatments of the subject that address many of the claims and objections raised by the computationalists. That being said, I find "AI" (arguably a misnomer) to be a very interesting field.
That is face-palmingly bad logic. Can we pretend that famous ancient sages weren't so blindingly, obviously dumb as to believe one piece of stuff can't do multiple things?
>The general problem here can be related to the problem of qualia (and intensionality) and thus the mind-body problem introduced by Descartes' metaphysics and haunting much of philosophical discourse since (even when the mind is dropped and the body endowed with the powers attributed to the mind).
I don't see how the acquisition of abstract knowledge has anything to do with qualia or the mind-body problem. Abstract knowledge is just hierarchical modeling.