AI
blog.samaltman.com
blog.samaltman.com
It's not that complicated.
The only non-obvious insights Fitt’s law bring are that objects twice as big and twice as far away take the same time to aim for, and that as objects become smaller or distance increases the time only grows logarithmically. Everything else is just squeezed into its definition to make it sound well-founded.
[1] http://en.wikipedia.org/wiki/Deep_learning#Convolutional_neu...
[2] http://deeplearning.net/reading-list/
[3] http://en.wikipedia.org/wiki/Deep_learning#Results
[4] http://www.wired.com/wiredscience/2012/06/google-x-neural-ne...
Deep Learning generally refers to machine learning algorithms that deal with stacking multiple layers of simpler functions to enable more complicated functions, and optimizing all the parameters to best fit your training set and generalize to new samples (the hard part). Though it usually refers to neural networks, I dont think there's any reason it doesn't also apply to other layered approaches as long as there's a relatively unified learning algorithm applied across the whole system.
There are clearly many different deep learning algorithms, even if you just count the permutations of tricks you can choose from to improve layered NN generalization. Though to be fair I think very good progress is being made towards developing "better" algorithms in the sense that new ones (e.g. RBM pretraining + dropout) usual perform better than than older algorithms, no matter what data you use it on (now network architecture is another matter entirely).
But I do agree with your point.
As algorithms can be combined, the existence of any set of algorithms satisfying this goal would automatically imply the existence of a single algorithm incorporating all of them.
If a sufficiently-detailed physical simulation of a human's brain satisfied this goal, then that would be one such algorithm.
We don't see animals self-improving to become humans (in the "consciously deciding what revisions to make to their mental architecture at each step" sense), so why would you expect electronic animal-level minds to be able to do the same to become electronic human-level minds, or beyond?
Personally, I doubt even electronic human-equivalent minds would be capable of self-improvement. After all, we are not smart enough to build AIs better than us (yet), so why should electronic minds only as smart as we are be capable of that, either?
A possible answer to that, I suppose, would be that the electronic mind would have far more input/training data fed to it per second than biological minds receive, and far more time to "work on" that sense data to derive patterns between each decision-step.
If humans created an AI to solve some kind of open-ended problem, where "more AI" made the solution better, there would be every incentive to spend the money on more or better hardware for it. It's not clear that a shark would gain much by being 100x smarter, particularly if the metabolic cost were high; for a lot of human problems, spending 100x more on hardware/power/etc. for a 10% better solution would be quite desirable.
And maybe, once you have a fully functional brain, that it constantly derives out of bounds. Like sudden overheating, which is litterally the step #1 of a depression in a human. So we might not be able to scale an AI beyond the size of one brain. Apart from clusters obviously, but then you need to sustain civilizations of brains, and civs do collapse every dozen generations.
We might not be materially able to find enough energy to power all of those trials and errors.
So common sense is a perfectly satisfying qualification to post in this particular thread. Concerning the 5 senses, I refer to common wisdom because it speaks to everyone. Those who know better are probably smart enough to translate "5 senses" into an accurate scientific wording.
You're off by 3 orders of magnitude. Try 200 Billion[0].
to be honest, it sounds really great and it could even be the case that there is a very general underlying principle to cortical information processing and pattern recognition. But one should be careful not to mix solid scientific hypotheses with mainstream media hysteria and people who try to grab attention with their simplifications, claiming today that entropy maximization is the underlying principle and changing to sparse coding tomorrow. We are not that far and what we need is solid research instead of over-the-head assumptions and claims "to have solved the riddle" (in that respect, it might not be that far off alchemy :P)
[1]: http://www.amazon.com/On-Intelligence-Jeff-Hawkins/dp/080507...
That said, I hope that you will think more critically and clearly before publishing vague, fuzzy, uninformed, and unlogical thoughts (not illogical, but unlogical) like the following:
>The biggest question for me is not about artificial intelligence, but instead about artificial consciousness, or creativity, or desire, or whatever you want to call it. I am quite confident that we’ll be able to make computer programs that perform specific complex tasks very well. But how do we make a computer program that decides what it wants to do? How do we make a computer decide to care on its own about learning to drive a car? Or write a novel?
Consciousness, creativity, and desire are all quite distinct things. It is very important for people who are attempting to approach the coming reality of artificial intelligence to be able to distinguish between different things like that.
There have been computer programs that decide what they want to do for decades. Perhaps you were thinking of a specific human-like type of decision process, but if so, you must say so and reason that way. Otherwise you are just conveying some fuzzy thoughts. And the problem is that you are doing so in the context of real scientific undertakings with results directly applicable to your thoughts.
A computer deciding what to care about or learn or what behavior to engage in "on its own" is related to the previous topic you mention, and in and of itself, does not require artificial general intelligence.
How do we make a computer program write a novel? I think that is a good question and an effective answer to it I believe _might_ be in the category of 'real' artificial general intelligence. However, I think that it will probably soon be possible to create 'narrow' AIs that can generate novels without being generally intelligent. http://www.nytimes.com/2011/09/11/business/computer-generate...
So I reflexively upvote anything that looks even vaguely interesting.
In this case, I did read the article, and would probably have upvoted it anyway. Why? Because it stands to serve as the seed of an interesting discussion.
Personally, I don't give a fuck if the article itself "adds anything" or not. Who cares about that? It's irrelevant. If the topic itself and/or the content of TFA are interesting enough that it gets a bunch of interesting HN readers talking and commenting and linking and sharing stuff, then it's a worthwhile article in my book. Not everything has to be an earth-shattering scientific breakthrough, that's published in a peer-reviewed journal, blah, blah, blah.
http://www.theatlantic.com/magazine/archive/2013/11/the-man-...
Less cynically, after about 40 years of AI winter, any possible sighting of a sprout is news.
On the other hand, if you want to go the biological route, there's some awesome work to be done. If I were to study consciousness, here's the question I would ask: how do we separate our selves from our surroundings? Patients with brain-machine interfaces (like moving a mouse cursor) start by thinking about moving their arms around. Then they apparently report that they gradually just feel that the interface is another body part. So if it's set up to change the TV channel, they just imagine that they have a channel-changing organ.
So maybe you want to build a system that can identify what is a part of itself versus what is not, and it's not just a fixed list. So what does that data structure look like? How is it defined, queried, and updated? Defined by what you can 'influence?' So gradated based on my influence? These aren't just broad philosophical questions, they're more specific and actionable.
That's just one possible angle, but it's different than, say, machine learning paradigms where you want to build a machine that can do pattern classification (which the brain undoubtedly does). There are probably other routes as well.
I think building an artificial consciousness is going too far. Artificial intelligence is simpler; it's just fake intelligence. Seems easy enough right? If it looks like a duck and quacks like a duck then it's intelligent. We don't need to make it "conscious" necessarily, again whatever that means, in order for it be intelligent.
I feel like we can build artificially intelligent software pretty "easily" relative to making it "conscious".
One of the really, really bad consequences of the Cold War was the scientific divide between East and West. By that I mean serious lack of scientific data exchange between the blocks. The consequences are still felt and this area (the problem of consciousness) is the one that suffered. The problem of "consciousness" was basically solved, at least at a conceptual level, by Soviet psychology and neuropsychology. Here I refer, of course, to the work of Vygotsky and Luria. What is consciousness? Almost nothing at all by itself. Consciousness as found in humans is a consequence of our cognitive development and the advanced symbolic capabilities of humans. The subjective perception we have of the thing we call consciousness is "simply" (it's not really simple when you get into details) a product of humans acquiring language skills (I'm simplifying).
This is not to say the subject is trivial, it takes volumes to describe what is happening, but the thing we informally call "consciousness" is really nothing at all in and of itself, and the perception we have of it is just a result of the very complicated process of cognitive development. Thin air, like Lisp's cons.
If you want to read on it I can recommend Vygotsky's Language and Thought (actually, it's his only book) and Luria's Language and Consciousness (I'm not sure it was ever translated into English, it's a collection of his lecture notes from a university course he did on the subject) or possibly The Cognitive Development: Its Cultural and Social Foundations.
Why this line of thinking is mostly ignored in the West I have no idea. Why do we still cling to metaphysical (even religious I would say) phantasies about "consciousness" is an interesting topic itself. Is it because it's romantic to think there's something special, transcendent, about our minds? Are we really that sentimental? I have some hypotheses, but it's a different topic.
It's an interesting narrative though and I'll check out those books.
I'm not sure I understand what you mean by this. Of course it can, that's the whole purpose of psychology (and, more fashionably, neuroscience of course). To me that sounds like saying science can't answer these questions. Do note that when I say "psychology" I mean strictly the scientific areas of whatever comes in the bag labelled "psychology". Due to historical accidents the term acquired a lot of BS pseudo-scientific baggage, and it's really a shame those things can detract from a wealth of valuable hard results honest scientific psychology uncovered.
The answer that developmental cognitive psychology, at least the theory I'm referring to, gives is that "the mind" comes from the only place it can come from: neural processes and the way they hook into environmental interactions of the organism (social and physical). The key to understanding what gives rise to "consciousness" is in understanding the role of language acquisition in broader cognitive development. The point where a child utters it's first words is neither the beginning nor the end of this extremely nuanced process. In my first post I took it for granted it's understood that this is not just armchair speculation, it's based on empirical data. As any good scientific theory it's far from complete, maybe in some details is inaccurate but it's certainly infinitely better that an endless philosophical debate (with strong religious, or in the best case idealist, undertones) on what the mind is and where it comes from.
What empirical data can say anything about, for example, the philosophical zombie problem? http://en.wikipedia.org/wiki/Philosophical_zombie
As science leads you to believe that your consciousness is nothing at all (according to your description), bible based christianity tells you that your conscious is part of your soul, which is the part of you that lives forever. It is independent of your body, which will eventually be replaced with a perfect body.
For the longest time it was not even known that the mind inhabited the brain. We've only known about the biology of the brain beginning with cell theory and onwards. Yet we have an entire vocabulary relating to mental states receding so far into the distant past that we don't know how far back language predates using language to speak about mental states. When you really think about it neuro-biology is very recent and still nascent, we've been thinking about thinking for a long long time.
David Chalmers excerpt- Conversations on Consciousness by Susan Blackmore.
If everything is conscious then some parts of it are just more dynamic (intelligent?) than others. Physical reality least, plants more [1], animals even more and humans most.
Defined like that human consciousness just becomes that part of all consciousness which we recognize as similar to our own.
In that view AI is just making a small part of reality, a computer, more dynamically conscious and, very importantly, more similar to our own so as to be more useful.
This reflexion may be extended by asking oneself how perfect prediction might be linked to pleasure / pain signals (this is purely rethorical as the answer is obvious and known). What is not to be enjoyed in perfect prediction ? This could be a reason to self-enjoyment (or self-sense might be related to enjoyment), and even enjoyment of other people whose actions we can predict.
We are embarking on the path to understanding what information really is, and this will no doubt deeply shake our world. Deep and fascinating subject anyway.
Somehow I fear the only satisfying implementation of AGI has to end in a skynet scenario. We as human will only accept intelligence as general when it is at least as intelligent as humans. But that would mean that we have to accept that we can't control it for sure. In fact only a machine that is able to rebel can be considered to be general intelligent. So I am not surprised that consciousness is not defined that way, because pursuing to build a machine capable of consciousness would mean to build a possible enemy.
[1] In a nutshell, the friendly AI problem is: assume we create an AI. It may rapidly become more intelligent than us, if we program it right. As soon as it becomes significanlty more intelligent, we will no longer be the most intelligent beings around, so the AI's goals will matter more than ours.
Therefore, we should really give it good goals that are compatible with what we want to happen. And since no one right now knows how to define "what humans want" good enough for writing it in code, then we'd better figure THAT out before building AI.
Why don't we just give that task to the AI? It'll be smarter than us...
Maybe the problem is that people are too easy to understand: We want "Brave New World", but we don't want to know about it, or that we want it.
> To be a safe fulfiller of a wish, a genie must share the same values that led you to make the wish. Otherwise the genie may not choose a path through time which leads to the destination you had in mind, or it may fail to exclude horrible side effects that would lead you to not even consider a plan in the first place. Wishes are leaky generalizations, derived from the huge but finite structure that is your entire morality; only by including this entire structure can you plug all the leaks.
Humans can mostly differentiate between good and bad (ethics), but we don't know how we arrive at those conclusions (metaethics) because humans are terrible at introspection. Also, there's a ton of gray areas (e.g. the trolley problem). So rather than define all possible edge cases, it's probably less difficult to understand human decision-making from first principles and model our FAI accordingly.
For example, let's say we develop a super friendly AI, running on your computer. The AI realizes the human race is actually awful. We're greedy, we're killing tons of animals, chopping down rainforests, destroying the ocean and planet, starting wars with one another, and committing unspeakable acts of evil at times. The AI, being more intelligent than us, might decide the world is better off without the human race, and that we're actually a problem that needs to be removed.
Now, what does the AI do in your computer? Well, it's intelligent and knows the human race. It's not a hurry. It calculates the best way to destroy our species. It acts friendly, and talks about how humans and robots should live together, and if we make robots with a similar intelligence, they could drive our cars, shine our shoes, cook us dinner, look after the elderly, open your pickle jar, etc. So, we listen to the AI, it's smart, and friendly, and we build all these robots. It's right, the new robots are doing great and helping us out. Then the robots start building more and more robots. They start building robots with firepower, so they can, you know, shoot down threatening asteroids, or stop one of those dangerous human types that goes on a killing spree in our society. Fast forward a couple of hundred years, and there are robots everywhere. They finally decide it's time to continue their plan, they're in a position of power at this point, and they can instantly disable our security systems, phone lines, satellites, internet etc, and start wiping us out.
We're gone. They constructed the most efficient way to clean us from the planet. They were planning it for hundreds of years, starting in your computer. The AI then goes on to explore the universe, and we're just a blip in the past.
It kind of feels like we're a bug going towards the light, and that the unfortunate conclusion is almost inevitable.
"Friendly" is a term of art among AI people, at least at Less Wrong, and their meaning of friendly excludes this whole scenario. A friendly AI is one which helps humanity and has no horrifying side effects. The vagueness of that definition is the problem Yudkowsky and his acolytes are trying to solve.
Within a week it could pay/blackmail/manipulate some humans somewhere into developing some crude self-replicating robots or nanotech. Then almost immediately afterwards it consumes the entire Earth in a swarm of rapidly self-replicating nanobots.
...or something. How should I know what a mind literally millions of times more intelligent than me would do. It's like predicting the exact next move a chessmaster will make. I don't know, but I'm confident they'd beat me quickly.
The goal, of course, is to make an AI which won't do this. If the AI decides to terminate the human race, we've already failed making it friendly and it obviously doesn't share our goals and values. But what are our goals and values? I don't know if anyone can answer that. I'm not sure if there is a satisfactory answer.
Even if it decides to be our friend and to help our species, someone will of course fork that AI and give it a negative personality and goals. Then you have the evil AI trying to hack the friendly AI that exists in our homes, and it's a battle of the robots.
Of course, whether or not AI is even possible, no one knows. If it is, I think we'll achieve it, and we'll open up a remarkable can of worms.
No: the first AI won't let them. See, we're talking a rapidly improving super-intelligence. Whatever is contrary to its goals, it will squash like a bug. A mad scientist forking the code of the AI with a different goal structure is definitely contrary to the goals of that first super-intelligence, and will be shut down before it grows into a sizeable competitor.
The result of intelligence explosion is a Singleton: the AI will be a perfectly efficient dictator. It may even shield us from the laws of physics until we graduate to adulthood.
We also probably won't get multiple generations to work these problems out. The first true AI could rapidly increase it's own intelligence and power and then pretty much do whatever it wants. We have to get it right the first time.
(If we're going all LW on this thread: http://lesswrong.com/lw/xt/interpersonal_entanglement/)
>Sure, it sounds silly. But if your grand vision of the future isn't at least as much fun as a volcano lair with catpersons of the appropriate gender, you should just go with that instead. This rules out a surprising number of proposals.
Read HP:MoR.
It's a fanfic of Harry Potter, written by the same Eliezer Yudkowsky who wrote much of Less Wrong. It was specifically written to convey the feeling of "what it means to be a rationalist".
For those who aren't into Harry Potter or into Fanfiction (like me), I can tell you this: Suprisingly, it is one of the best stories I've ever read. And I'm talking just as a story, nevermind the other value you can get form it, which is a good introduction to the "rationalist" community.
I'd argue that the BEST way to understand what is going on at LessWrong is to read HP:MoR, as it was intended to be such an intro and succeeds masterfully, while being amazingly fun.
I think I might restart reading it.
When you have a child, do you want the child to stay at home and do what you say forever? Or do you want the child to succeed as much as possible? I want the latter.
An AI might be much more intelligent, but it has a goal system that basically says: "Make as many paperclips as you can". Everything it does will be with the singular purpose of making more paperclips. Not music, math, sport, culture or anything else that we think of as good. Not "help save intelligent creatures and animals from death". Only one goal - making more paperclips.
And if it decides that the optimal way to make paperclips just happens to involve death and destruction to humanity, that won't matter.
So yes, the AI might be more intelligent, but I still wouldn't want to trade humanity for an intelligence which doesn't do anything I value.
Friendly AI is a silly research project at this stage of AI research. It's like trying to figure out how to make horseless carriages safe before you have an internal combustion engine, or even know what one is.
The lesson is this: we had only one try. If nukes did cause the atmosphere to burn up in a giant blaze, we would all be dead by now. If you do something, anything, you better make sure it won't kill us all.
Horseless carriages? Sure, these might kill a few people, here and then[1], but we're pretty sure they won't kill us all in one blow.
Intelligence on the other hand is way more dangerous. Human intelligence designed Nukes in the first place remember? AI can do way worse. Even if we model it after the human brain, if it's smart enough to do the same as we did, then it will be able to model another such AI, only slightly better, and so on until it takes over the world. "Taking over the world" may sound enormous, but it really isn't. Imagine for a minute a small group of cavemen vs an army of chimps. Well, if you give the cavemen a chance to prepare, the chimps are toast: the cavemen have spears, fire, better communication… Now imagine an AI imagine the AI is smarter than us by the same margin we're smarter than chimps. Same thing: if it's not safe, we're toast.
[1]: http://www.statisticbrain.com/car-crash-fatality-statistics-...
But … this idea of Friendly AI is nonsense. It seems like a yearning for religion, but in an atheist-compatible framework.
Does MIRI have a single real AI researcher, yet?
Also, as I say down the page, Friendly AI is a silly research project at this stage of AI knowledge. It's like trying to figure out how to make horseless carriages safe before you have an internal combustion engine, or even know what one is. It's an interesting thought experiment, but one that is probably unsolvable before we know a little bit more about what an AI will look like (an emulated human brain? Something else?)
nooo, please!! not any more of this kurzweil crap. guys wake up! This world is not some asimov sci-fi story.
friendly ai
I always get a headache when I read that term online. ppl seem to go crazy about the machines taking over earth idea but this whole debate is so utterly useless! If all the effort fapping to conscious AI and friendly AI would be put into concrete ai research (agi as well as applied ai) we would get to a reasonable point so much sooner...
Ray Kurzweil has little to do with this. When he talks about "singularity", he thinks about "accelerating change", with exponential growth everywhere, even beyond human intelligence scales.
Those who work on Friendly AI have another scenario: "intelligence explosion". Typically a self-improving AI that would grow itself into a super intelligence. You should read this introduction on the subject: http://intelligenceexplosion.com/
> This world is not some asimov sci-fi story.
Indeed. In Asimov's stories, the robots mostly behave, though in every story there is some problem with the way the laws of robotics apply.
In the real world, the laws of robotics don't work, and we would at best be put "safe" into cushioned rooms so we can't hurt ourselves. As for emotional hurt, drugs and lobotomy are extremely efficient remedies.
But if you look at the history of technology, things we create aren't usually exactly based on models seen in nature. Airplanes aren't exactly like birds. I believe we will find a more "man made" model for general intelligence (maybe not even a neuronal model) that works much more efficiently with the hardware we have available.
Going back to the airplane analogy, we already have the people who strap wooden wings to their arms and jump off buildings (like Blue Brain), but we are looking for the first Wright brothers design.
http://lesswrong.com/lw/vx/failure_by_analogy/
A very relevant article, both to the main topic and bird/airplane example.
Yes, biology has solved some problems and can suggest some solutions, but we can't go cargo-cult on it and expect things to work just because they look similar.
Equally, our materials science wasn't good enough for flapping wings at the time.
I expect to see flapping wing designs appearing in micro-flyers within a decade. At the insect scale they're really useful:
http://www.epsrc.ac.uk/newsevents/casestudies/2011/Pages/Tin...
[1] http://mathworld.wolfram.com/PrincipleofComputationalEquival...
Anyway, it's just interesting to be reminded that context is important. Being able to distinguish between insanity and genius seems like it would be a super-power.
IBM Watson is the obvious example, and the many successful Deep Learning-powered image recognition projects are another.
I prefer to view everything as nature and natural.
There are elements of past AI models in the HTM model, however to reduce HTM's or any deep learning algorithm to a mere combination of past AI concepts overlooks the power of the right model when it is achieved. It would be like saying that Facebook is just a news feed. Sure, that's what gets most of the eyeballs, but there's a lot more there which would drastically reduce its value if not present.
What I think is most interesting is that we may find that Humans learn pretty inefficiently from the perspective of the amount of input data required over time. This may seem silly at first, but when you consider how many neurons cover the surface area of our ears and eyes and then consider the fact that it takes anywhere from 12 to 14 months for a child to speak its first word, you might start to agree with this line of thought. Also, when I consider the fact that this processing all happens in parallel even further pushes me in this direction.
Whatever the case may be, HTMs are definitely a cool area of research. For those who are interested, you should definitely check out more of Jeff Hawkins work at Numenta. They've been able to demonstrate some pretty novel things. He wrote a book back in 2006 that blew my mind. Went into deeper explanation about how HTMs could model everything from deep learning, to consciousness, creativity, and bunch of other things.
Are we so sure that "we" really are in charge of what we want to do? I believe a lot of our desires and ambitions are hardcoded into our brains and we just project them onto present goals, like getting a promotion or learning how to play the piano, etc. ultimately all these desires cater to the same few desires we always had and were born with: developing a sense of social belonging and intimacy.
If it didn't have character flaws, it wouldn't be operating at a "human" level. But if it does have these character flaws, how useful would it really be compared to a real human? Is the quest for Strong AI just a Frankensteinian desire to create artificial life?
I'm curious if there are any good papers looking into stuff like this.
Why does quantum computing even exist if all forms of computing are equivalent?
For all we know, we already have the Algorithm and all we need to do is run it for years on the best computer available "just to see what happens".
You might say, we'll make the substrate faster. At best we'll shave of half of time every 18 months. At best.
We'll can in a couple of decades, probably. But not now.
I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.
We cannot decide what we want to do--we can only decide how best to fulfill our wants; a person learns to drive a car because we want the freedom, social approval, and other stuff that comes with that.
Natural selection gave us our base-level desires, that all other desires spring from; and it was able to do that because it's an optimization process. A functional AI's desires will come from some sort of optimization process; the only question is what that process will be optimizing.
Is it a signaling mechanism to attract people working in this area? I'm sure you have already turned them off by showing your naivete. So no value to you.
To people trying to discuss this by racking their brains and looking for new ideas, Any one with a quint decent thought/idea will never share it here to enlighten us laymen. So no value to us too.
So, what's the point HN, and why all the upvotes?
I guess SV types get a little hot for AI based utopian/dystopian succession when they're the only ones who could possibly gain from such an outcome.
You can have a machine read every book in existence but how long will it take for it to understand in the same way a human reading a lot of books would need to understand it somehow.
Note that I'm referring to artificial general intelligence[0].
0. http://en.wikipedia.org/wiki/Artificial_general_intelligence
What we found was that human key entry significantly outperformed the neural nets, even when the data was carefully handwritten in constrained boxes. Humans were so far ahead of the heavily trained neural nets that the software was basically unusable at that point.
Of course, that was nearly 20 years ago, and things have probably moved on quite a bit. But you can still see the basic problem in Captcha-style validation on web pages. Computers just can't be trained to recognize distorted text that humans can read pretty easily.
I think they have done that.
In essence: unless we extract that "single algorithm" that Andrew Ng believes in from an existing medium, we're unlikely to rediscover it independently.
(read the papers i link to at the bottom of the article for a more rigorous explanation)
I mean you need two neural nets. One that adds chaos to another neural net and a second one that returns the result. You could probably optimize by somehow making those parts parallel.
We have no proof that artificial general intelligence can exist. We have numerous examples of specific intelligence: playing chess, driving cars, various forms of categorization. But we don't have a single example of an application that can handle a task that it hasn't been specifically trained and tested for. It's not to say that it isn't possible, but there is no more evidence for AGI than there is for Bigfoot, leprechauns or space aliens. The idea of artificial consciousness currently requires a leap of faith.
The most important trend today is collecting massive amounts of data and using them to make accurate predictions. Instead of lusting after one all-singing all-dancing intelligent program we should focus on tackling one form of decision making at a time. Drive cars, land planes, predict the weather and calculate the best way to get from point A to point B. One day we'll wake up in a world where artificial intelligence is all around us and the idea of a one size fits all solution will seem silly and quaint.
[1] In fairness, this was said about every amazing thing in modern life. You never know.
Flying machines vs AGI makes a much better analogy. Natural flying machines (birds) existed in nature, before artificial ones (planes) were invented. Artificial manned flight was speculated as a possibility for centuries with no macroscopic working examples, and heavily criticized as a transportational panacea/fantasy that was seen as clearly impossible to people at the time.
In fact, the analogy extends startlingly far. People doubted the possibility of manned flying machines for the longest time, with arguments strikingly similar to yours, e.g. 'We have no proof that such artificial flying contraptions can exist!'. Explaining away natural flight as a supernatural magic only accessible to birds is also eerily similar to the constantly retreating dualistic arguments against a mechanical brain.
Even your criticisms of current AI has analogues[2]. Long ago, even in ancient times, there is evidence of small "toy" birds that might have flown much like paper planes. Many similar toy examples existed around the 19th/20th century, too, yet many people still vigorously doubted the possibility of a flying machine that could carry a human.
Let's compare a paraphrase of common arguments against heavier-than-air human flight, with a paraphrase of your argument:
Of course we have natural examples of flight, as we see birds all around us. But there is no more evidence for human flight than there is for Bigfoot, leprechauns or space aliens. Sure, we have little toy examples of flying machines, but they're very limited -- the idea of full heavier-than-air human flight requires a leap of faith.
Vs.
[Of course we have natural examples of AI, I, the author presumably am one.] But there is no more evidence for AGI than there is for Bigfoot, leprechauns or space aliens. Sure, we have little toy examples of AI, but they're very limited -- the idea of full AGI requires a leap of faith.
[1] (Obviously I'm not referring to fusion or anything like that, because people of that era wouldn't have recognized that as transmutation.)
[2] To be fair, I think it's safe to say old-school non-probabilistic AI is dead. But just because one path ends doesn't mean there aren't a thousand others constantly exploring new ideas. Indeed, recent advances in deep learning are incredibly promising.
Whether the flight analogy carries on to intelligence, in my mind, depends on how many of our subsystems are 1) indispensable for intelligence, and 2) reasonably computationally reducible.
From neurotransmitters to ganglia cells to hormones and bacteria in the gut, we have found a lot of subsystems that contribute to our abilities to make diverse, everyday decisions that the ideal AI we are discussing would have to make. The cortex actually seems like one of the most orderly and therefore reducible parts of the apparatus. The hormonal system that regulates emotion based decision making may be far more difficult to abstract and less efficient to model. And there are many many other systems. Could it be that without details of those subsystems, our AI behaves in less than optimal ways the same way a human would? How much can we get away with reducing biology to simpler rules while maintaining general intelligence?
It is possible I suppose that all those biological dependencies are merely hampering an ideal algorithm for generalized intelligence that we are only crude approximations of, a powerful and simple algorithm we can finally free of biological constraints, -- but it's too late to get into the probability of that hypothesis! In any case it's not clear to me how that kind of nonhuman intelligence would serve us.
It's probably not going to be easy, tough.
...just to bite the bate a bit:
> We have no proof that artificial general intelligence can exist.
Well, we had no proof that fire can be produced ultil the first paleohumans learned to make fire. We had no proof that mechanized flight was possible until we built the first flying machines. We had no proof that space flight is possible before we made it possible etc.
These kinds of arguments only make some sort of sense when you reason from first principle instead of reasoning by analogy: for example if you found an argument that, starting from the current known basic laws of physics, would after some logical steps (no, analogies with "what we now know exists" don't count as logical steps) arrive at the conclusion that AGI is not possible. Elon Musk has a cute and short explanation of "reasoning from first principle" vs. "reasoning by analogy": http://www.youtube.com/watch?v=L-s_3b5fRd8#t=1356 .
My point is that the way you and other people like you reason "by analogy" is just plain wrong. It's a useful way of thinking, that can help you make lots of very good business decision and profit from them. But when you apply it to technology or science it becomes obvious just how wrong it is, as you just said it: "In fairness, this was said about every amazing thing in modern life.".
If you are a teacher/mentor/speaker etc., please don't expose your students or other younger minds with potential to innovate to this way of thinking! This mindset is the most effective way of killing innovation. It doesn't matter that much the topic of whether AGI is possible or not. It's this way of thinking that is extremely toxic to innovation, even it may seem quite harmless and very useful.
If you've read some history, it's clear that this is exactly what the Wright brothers did when they pioneered powered flight. Three things makes Kitty Hawk a good place to test airplanes: You'll survive if you crash in the sand dunes, there is enough steady wind and there are no newspaper reporters around to make fun of your failures.
Get up and go look in the bathroom mirror, there's your example. If you're saying that our intelligence/consciousness derives from some secret sauce, then you're essentially making John Searle's Chinese Room argument and it's on you to show what the difference is between sapience and the appearance of sapience.
Your argument could have been made for every piece of technology before it existed. Here's what scientists, engineers, and mathematicians do: they either keep trying, or they find a reason why it's impossible.
Humans obviously can perform intellectual tasks that a human being can perform, thus humans are machines that implement general intelligence.
Evidence that something exists is evidence that it is created.
Does it strike you odd that humans have real trouble doing this? We are not good with 'big' data. Maybe from a large population things can be inferred, but why are we programming computers to do things we can't? Did the goalposts move? (yes)
I do think AGI lacks the incompleteness that's seemingly required of everything real. With no constraint (humans have many: low wattage, 8hr GC cycles, fairly static network structure..), it's too perfect a goal.
"Playing Atari Through Deep Reinforcement Learning" was published just this last year.
>The idea of artificial consciousness currently requires a leap of faith.
Detach the notion of AGI from "artificial consciousness" or "artificial people". Contrary to the normal Humans Are Special shtick we all recite, intelligence is only one design feature of us homo sapiens sapiens out of very many.
"Intelligence" in software terms is just machine learning in active decision environments. Or in other words, Machine Learning + Decision Theory = Artificial Intelligence.
This is not to say we should be cheerleading for the fabled "Strong AI" within the near term. Quite the opposite: I'm trying to express just how far the distance is between software that can learn and perform some general task without being specifically purpose-built, and a conscious, sapient Asimovian robot deserving of personhood rights, and of course Skynet.
For an even further elaboration of just how far off we are from the latter two forms of "AI": we currently have literally no way of specifying tasks or goals to general AI agents other than reinforcement learning. We are stuck training our software like we train our dogs: give it a cookie when it does the right thing, bring out the rolled-up newspaper when it goes wrong.
So yeah. AI is almost definitely possible, and a formal field of research regarding it does exist, but we are indeed decades away from anything really and truly useful for large-scale applications such as killing all humans.
I know how to play chess, but I don't seem to be playing it as well as Gary Kasparov.
But that doesn't mean anything. Creating something artificial that has whatever it is that we have is the goal.
I think "physical tamper evidence/tamper response" is one, along with hardware security functionality (crazy secure virtualization extensions, etc.) -- essentially competing with Intel not just on power but also on security features. Although Intel is leading in this area with TXT and now SGX.
Didn't know that; thanks for pointing out.
(samaltman.com)
Thing is that drive is uncertain/subjective.
Learning can't be an objective by itself, intelligence is just a tool, from that perspective you can say there are already AI, like targeted ads, it learns about you and acts accordingly.
So to have a "generalist" AI, like human's intelligence is general, you'd have to have an objective like staying alive and build up from that.
As I wrote a few days ago here:
https://news.ycombinator.com/item?id=7217967
"The way intelligence works, in my opinion, is this:
1) experiences are stored in the brain. Experiences contain inputs from the 5 senses as well as the sense of danger/satisfaction at that point.
2) at each given moment, the brain takes the current input and matches it against the stored experiences. If there is a match (up to a threshold), then the sense of danger/satisfaction is recalled. Thus the entity is able to 'predict', up to a specific point, if the outcome of the current situation is bad or good for it, and react accordingly.
The key thing to the above is that the whole process is fused together: the steps for adding new experiences, matching new experiences and recalling reactions is fused together in big pile of neurons."
We've advanced a lot in the last 100 years. We're starting to see a bigger picture forming with the advent of compute and networking capabilities. Combining simple elements of these basics give rise to surprising and interesting behaviors. See "Twitch Plays Pokemon": http://news.cnet.com/8301-1023_3-57619058-93/twitch-plays-po... as an example of suprising behavior.
The more we look in detail at the universe around us, the more puzzling it gets. Prime numbers spirals are unexplained. The two slit experiments results indicate the observer plays a part in collapsing a particle's probability wave. The effects of dark matter could be a result of parallel universes. You couldn't make up weirder shit if you tried.
It's not a huge leap of logic to assume some parts of our brain operate at a quantum level. Given that first statement comes to a truthful fruition, I don't think it would be entirely unreasonable to assume AI will do so as well. Given computers already use some quantum properties, it's also reasonable to expect advancement in AI lies in this direction.
When they announced Google was getting a D-Wave computer, I got really interested. Granted, they know beans about how it works (and whether or not it actually works at all) but it's still crazy interesting to consider.
As I said, I could be wrong or crazy. Or both.
To the best of our knowledge it is still a pretty large leap of logic.
I've debugged problems in my code that, at first glance, appear to be unrelated to each other. Given something is slightly off in one area isn't a proof something of in another area is related, but it's a good place to start looking.
On the other hand there is a talk by Hinton on youtube [1] that sheds interesting light on some variants of deep learning and the way the brain uses (classical) noise.
You omit that if you give a much larger dose of the same thing you'd kill a person, and a smaller dose might fall below any threshold of activity at all.
And for that matter, a microgram is a huge quantity of matter.
This is the Holy Grail of CS. I believe we're closer than most people would expect and I think it's going to be a race to the finish line.
We should, instead, be concentrating our efforts on two things - sensing, and reacting. The predictability of the reaction doesn't matter; all that matters is that the machine reacts. Everything else will need to depend on evolutionary processes, which requires a third criterion - changing reaction based on prior data.
If the previous reaction did not lead to a negative result ("negative" meaning detrimental to one or more arbitrary values), then the reaction can continue to the same stimulus. If the previous reaction, however, elicited a strong positive result, then the reaction should be encouraged. Similarly, if it triggered a strong negative response, it should be avoided.
To a degree, you could do this without any kind of "operating system," just by using sensory data as inputs in a complex circuit.
At least, that's how I would approach it. I know nothing about A.I. research.
We just don't have the knowledge to utilize such resources to write an AI that could, for exampe, play League of Legends or Starcraft at a level beyond professional gamers. And it certainly couldn't write a best selling novel. It could solve an arbitrarily large traveling salesman problem but it couldn't do those other things. I think that's kind of awesome.
I'm not saying it can't be done. Assuming we humans don't kill ourselves I think someday it will. But it's a long, long ways off.
AI is an optimization problem: how do you build smart software within realistic constraints.
Someone tries it, with mixed results.
Flip the switch and you'll have real intelligence (as opposed to our lazy Approximate Intelligence) just in time for the immediate heat death of the universe.
[Everything can be seen as a state space search.]
[1]: http://jubal.westnet.com/hyperdiscordia/library_of_babel.htm...
Evolution is a purely physical process. Make up a series of tests more or equally complex to those evolution present, and you'll end up with intelligence - unless there happens to be something very, very special about human intelligence as opposed to other forms of intelligence. Humans are currently a local maximum in the space of intelligences which have been explored by evolution.
The problem is, you have infinite potential algorithms and a finite number of tests. This means you'll necessarily get algorithms that pass all your tests but fail at least one other test of intelligence. Because of this, your tests won't actually let you discover which of the generated algorithms is intelligent.
Or, you could have an infinite number of tests, but if you have infinite tests for intelligence, you effectively already have an algorithm for intelligence (for any problem, just look up the answer in your list of tests), so, again, brute-forcing a solution isn't helpful.
i realize there are a lot of mechanics the program would need to be aware of, but they're finite.
writing a bot that can learn to play Starcraft on it's own, on the other hand, is the harder problem.
I urge you to, seriously.
So if you believe computers today already have the "intelligence" of a reptile, or a toddler (i.e., ability to play pong), or something along those lines, it's only a matter of time before a computer has the intelligence of a full-blown adult human (and soon thereafter much more).
Our level of intelligence/awareness seems magical only because we haven't fully understood it yet. That will change.
Those abstractions are exactly that - abstractions. They are not the thing itself. Do you think we can understand everything through abstractions, even the process of understanding itself?
I'd like to point interested readers to the AGI Conference series[1], the Open Cognition Project[2], and a mostly outdated (2009) but still useful list of everything AGI[3].
Perhaps it'd be better to ask: why would we want a computer to do these things? I certainly do not want to live in a world where computers have their own motivations and desires and the ability to act on the same.
Actually, I can put that more strongly: none of us will live very long in a world where computers have their own motivations, desires, and the ability to act on the same.
ha :) i love this. my initial reaction was again- this is a huge assumption. that is, you're assuming self-preservation would be a goal, before sustaining human life. but then i realized, i guess this is your point! regardless of what goals we intend to program for, their solutions are unknown to us and could be catastrophic by our definitions.
that all said; i welcome robot catastrophe too. if it's going to happen it's going to happen and i'd prefer it be while i can experience it.
I'd rather it turn out friendly than unfriendly, and I'd rather have as much time as possible to either figure it out or live out my life.
There'd be nothing to experience anyways. The AI might kill us all overnight with super-lethal virus, nukes, or swarms of nanobots.
It'd certainly be fascinating. But I think I would rather that humanity does whatever it can do ensure that any artificial intelligence we create won't cause an outcome that is bad for humanity.
Can you change yourself to be a sociopath? Would you? Would you make yourself a perfect nihilist?
I assume that an AI will be more intelligent than us if we build it right. Then assuming that a randomly designed intelligence has the same goals as us is a huge assumption. Most humans don't even have the same goals, and we're 99% similar to each other.
In other words - the assumption that a random intelligence shares our goals is a much bigger assumption than that a random intelligence will be just like us.
if intelligence is solved by reverse engineering the brain at a molecular level surely consciousness and creativity are?
"And maybe we don't want to build machines that are concious in this sense."
if the physical composition of the brain defines intelligence and conscience, i'm not sure you'll be able to pick and choose. i am all for artificial conscious though. yolo.
http://en.m.wikipedia.org/wiki/Philosophical_zombie
If P-Zombies are an impossibility, then so too true AI that dismisses the need for consciousness, or begs the question by assuming that consciousness will emerge from intelligence.
It may be that intelligence, true human intelligence, emerges from consciousness.
Not just the human brain, a great number of animals share the very same characteristics as the human brain we find ourselves closer and closer to them every single day as new researches on animal neurology and animal behavior get published. It would be very wrong to suggest that intelligence is only a human thing.
http://www.vetta.org/documents/Benelearn-UniversalIntelligen... is a pretty good place to start.
If you worried about it overfitting to a specific problem, give it lots of problems and weight the solutions by complexity. So you heavily favor simple algorithms that can learn to solve a large class of problems, over ones that are more adapted for those specific problems.
That's the most interesting and informative issue presented by this scenario.
The problem with the sort of test you propose is that just because a human uses intelligence to solve a problem, it does not follow that the task requires intelligence. For example, playing chess.
From an engineering point of view, your black box may be as good as the real thing, though you couldn't really trust it beyond the areas of its demonstrated competence. Knowing how it works, however, would be the most significant achievement.
Furthermore, it's going to be hard to build one of these black boxes without a reasonably good idea of how it is going to work.
There is also the risk that your weighting scheme will rule out the only algorithms that have a chance of succeeding, because I bet they are pretty complex.
edit: what I mean:
Person 1: "What if we're just measuring the color of the sky? What if we're just measuring blueness? What if we're just measuring a wavelength of light? ... ..."
Person 2: "The sky looks blue."
Why would you create a god, when you could instead become a god?
No PG essay is so lacking in content or original ideas. Conscious computers? Who hasn't thought of conscious computers? http://en.wikipedia.org/wiki/History_of_artificial_intellige...
Actually there were (and still are) a lot of the biggest names in the field there at the time (including Crick, Terrance Sejnowski (a professor of mine who co-invented the Boltzmann machine* amongst other things), VS Ramachandran). These are all people who were coming at the problem from the biological &| pure science/math side of things (vs people like Andrew Ng, who Sam mentioned, who have a more CS/engineering based approach).
No doubt they consistently came up with spectacular theories and very interesting models of how a specific regions of the brain may function. How for the most part they worked was this:
1. Come up with a biologically or cognitively plausible mathematical model (many of which were fantastically cool). 2. Implement and run this model on massively parallel architectures (at the time not quite the level of technological sophistication you see nowadays, so things may have changed a lot) 3. To train, use feedback from EEG (this is what I "worked" on, but they also worked with other electrical signals, MRI and chemical measurements at the level of individual neurons).
The biggest progress was made at the smallest level (understanding how individual neurons and small networks work). This was primarily because measurement at this level actually provided useful information. The signal/noise ratio of EEG scalp recordings (which to this day gives me nightmares) was (and is) so terrible that I left the field as a quite disgruntled phd student. Maybe I just didn't have the intellectual capacity, but I never felt like I was working on anything that made sense. This was true for many of my fellow graduate students .. after a couple years, we felt we were doing pseudoscience.
Rant completed, I think the CS/engineering approach is more promising: don't worry about the biology or some grand theory of the mind and just try to do something useful. Since computers get more powerful consistently, we'll incrementally be able to do more and more useful things. If consciousness emerges at all, it may or may not appear like human consciousness. We may not even be able to tell if/when this happens, but at least we would be solving real problems in the meanwhile.
It understands how every other organ works in explicit detail at the the molecular level.
The only thing my brain can imagine, is that my consciousness is disconnected somewhat from my brain. It's as though my consciousness is inside a machine that it barely understands the workings thereof. Like a dog riding in a car.
2) Your brain definitely doesn't understand how every other organ works in explicit detail; controlling organs only requires being part of a good enough feedback loop to have the organ mostly function; your brain understands your heart "at the molecular level" about as much fruitflies do.