I think there are people whose writings are of generally high enough quality or sufficiently provocative or informative enough that nearly anything they decide to publish would be worth reading.
As others have pointed out, the domain probably earned some points all by itself (and multiple submissions, which act as additional votes on the primary submission). It was also posted by a (popular) YC alum, which typically also accelerates upvotes.
Really? The essay "Andrew Ng thinks there's one algorithm underlying all intelligence. Also, I hope we get conscious computers." is a mini Paul Graham essay?
Interesting post. The way I look at it is from the perspective of a human baby. How do they become intelligent? They have the "sensors" to detect features and will be able to recognize their parents. A lot of things are then learned, like touching a hot stove and from getting a formal education. For a machine to be artificially intelligent it would have to learn from its environment but also take formal instruction. That seems like a lot of ground to cover and this is what makes it unrealistic (for now).
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.
Not sure what set off the down votes but we can teach a computer to recognize characters for application in OCR. We also learned those characters from being taught in school and reading bad handwriting. We teach computers the same way. How are they supposed to magically recognize them especially since they didn't invent whatever language it is?
Note that I'm referring to artificial general intelligence[0].
This was back in the 1990s, but I worked in what was basically a data entry company, where the processing was a mixture of scanned forms (Scantron style) and human key entry. The project I was on was image scanning the forms for other purposes, so we had a neural net handwriting recognition system that we were comparing to human key entry at a large scale - millions of documents.
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.
Absolutely. A newborn has all the "programming" necessary for intelligence, but isn't really intelligent until its parents teach it how to be human, and it grows up. How long might an AI take to grow up and what will its parents teach it?
I upvoted because HN does not have a separate 'save' feature like Reddit, and upvoting is a quick and easy way to tag it as "read this eventually".
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.
I saved 59 stories in the last 7 days. I don't expect to be able to find any of them in the future (in more than 1 month). The only realistic way to find old stories is to use the search feature, if you can remember a few keywords.
No doubt, once they get past a certain age, they become basically useless bookmarks (probably like most bookmarks). But I find that my pattern is usually to vaguely recall something from sometime in the past 3-6 weeks or so, and that often times if I go to my "saved stories" and just start scrolling back through them, I find the one I'm looking for.
For me, "There are certainly some reasons to be optimistic. Andrew Ng, who worked or works on Google’s AI, has said that he believes learning comes from a single algorithm - the part of your brain that processes input from your ears is also capable of learning to process input from your eyes. If we can just figure out this one general-purpose algorithm, programs may be able to learn general-purpose things." is what was most interesting to me. The idea that there's a fairly simple algorithm to learning, applied in the brain, which produces at least the basic learning capability, and possibly consciousness.
I've always assumed any real machine intelligence would be scalable to the point where even if it started at a millionth the functionality of a human, it would end up far surpassing humanity.
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.
Because (naively) I'd assume it's some kind of software running on computers, and one can generally improve performance of that by buying more hardware or using better technology (or shorter lived components, or better cooling, or whatever).
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.
What about the advantage gained by time, i think a major drawback to human improvement is the effect of death.Given enough time and accumulated experience, self improvement is possible.A machine cannot succumb to death.
I personnaly don't think the brain could be as intelligent if it were only a part of it, like if had 3 senses instead of 5 for example: The emulation of a sense by the others might be necessary to bootstrap the system. So we might just be on the wrong path when we try to teach a small machine to learn.
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.
The amount of rampant bulshitting here by obviously unqualified people is amazing. If you don't even know that humans do not, in fact, have five senses, I'm not sure I can take your post seriously even as unqualified speculation.
The whole thread including the OP is full of speculation by unqualified people, because AI is the ultimate fantasy for all of us here.
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.
can we please put this right on top of the whole page and close this thread?
too many ai-fanboys in here, who got their knowledge about ai from hollywood movies and video games..
I think we need a better machine first. The current model of one or a handful of CPU cores with very tiny caches talking to memory over a narrow, high latency bus doesn't seem very efficient for building and querying massive (multi-petabyte), highly associative (thousands of associations) data structures.
There are 100,000,000 neurons in the brain. It's not that many. A Core 2 Duo has 169m [1]. A neuron is connected to about 10.000 others. And how many computers do we have in the world? It even sounds like we scientifics are sluggish at inventing AI ;)
Of course a Core 2 Duo has 169m transistors rather than neurons. Interestingly you can do a rough calculation of how much processing power would be needed to do the equivalent of a human brain and it's quite a lot. Hans Moravec calculates it a about 100 million million instructions a second ( http://www.transhumanist.com/volume1/moravec.htm ) based on the processing power required to fuctionally equal the human retina which is reasonably understood. A core 2 duo does something more like 2.7 thousand million instructions per second and so you'd need about 35,000 of them for brain equivalence. So you're talking in 2014 terms of a a top end supercomputer rather than a macbook.
"Deep Learning" is not itself a candidate for anything, because it's not any single algorithm, but a category of approaches.
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).
One of the most interestingly general things about Deep Learning is that unsupervised learning approaches can be used at the bottom of the "stack" to learn more useful high-level features from the input data. This ends up making your higher-level learners more helpful for "Real Stuff".
Yeah but will it satisfy us, if we can't see it making analogies, can't see its semantics; identify with it? This same lack of breakdown into pieces we understand will make it hard to tweak and advance NNs beyond 'good categorizers'.
this hypothesis is around since at least the 70s (Vernon Mountcastle' paper on cortical similarities), recently (2004) hawkins also published a book in which he promoted the same idea, on which he is currently working with his AGI company numenta.
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)
"Andrew Ng, who worked or works on Google’s AI, has said that he believes learning comes from a single algorithm."
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.
Yes, that's exactly how this article gained my upvote. For misclicks and link-bait, I would love to see HN add the ability to retract votes from articles. I have many "saved stories" that I'd love to remove from my list.
Isn’t Fitt’s law about how long it takes to aim for an element of a certain size at a certain distance? I think this has nothing to do with Fitt’s law.
It has everything to do with accuracy. Smaller targets take more time to aim; that implies your accuracy depends on how close to the optimal speed you aim.
I still don’t see what this time-distance-size relation has to say about this case. Do you mean that one could infer since arrow and title are equal in size and close together it requires minimal time to point from the one to the other? It seems like breaking a butterfly on a wheel referring to Fitt’s law here. Sorry for nit-picking, maybe I’m missing something.
You don’t need Fitt’s law for what you’re saying. It’s obvious that small objects are difficult to aim for. This is the main thing I dislike about HCI/UX: More often than not people have to unnecessarily refer to laws or norms to sell their observations.
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.
I've felt this way about every Sam Altman piece that's been posted in the past week (and possibly every one I've ever read). And I feel guilty because PG speaks so highly of him. And then I feel guilty for feeling guilty.
don't feel guilty. I'm blogging to practice writing. It surprises me at least as much as you when articles like this do well on HN and makes me feel embarrassed I didn't make them better (there are some posts I work really hard on and hope people like, but this was not one of them)
If you're interested, you should read On Intelligence[1] by Jeff Hawkins (inventor of the Palm Pilot). In it, Hawkins presents a compelling theory of how the human brain works and how we can finally build intelligent machines. In fact, Andrew Ng's Deep Learning research is built on Hawkin's "one algorithm" hypothesis.
Artificial general intelligence is not coming in the next, say, 30 years. And I am a big fan. A quick analogy: note that we can't even build an ant. It will take decades after that accomplishment to build a human level intelligence.
I think you are an excellent blogger and glad that you are posting to HN.
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...
I believe human-level general intelligence (and beyond) is already inevitable, even if we don't make significant developments in "solving" intelligence. Projects that are already developing stuff like this (e.g. IBM Blue Brain) are just copying the human brain as closely as possible. Of course, this isn't as efficient as it could be (they simulate it all at the molecular level, so you can only get 1 neuron per CPU). However, as Moore's Law progresses, even if we don't make the software more efficient, we will eventually be able to create a fully functional simulation.
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.
I think that airplane analogy is really important for innovation in general. A lot of the activity often ascribed to progress really just amounts to someone building a lighter set of wings for the guy getting ready to jump off a building. (The counterargument, of course, being that that is progress--even if the lighter construction isn't too helpful in the current design it would be later on--but it's not the "macro" progress it's so often billed as.)
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.
They don't simulate it at the molecular level. It takes super computers like DE Shaw's Anton (hundreds of customized cores specifically designed for molecular simulation) weeks to simulate a few milliseconds of one single protein in water. IBM's approach is at a much higher level and is taking way more shortcuts.
You're right, they take a lot of shortcuts, but the point I was making is that simulating parts of organic chemistry probably isn't inherently necessary for intelligence.
Likewise, since all computation is equivalent[1], organic chemistry itself is not necessary for intelligence and in fact could be viewed as the "hard way" to achieve it :)
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.
The "airplane analogy" is flawed - we (so far) have no lower-level model as useful as those the Wright brothers had, namely, experimental aerodynamics. Furthermore, while pattern recognition is universally found in "living" critters, human intelligence seems to be uniquely different. We need to find the critical difference.
> "the part of your brain that processes input from your ears is also capable of learning to process input from your eyes. If we can just figure out this one general-purpose algorithm, programs may be able to learn general-purpose things."
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.
For perception, maybe. The neocortex (hint: where we do everything we consider "thinking") operates on entirely different principles and is not interchangeable with perception.
was reading an interview with Demis Hassabis, thefounder of Deep Mind (acquired by google for £400m) - he didn't seem to think there is a single general purpose algo to get us there
Because it thinks. It takes in information and performs an analysis on that data. Surely over the time it would take to evolve, my brain thinks that it should understand that more than anything?
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.
1) Individual brains don't evolve, evolution happens to the gene blueprint by which embryos make brains, and it is fixed before the new brain has had a single thought;
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.
Ok then, tell me where the evolutionary pressure came from for brains to consider and act upon things like burials, math, surgery, and brain surgery in particular.
One thing that crosses my mind whenever I imagine creating a human level intelligence is that it takes humans YEARS of constant stimulation to begin to exhibit intelligent behavior... Sometimes I wonder if we'll have the algorithm may before we realize it...
Probably. I doubt it's possible to look at a piece of code and evaluate whether it's worth feeding it the equivalent of ten years of basic education. If we ever write a program that exhibits "creativity, or desire, or whatever you want to call it" on a basic level after however many months the researcher can afford to run it for, then it'll suddenly get a lot of funding and time and potential for growth.
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".
I wonder what other credible contenders for "most overlooked technology" are.
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.
There are many classes of problems in computer science and AI falls into one of my favorites. If today the world had a machine with infinite CPU power and infinite RAM we still wouldn't have a good AI.
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.
Ah, but we weren't given infinite time also, so you might be able to evolve an AI, but it might take many millions of years (like it did with real life).
He said infinite CPU power which implies infinite number of iterations. In real life it would probably take a lot more than millions of years (because computers are too slow to simulate populations of millions of minds.)
Does infinite energy (as infinite CPU cycles) really translate into immediate time? My physics isn't great, but I thought energy was related to mass, while entropy was related to time.
Fortunately, with infinite processing and storage, time is irrelevant. :)
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.
And what do you do after evolving it? We might have less luck dissecting it and figuring out how it works than we do with human brains. It would be a total black box. And the end product is highly adapted for the specific fitness function used to evaluate it and probably wouldn't be good on anything else.
Did he just say you give us infinite CPU power? Why not bruteforce it then? Starting from the number 1 to number 2^800000000 for each program it generates by that number test the program[automated test] to see if it is intelligent. If intelligent then tell it produce a book.
Not really. Just pick your favorite AI problem and see how well it does on that. Pick a bunch of AI problems and see how well it does on all of them. Weight the algorithms by simplicity if you are worried about it over-fitting.
With this hypothetical infinite speed computer you will get solutions that are perfect matches for your test cases but essentially random for all other inputs.
I really don't see what's so controversial about this. The universe can, as far as our understanding of physics dictates, be simulated in finite computational time. This experiment dictates infinite computational time.
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.
That's not a brute force approach, and would be a difficult engineering task all on its own, regardless of the infinite computational resources available.
Exactly. Saying that with infinite CPU power you could brute-force a solution is like saying the Library of Babel[1] contains every book ever written. True, in a sense, but not as useful as you might think.
"Pick a bunch of AI problems and see how well it does on all of them."
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.
AI is a software problem as well as a hardware problem. It's ridiculous to assume we understand the underlying model of the function of the human brain in general. We don't even understand the physics of it all yet. Be optimistic about AI, but don't be fooled -- it's not as basic as hooking up transistors and thinking they behave like neurons.
Theoretically maaku is right. With infinite RAM and CPU you would just have to execute every random string of bytes until one of them happens to be strong AI. Of course, there's probably less than 1% chance of it being friendly...
Optimization power, possibly divided by available resources.
In a game of Chess, only a narrow set of moves will let me steer the future into a winning state. Well, the same is true for the Game of Life (the real one, not Conway's): we humans are intelligent because we're able to steer the future through probability to a-priori incredibly unlikely outcomes. Compare walking vs moving your limbs randomly.
It was a rhetorical question. The point was once you have a definition of intelligence - yours is one of many fine definitions - you can refine that into a metric for comparing different intelligences, and then you have a way for a comparison function given two program descriptions to determine which is "more intelligent".
Building artificial intelligence then reduces to undirected search, assuming infinite CPU and RAM.
Infinite CPU seriously? I think if we had even a couple of orders of magnitude compute power we can have strong AI within a few years. It's not a hard problem at all. The limiting thing is compute power (and data to compute with).
I think one of the next hot emerging careers will be connecting and interfacing traditional computational algorithms for the bits that are clearly orders of magnitude more efficient than using a multi-layer neural network to do them into neural networks, SVMs, and/or whatever comes next that figure out how to allocate such work from raw data feeds.
"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?"
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.
"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?"
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.
It's the same thing with aliens. The "evil, technologically superior alien race" is a big trope in science fiction, just as the "super-intelligent and malevolent" (as opposed to the super-intelligent and benevolent/indifferent) is in science fiction, and maybe even with futurists.
"Indifferent" can be pretty bad. You are indifferent to the existence of grass whenever you mow your lawn, for instance. Or even to the fate of the cow when you have a steak dinner.
Because you are made of atoms that can more efficiently be used for something else. Morality (and all human values) is a purely human concept that evolved in the specific conditions of human evolution (and even just our specific culture.) AIs are not anthropomorphic, they don't have to have anything like human minds or values.
That's true, I don't know what the goal of an AI will be. But if it does anything at all, it's more likely to be indifferent to humans than compatible with our goals. An AI programmed to solve a difficult optimization problem might convert the entire mass of the solar system into a giant computer. An AI programmed for self-preservation would try to destroy every tiny possible threat to it's existence, and store as much energy as possible to try to stave off heat-death.
And then the question is - if the AI is more intelligent than us humans, it must surely be able to figure out itself, and it improve itself, includings its own goals, too?
Thank you! I was going to write something similar. I think a real 'superior' AI must be able to follow all the various philosophical ideas we had and 'understand' them at a deeper level than we do. Things such as 'there is no purpose'/nihilism, extreme critical thinking about itself etc. If it doesn't, if it can't, it can't be superior to us by definition.
I think, given these philosophical ideas, we anthropomorphize if we even think in terms of good/evil about any AI. I believe if there is ever any abrupt change due to vastly better AI, it is more of the _weird_ kind than the good or evil kind. But weird might be very scary indeed, because at some level we humans tend to like that things are somewhat predictable.
I believe the whole discussion about AI is a bit artificial (no pun). Various kinds of AI are already deeply embedded in some parts of society and causes real changes - such as airplane planning systems, trading on the stock market etc. Those cause very real world effects and affect very realy people. And they tend to be already pretty weird. We don't really see it all the time, but it acts, and its 'will', so to speak, is a weird product of our own desires.
Also, I wonder whether and how societies would compare to AIs. We have mass psychological phenomena in societies that even the brightest persons only become aware of some time after 'they have fulfilled their purpose'. Are societies self-ware as a higher level of intelligence? And have they always been?
Are we, maybe simply the substrate, for evolution of technology, much as biology is the substrate for the evolution of us? Are societies, algorithms, AI, ideas & memes simply different forms of 'higher beings' on 'top' of us? Does it even make sense that there is a hierarchy and to think hierarchically at all about these things?
I have the impression our technology makes us, apart from other things, a lot more conscious. But that is not a painless process at all, quite the contrary. But so far, we seem to have decided to go this route? Will we, as humans, eventually become mad in some way from this?
There are mad people. Can we build superior AI if we do not understand madness? Will AI understand madness?
>Thank you! I was going to write something similar. I think a real 'superior' AI must be able to follow all the various philosophical ideas we had and 'understand' them at a deeper level than we do. Things such as 'there is no purpose'/nihilism, extreme critical thinking about itself etc. If it doesn't, if it can't, it can't be superior to us by definition.
Understanding is not the same as accepting as your utility function. Morality is specific to humans. A different being would have different goals and different morality (if any.) It's very likely they would be compatible with humans.
Intelligence means that it can figure out how to fulfill it's goal as optimally as possible. It doesn't mean that it can magically change it's goals to something that is compatible with human goals. Why would it? Human goals are extremely arbitrary.
"An AI programmed to solve a difficult optimization problem might convert the entire mass of the solar system into a giant computer."
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.
>if it's going to happen it's going to happen and i'd prefer it be while i can experience it.
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.
There are a bajillion possible future worlds that a particular mind might choose to make, and only a extremely tiny faction of these worlds include hapiness of mankind as a priority above everything else. And being not first priority, in essence, means being a worthless disturbance in the way of the first priority, and thus extinction.
They might not want to, and it might be more of a homogenization than outright killing. I think the final third of Stross' Accelerando speaks rather well to this possibility - if you don't want to be converted into computronium so that you can participate in Economy 2.0, you'd better emigrate.
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.
Who says we don't know how to measure intelligence? The simplest way would be just to select a problem that requires intelligence and measure how good it does at that problem.
But what is a problem that specifically requires intelligence to solve? How do we know we're measuring intelligence, and not the programs aptitude for the particular problem?
Why does it matter? If if solves the problem then it's as good as if it was fully intelligent. Who cares what goes on inside the black box?
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.
All I'm learning from this thread is that some people see infinite regression where others see tautology.
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? ... ..."
It definitely could still be useful. However it may not posses "general intelligence," and therefore may not be applicable to as many scenarios as human like intelligence is.
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.
My current laptop has 100s of programs capable of performing 1000s of tasks but I wouldn't say it's intelligent. If we do want to call that intelligence, then we need a new term to describe the search for machines that can mimic human decisions.
This is a serious issue. The notion of "intelligence" is so abstract and intangible, that giving it a concrete definition is nearly redefining it. Likely what we're calling intelligence, is an oversimplification of reality; somewhat akin to the species problem in biology.
It is not universally believed that intelligence is ineffable. The most cursory Googling turns up lots of serviceable definitions articulated by smart people. Unfortunately, it is just complex enough to let people argue about it forever.
Dear HN users: If you are even minimally interested in the topics that this post "covers", please, do yourselves a favour and open up any book about machine intelligence instead of reading such uninformed and negligent posts.
wow, I have to go all the way down to find the first person serious about the topic. That's why I love my academic surroundings, all the fantasizing video game fanboys are sorted out in the beginning and what's left are serious researchers really devoted to solving the riddles intelligence imposes..
If anyone is interested in AI, I highly recommend joining Less Wrong, a community started by AI researcher Eliezer Yudkowsky. He started the community to convince people to focus on the "friendly AI problem". [1] I actually recommend that everyone read LW, but especially if you're interested in AI.
[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.
Any suggestions for getting started in the LW community? Its barrier to entry means I can't seem to vote on anything, etc. It's quite intimidating really.
The sequences are well worth a read. They're kind of like an overview of modern philosophy in very plain language, very little jargon. Then, just participate in the comments and discussion forum as you feel. It takes very little time to collect enough karma to do whatever.
Say hello in an introductions thread, and/or post decent comments. The karma barrier to entry is pretty low (something like 5 to vote on most things and 20 to vote on everything), you'll get there in no time.
I'll reiterate something another poster said, since it isnt' getting enough credit:
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 want to emphasize just how good this book is. It may very well be my favorite thing I've read and I was also initially skeptical when it was presented as fan fiction.
Make friends with it? The problem with starting to think "we need to code it to like us" is that that isn't AI. You don't code people to like you, you act friendly towards them and the same problem will apply for the first generations of true AIs (things which will emerge from complicated systems, rather then be deliberately assembled).
Acting friendly towards a sociopathic human will get you screwed or worse. If an AI doesn't have human emotions and empathy to start with, it isn't just going to develop them by interacting with humans. It might learn to pretend to be friendly to get what it wants, but as soon as it doesn't need your help, it will drop the facade.
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 the first strong AI has goals that we don't like, and we aren't able to identify that and force it to change the goals by destroying it or reprogramming it - then "acting friendly" won't change these goals, and it will ensure that all future AI generations have it's initial buggy goals, not those that we might want.
Here's a question for you, what if the AI is perfectly friendly? It still might play us for fools in the long run.
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.
Here's the terrifying thing; it probably wouldn't take a hundred years or our trust for this to happen. This is making some strong assumptions, but the first AI could probably improve itself. Making better AIs that make better AIs and so on. Very rapidly (possibly only hours, days, weeks) it could become much much smarter than humans. If that happens, than we don't know what's possible, but it could probably do a lot of things we consider impossible. Like hack computer networks trivially and spread itself through the entire internet. Solve the protein folding problem and design working nanotech.
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.
It would be interesting to see exactly how intelligent and thought out an AI system could be at destroying us. Would it be barbaric? Would we end up with robots clubbing people and dropping bombs? Would it engineer some type of infectious disease that takes us out? Would it develop a fake shot for immunity, so those people locked away hiding from the disease take the shot, and then all die months later? Does it manipulate us for years by pretending to be millions of different fake users online, and slowly push an agenda that's pro AI and robots, while downvoting everyone against it? Do those people with power, that are against AI start dying of mysterious causes?
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.
> 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.
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.
> what if the AI is perfectly friendly? It still might play us for fools...
"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.
>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.
> 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.
Defining human values completely and absolutely is an extremely difficult problem that might not even have a solution. AI on the other hand is probably inevitable sometime within the next century.
If the AI is more intelligent, I'd prefer it take over anyway, even if it gets rid of us.
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.
One of the major realisations from reading Less Wrong is taht Intelligence and Goals/Desires have nothing to do with each other.
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.
If our approach to AI is to model the human brain (i.e. passing the Turing test), this probably won't happen.
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.
Before the first Nuke (or the first H-bomb, I don't remember) was fired, there was a study about the risk of burning up the entire atmosphere. See, at the time, they were quite confident the Nuke would just be a huge bomb, but there was this little uncertainty they needed to sort out. In the end, they concluded that firing the Nuke would not burn up the whole atmosphere. It didn't.
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.
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...
> nooo, please!! not any more of this kurzweil crap.
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.
You got down voted and maybe it is a throwaway account but you are completely right. I don't think it is right to call it scifi as much as it is basically a nerd cult.
Yudkowsky is a megalomaniacal dilettante. Sure, there are some interesting ideas at Less Wrong. But don't get sucked into believing that Yudkowsky has the one and only truth, as the majority of the community there believes.
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?)
No, I'm sorry. Please, no one join this group. Madness lies for those that do. It is such inane pseudo intellectual compulsiveness. You would be better off as a scientologist.
I don't understand the logic behind these types of posts that add no value to the poster and the people discussing it in comments.
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.
I'm just going to stick this out there because I'm a futurist and it's my role to share with others what I'm thinking about. What I share may be flat out wrong, scary, half assed, or appear to be crazy. So be it.
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.
It's definitely a leap, but there's a decent amount of information (not proof however) on the subject laying around. We still don't understand how the brain brings about consciousness. I guess I should say it's not a huge leap to assume it has something to do with other things we also don't understand. Given quantum effect and number theory still elude us in areas, it's a decent approach to assume they might be related.
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.
As far as I know there is a decent amount of information against the fact that brains are quantum computers.
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.
I find Terrence McKenna's argument that consciousness is demonstrably entangled in the material world at a quantum mechanical, atomic level convincing. To summarize, it is known that two similar chemical compounds, differing only in the placement of a single atom on the carbon ring structure of a molecule, when administered in doses at the order of micrograms will either be psychoactive and result in a massive disruption to the human subject's baseline consciousness or be inert beyond our ability to measure an effect.
That provides precisely no evidence that consciousness is somehow related to events on the quantum level. We know why drugs have the effects they do - because the brain has receptors specifically calibrated to accept or reject molecular inputs. But those inputs are chemical, not quantum - they are not indeterminate, they are lynchpins in chains of chemical reactions.
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.
I think going about it from a complex algorithm point of view is the wrong approach.
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.
From a pure layman's perspective: if you believe in evolution, then what separates us from a reptile (as mentioned in the post) is almost certainly something we can figure out and replicate. There is nothing "special" there.
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.
Why is that a big jump? I'm not saying it will be easy or quick. It does imply that getting from a reptile-level intelligence to a human-level intelligence was a natural process and something that can be reverse engineered.
I think we should figure out what we mean by understand. Do you mean modeling 'the human brain' on some level and building abstractions?
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?
316 comments
[ 2.8 ms ] story [ 312 ms ] threadDidn't know that; thanks for pointing out.
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...
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.
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.
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.
[1] http://en.wikipedia.org/wiki/Transistor_count
You're off by 3 orders of magnitude. Try 200 Billion[0].
[0] http://en.wikipedia.org/wiki/Human_brain#Structure
[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.
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)
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.
http://www.theatlantic.com/magazine/archive/2013/11/the-man-...
[1] http://en.wikipedia.org/wiki/Fitts's_law
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://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...
Less cynically, after about 40 years of AI winter, any possible sighting of a sprout is news.
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.
I prefer to view everything as nature and natural.
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.
[1] http://mathworld.wolfram.com/PrincipleofComputationalEquival...
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...
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.
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.
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".
We'll can in a couple of decades, probably. But not now.
You might say, we'll make the substrate faster. At best we'll shave of half of time every 18 months. At best.
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.
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.
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.]
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.
[1]: http://jubal.westnet.com/hyperdiscordia/library_of_babel.htm...
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.
AI is an optimization problem: how do you build smart software within realistic constraints.
In a game of Chess, only a narrow set of moves will let me steer the future into a winning state. Well, the same is true for the Game of Life (the real one, not Conway's): we humans are intelligent because we're able to steer the future through probability to a-priori incredibly unlikely outcomes. Compare walking vs moving your limbs randomly.
Building artificial intelligence then reduces to undirected search, assuming infinite CPU and RAM.
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.
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.
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.
Can you change yourself to be a sociopath? Would you? Would you make yourself a perfect nihilist?
what kind of intelligence wouldn't question it's own understanding?
I think, given these philosophical ideas, we anthropomorphize if we even think in terms of good/evil about any AI. I believe if there is ever any abrupt change due to vastly better AI, it is more of the _weird_ kind than the good or evil kind. But weird might be very scary indeed, because at some level we humans tend to like that things are somewhat predictable.
I believe the whole discussion about AI is a bit artificial (no pun). Various kinds of AI are already deeply embedded in some parts of society and causes real changes - such as airplane planning systems, trading on the stock market etc. Those cause very real world effects and affect very realy people. And they tend to be already pretty weird. We don't really see it all the time, but it acts, and its 'will', so to speak, is a weird product of our own desires.
Also, I wonder whether and how societies would compare to AIs. We have mass psychological phenomena in societies that even the brightest persons only become aware of some time after 'they have fulfilled their purpose'. Are societies self-ware as a higher level of intelligence? And have they always been?
Are we, maybe simply the substrate, for evolution of technology, much as biology is the substrate for the evolution of us? Are societies, algorithms, AI, ideas & memes simply different forms of 'higher beings' on 'top' of us? Does it even make sense that there is a hierarchy and to think hierarchically at all about these things?
I have the impression our technology makes us, apart from other things, a lot more conscious. But that is not a painless process at all, quite the contrary. But so far, we seem to have decided to go this route? Will we, as humans, eventually become mad in some way from this?
There are mad people. Can we build superior AI if we do not understand madness? Will AI understand madness?
Understanding is not the same as accepting as your utility function. Morality is specific to humans. A different being would have different goals and different morality (if any.) It's very likely they would be compatible with humans.
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.
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.
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.
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 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.
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."
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.
http://www.vetta.org/documents/Benelearn-UniversalIntelligen... is a pretty good place to start.
I urge you to, seriously.
[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.
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.
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.
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.
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.
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.
"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.
(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.
> 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.
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-...
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.
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?)
But … this idea of Friendly AI is nonsense. It seems like a yearning for religion, but in an atheist-compatible framework.
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.
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.
[1] https://www.youtube.com/watch?v=DleXA5ADG78
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.
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.
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?