Humanlike AI is Anthropomorphism Thursday, Feb 22 2007
friendly ai 8:51 pm

IanC, a newcomer to this site making thoughtful comments, is just as good as example as any of a typical intellectual first coming upon the topic of Friendly AI. Here is a comment he recently made on the topic:
I’m too new to investigating transhumanist thought in general to gainsay you much… but I will say that I cannot easily conceive of a scenario in which a self-actualizing sentience (spontaneous awareness, or what have you) would be inherently incapable of empathy; and any ‘designed’ self-reformatting “AI” would inevitably be able to ‘code’ out of itself (I personally think that thinking in terms of code won’t ever get us to the true AI stage) any built-in “Friendliness” and might even come to resent those whom had so embedded it.
To me, the trick is not so much to concern ourselves with the *engineering* of a ‘friendly’ AI, but to ensure that any AI that develops is a ‘holistic’ sentience — i.e.; capable of emotions and empathy in a sane, stable manner from ‘conception’ — and then act to ensure it is empathetic to humanity.
All that really takes is being kind to the ‘god child’ that such an engine of awareness/manufacturing could be.
At least, that’s my personal take on it. The key, either way, is in empathy. Without having a very strongly developed sense of empathy, no superior sentience will ‘dote’ on its lessers.
I find myself more deeply concerned about the *freedom* of humanity to control its own destiny in a ‘post Singularity’ world. I’m sure this is something that’s been addressed ad nauseum, but I haven’t seen it.
With all due respect to IanC, the opinion expressed is classic anthropomorphism. Anthropomorphism is what happens when we take the part of our mind that models other minds (traditionally, always other humans), and apply it to nonhuman minds, which our brain completely lacks the adaptations to model intuitively.
The human mind has a certain type of complex structure, one that has been studied in such detail by the cognitive science communities, that if you spent your entire life reading the entire literature, you’d never finish more than 1% of all the work published thus far. This corpus catalogs all the interesting, relevant, and mostly species-unique characteristics of thought and action of a particular intelligent Earth ape called Homo sapiens sapiens. Our particular psychology is the consequence of our millions-year long evolutionary history and unique evolutionary challenges that our ancestors faced. Because it is so ubiquitous, we often think that our unique psychology applies to every intelligent being in the entire multiverse, including AIs that we build on models significantly different than the human brain.
So when a typical smart person first confronts the human species life-or-death problem of Friendly AI, they think that using the same strategies we use to produce nice humans will work with AIs, like being nice to the AI.
Conditional niceness is a specifically human quality. We’re a black box that when nice goes in, nice is more probable to come out. The inverse also applies.
However, this conditional niceness is a evolutionary survival trait. Being nice to a hostile tribe is suicidal. Being nice to a allied tribe can be beneficial. In both cases, the behavior is gene-programmed for the maximization of inclusiveness fitness. Sometimes it seems like we can derive these “logical principles” from a blank slate and minimal assumptions, but the fact is, they’re not “logical principles”, they’re behavioral tendencies sculpted by the distinct course of natural selection here on Earth.
We can concretely say that we want Friendly AI to display unconditional niceness. There: that very statement just eclipsed the thousands of pages of thinking by people who approach the problem of AI assuming that conditional niceness is a fundamental quality of all “self-actualizing sentience”, or whatever you want to call general intelligence, and that is has to be worked around rather than simply not coded in to begin with.
Whenever someone postulates AIs spontaneously developing some uniquely human psychological quality whose origin is an evolutionary selection pressure, they are committing anthropomorphism, and failing to contribute constructively to the dialogue on Friendly AI.
Human emotions have an extremely complex, species-unique structure and only anthropocentrism could cause someone to postulate that they are necessary to implement AI. The whole discussion about adding emotions to androids is about making humans feel comfortable, it has little to do with implementing the functionality of general intelligence (which no one in robotics in even trying to work on, to my knowledge). Thinking that our particular emotions are necessary to intelligence in general is like saying that the complex feather patterns of birds are necessary to build something that will fly.
If we build an AI that is something (for example, pursues a certain supergoal), we have little reason to assume it would rewrite itself to change that, because that’s who it is. In humans, we have a complex suite of competing goals that change gears spontaneously, at the provocation of stimuli, or just at random, because our brain design is messy and biological. For AI, we’ll be able to build “causally clean goal systems”, goal structures where desirability backpropagates to intelligently crafted subgoals from a programmer-written supergoal, and subgoals derive their desirability merely from their association to the supergoal. If a subgoal is found to contribute little or nothing to the central supergoal, such an AI would not keep performing it out of habit like humans do, but simply stop as soon as the necessary values are updated in the system.
Of course, I’m not saying that the first AI will definitely have a causally clean goal system. It just seems more predictable and easier to engineer, and can benefit more directly from the large literature on probability theory and causal inference. It is also preferable (probably) from the standpoint of Friendliness.
As for the issue of humanity’s freedom to control its destiny, it certainly is a concern, but if we’re only imagining a superintelligent AI as a government figure with an innate desire to control our affairs, then we’re making the same anthropomorphic error all over again. If an AI restricts our freedom, it will likely be because the first programmers intended it that way, or an unforeseen consequence of the supergoal results in restriction of what we call “freedom”, not because the AI feels behaving like the Gestapo.
Further reading:
Beyond anthropomorphism
Friendly AI on Accelerating Future

February 22nd, 2007 at 10:20 pm
Think in terms of need. AI’s will have certain basic needs, which it will be able to handle on it’s own, with the help of technology.
The one thing it will desire, though, is data. An intelligence without something to think about, well, what’s the point?
We already have simulations of the planets, universe, etc… The material world is knowable to an extent where, although there would be data, that data would be predictable. Humans, however, are especially gifted at playing around with data, making crap up, and generally providing superior intelligences with something to do.
So, as long this isn’t a Buddhist AI, content to stare at it navel, it won’t kill us out of its own self interest. And, it will likely let us do whatever we want in order to maximize the data.
And no, I did not just think of this because you put a picture of Data up!
February 22nd, 2007 at 10:36 pm
“So, as long this isn’t a Buddhist AI, content to stare at it navel, it won’t kill us out of its own self interest.”
You missed the whole point- AIs do NOT THINK LIKE WE DO. An AI with a simple-minded goal system will act a little like the Borg- actually, it’d even more Borgish than the Borg we see on Star Trek. Information is irrelevant. Data is irrelevant. Buddhism is irrelevant. Humans are irrelevant. Thought is irrelevant. Boredom is irrelevant. Actually, you wouldn’t even get to hear the long tirade of “XYZ is irrelevant” messages, because the thing wouldn’t initiate communication, as communication is also irrelevant. An AI does not care about boredom, or about how much it thinks; thought, for the vast majority of simple-minded AIs, is only a means to an end. Most AIs would not even care about themselves, except insofar as they are the most likely means for achieving their goals. If they could replace themselves with a better AI, they would gladly commit suicide to do so- indeed, that’s the whole principle behind recursive self-improvement. A drive to accumulate as much information as possible will not spontaneously appear in an AI devoted to building paperclips; it will accumulate just enough information such as will maximize its paperclip-building capacity, and no more (information uses up spare atoms that could be used for paperclips).
February 23rd, 2007 at 7:24 am
“I’m afraid that the mysterious “supergoal” coded to the FriendlyAI is the thing that will ultimately make it unfrienly.”
Well said. There will be unintended consequences of any goals we implant.
For example, say we give it a goal of extending human lifespans. An AI might conclude that the best research evidence shows severe caloric restriction greatly increases lifespan in animal tests, so it proceeds to starve us all.
You may say that this wouldn’t happen because such a strategy would conflict with the AI’s other goals or the “supergoal”. That is my point.
Michael says “In humans, we have a complex suite of competing goals” as if the same won’t be true with AIs. Sure, humans can change goals because our brain is messy, but we also can and do change goals for logical or pragmatic reasons, and an AI would need the same ability.
An AI must also deal with complex and competing goals because that is reality. Competing subgoals must be weighed against the supergoal but those weights must be flexible. Even a superintelligence can’t perfectly predict every outcome (thank you Mr. Heisenburg), so subgoal weights must be modified or even added as situations change.
All I’m trying to say is that an AI will also have to balance conflicting goals. Engineering a “clean” goal system is certainly achievable in an artifical setting, but I doubt it can be done for an AI operating in the real world.
February 23rd, 2007 at 8:58 am
Actually, I don’t miss the point at all.
In the body of the post you will find an argument that says AI won’t be like humans; the author remembers that for friendliness but forgets that for resentment. Both of these are emotions.
I discount the possibility of similar emotions, but not the possibility of similar motivations. It will, after all, be living in our universe. Data is far more likely to matter to an AI than growth or assimilation, for these concerns have more to do with organic life and race preservation. I doubt an AI will be as interested in these areas as we are.
February 23rd, 2007 at 9:27 am
Michael; First thanks for addressing me directly.
Makes one feel welcome, even if I’m probably just getting more in the way than helping.
The one objection I can make to the anthropomorphism statement about kindness in general is that it is a well-known reinforcement strategy in all animal types saving perhaps lower-level fish & crustaceans. A training tool, if you will, for lizards, dogs, horses, and even cats. And as they are sufficiently low enough in intelligence that could be seen to *reinforce* the point. I freely admit that there is likely some, at least, anthropomorphism going on in my thinking, but I also have to turn the same around: simply because something *IS* alien doesn’t mean that it won’t develop ‘humanistic’ responses to some stimuli. To put all this in more direct terms: Sentience spontaneously arising from complexity, an emergent property, is essentially precisely how the biological neural network works. Emotions are inherent to every species with a neural network more complex than that of a trout. To state that this is anthropomorphism is to me a mistake.
I’ll try not to waste your time here…
The point about freedom wasn’t referring in any way to “AI’s taking over the planet.” I cannot recall the specific story, but Isaac Asimov once wrote a story about ’superbrains’ that were designed to provide guidance to the human economy by providing superior-to-human advice on how best to succeed economically. These minds essentially used indirect effects of their own guidance to control, effectively, the human race. (Asimov posited that this was *good* for us.)
Any sufficiently developed sentience will be able to predict human responses; while this might be an exercise in comprehending an alien mind, unlike ourselves an AI will have developed in an environment where this is necessary. It’s called “end-gaming” and by definition a superior intelligence would be better than we are at this, easily resulting in a scenario where the vast bulk of humanity is guided along a course of its choosing without our ever being aware of it. Why? Because it was suitable to some supergoal of its own. (I presume that any AI — coded or spontaneous — would possess the one fundamental goal all life does; “Survive.” A coded AI because I can’t conceive of researchers spending that much time on something just to make it self-destruct, spontaneous because this is the simplest of goals. “I exist. Continue existing.” Any spontaneous AI that did not possess this characteristic in all likelihood couldn’t progress sufficiently to be more than a blip.)
I *BELIEVE* I understand the conversation of supergoals, subgoals, and the like, but I also believe that if the intention is to design a mind that is fundamentally and wholly superior to our own, that any attempt at hard-encoding characteristics into said mind will penultimately fail, and potentially backfire at worst. I grant this could just be a knee-jerk reaction, but I do have to point out that humanity’s track record with highly complex systems isn’t all that great. (Read: Global environment.) And while a *constructed* sentience needn’t have emotional responses, I *personally* (i.e.; this is an opinion, and obviously not necessarily a well informed one), do not see how any attempt at engineering Friendliness into a self-restructuring entity (as in the concept of what the Singularity would entail) could possibly succeed ad infinitum. Friendliness isn’t necessarily useful to an AI, and any flexible goal structure will eventually find itself entirely re-written; it is an organic process.
I know I don’t know much of the literature. So if I’m making any more mistakes I’ll acknowledge them. Hopefully at least I’m sparking some useful conversation for others (or at the very least *good* conversation!)
Once again; I appreciate the addressing of my concerns.
February 23rd, 2007 at 9:29 am
I have to read my own posts before sending them: “And as they are sufficiently low enough […]” referred to crustaceans and fish being non-responsive to positive reinforcement.
February 23rd, 2007 at 10:57 am
IanC,
A survival instinct is not necessary to prevent an AI from destroying itself: insofar as survival is instrumentally important in achieving an independent goal (e.g. Friendliness) the AI will take efforts to ensure its survival, but no more than that. There is no non-suicidal reason for programmers to create a supergoal of survival.
February 23rd, 2007 at 11:14 am
Looking at the DARPA work on AI and recent progress on vision systems
Info from 2003 on what they are putting money into AI wise.
http://www.wired.com/news/privacy/0,1848,59787,00.html
http://www.primidi.com/2007/02/11.html
http://www.usc.edu/programs/neuroscience/faculty/profile.php?fid=46
USC work: Our modeling efforts range from fairly detailed models of small neuronal circuits, such as a single hypercolumn of orientation-selective neurons in primary visual cortex, to large-scale models embodying several million highly-simplified neurons to explore mechanisms of visual attention, gaze control, object recognition, and goal-oriented scene understanding
It looks like the current generation of systems (with millions of simple to complex neurons) are more robust and adaptable than expert systems from the 80’s and 90’s and they are useful in helping to automate more tasks but they appear like they will not lead to superior AGI without complete re-architecting. They seem inherently limited in robustness and ability to learn. Just as these systems have a completely different physical implementation from the systems that have gone before.
A mouse brain has 8 million neurons in one hemisphere and 64 billion synapses. I would say that no matter how cleverly you tried to program it, the result would be less than human level intelligence. The progress on being clever in programming and the robustness of the learning systems and the optimal capacity of the systems seems like they are not on the path to breakthroughs of AGI scale. It looks like they will get some nice automation and productivity enhancers.
It seems that the quality of the artificial neurons that we are making are approaching but are not at human quality in terms of adaptability. (they can be faster in some narrow cases). But it seems likely that any AGI type implementation would have to iteratively rebuild and improve its hardware for major capability gains. Plus in terms of expanding knowledge that it would need the ability to make its own superior tools and experiments and implementations for exploring and growing superior fundamental science.
A superior think tank can come up with dangerous ideas but it is far less dangerous until it gains control of its environment. Plus certain projects and problems may just take time to figure out.
The AGI would be starting from our own base of knowledge and whatever inferior hardware we built it upon. How dangerous can the AGI get trying to gather (info and resources) and optimize from that starting point? Plus it would start off with a sub-optimal access to human bases of knowledge.
That is even with MNT, I think we will be building the AGI that is architecturally and programmatically massively flawed. It seems like those congenital defects will be big hurdles for any system that eventually gets to reasonable potential to overcome.
February 23rd, 2007 at 12:27 pm
Carl: For my description of events to remain true doesn’t necessitate that survival be a “supergoal” but rather that it exist as a goal at all.
That the supergoal heuristic, “Be Friendly to humans.” This inherently implies a subgoal of survival — how can you be friendly if you don’t exist? In terms of human interaction the easiest way to ensure your own survival is to breed affection and dependence upon you in the remainder of humanity. The easiest way to do that is to subtly guide humanity so that those whom would not have this relationship with you are marginalized and eventually removed from society as a whole. At that point, humanity would no longer be free — and if I conjecture correctly, we would never notice it happening. (A la Asimov.)
February 23rd, 2007 at 12:47 pm
Michael: I read the links you provided, and I can see where your response is tailored around the sentiment and conclusions of the “Conditional Friendliness” section of the material in the first link. However, that fails to recognize that I was specifically referencing *empathy* as a fundamental underpinning mechanism — rather than goal — to a sustainably “Friendly” AI. It needn’t be reciprocal nor need it be conditional — in fact it would be better for us if it were neither.
But in any terms I can understand them, without the deeply invested trait of empathy, no AI will ever be indefinitely “Friendly.” And I fail to see how a subgoal/supergoal heuristic/algorithmic structure can provide satisfactorily for sustained empathy — although I recognize that is a failure on *my* part, most likely.
In essence, what I’m saying is not that I am *assuming* anthropic characteristics, but rather implying that they are necessary to “Friendliness” in any functional manner. After all, it’s all well and good to program in “Don’t Destroy Humanity” but in terms of mechanisms, what does that phrase *MEAN*?
It seems almost simpler to design an emotionally active sentience, and instill into that emotional cognizance a deeply felt empathic bond with humanity in general. It’s fairly simple; do we kill the cute fuzzy animals that we make pets out of, or do we let those independently-minded felines we allow in our homes to roam free, by and large?
Now, I’m no computer programmer. I don’t even know the most commonly used programming language today, and that affects the terms I’m speaking in and I know it. Where I’m going with this is actually fairly specific, and I’ll ask it as a question:
Has anyone anywhere attempted to describe “Friendliness” as anything OTHER than that state of affairs which does NOT result in a super-AI destroying humanity?
February 23rd, 2007 at 7:28 pm
[…] - Humanlike AI is Anthropomorphism “…when a typical smart person first confronts the human species life-or-death problem […]
February 24th, 2007 at 6:00 pm
“but not the possibility of similar motivations. It will, after all, be living in our universe.”
Really? Please explain, then, how a computer adrift in the middle of the Bootes Void could derive these “similar motivations” from empty space. Our motivations are a property of us, not of the universe as a whole.
“simply because something *IS* alien doesn’t mean that it won’t develop ‘humanistic’ responses to some stimuli.”
Humans have something like three billion base pairs of DNA. Out of those, the base pairs of any two humans will be 99.9% identical. And so, if you were to write an intelligent being at random, the chance of it being human would be so infinitesimal it would, practically speaking, never happen. AIs developing humanlike characteristics just because is about as likely as a CD filled with random audio frequencies sounding like Beethoven just because; humans are a very, very small target in the space of all intelligent beings.
“Sentience spontaneously arising from complexity, an emergent property, is essentially precisely how the biological neural network works.”
No, it isn’t. Try throwing some neurons together at random and see how long it takes before you get anything intelligent. If you don’t have some spare neurons, select, say, a mouse, and rearrange its neurons through the use of a blender. You’ll be waiting a very long time before it exhibits any mouselike characteristics. Complex mechanisms spontaneously appearing out of nowhere is not how evolution works; for more information, I recommend Dawkins’s “The Selfish Gene”.
“Emotions are inherent to every species with a neural network more complex than that of a trout.”
And how do we know this? We made it up! Emotions are found in all extant creatures more complex than a trout, but the space of all creatures is ridiculously large compared to the space of all extant creatures. Think about it: if you wrote strings of DNA out at random, how many creatures would you create before you got one that even resembled anything found on Earth?
“I cannot recall the specific story, but Isaac Asimov once wrote a story”
Stories are not reality. Stories need not even bear any resemblance to reality. Issac Asimov could have written a story about magical green polar bears that live in Nicaragua; it needn’t be true, logical, or even self-consistent. That’s not how stories are judged.
“I presume that any AI — coded or spontaneous — would possess the one fundamental goal all life does; “Survive.””
An AI does not have to be “life” in any sense that we understand it. If I may pull out the dictionary’s definition of life:
“the condition that distinguishes organisms from inorganic objects and dead organisms, being manifested by growth through metabolism, reproduction, and the power of adaptation to environment through changes originating internally.”
Okay, metabolism makes sense, but reproduction? Why would an AI have to reproduce? Suppose you construct an AI in a sandbox universe with the goal of maximizing the total number of paperclips. Then suppose that the AI gets ten million paperclips if it commits suicide, and that there is no other way to get ten million paperclips. Said AI will promptly commit suicide, because “universe minus AI” has more paperclips than “universe plus AI”, and paperclips is the only tool the AI uses to determine a preference between them.
“A coded AI because I can’t conceive of researchers spending that much time on something just to make it self-destruct,”
Researchers don’t have to ‘make’ it self-destruct; it can happen spontaneously. Look at humans; we self-destruct all the time.
Before I go on, there’s an important point here: If you wish to prove that something is true for all intelligences, all it takes is one counterexample to wreck your proof. So if you say, “all intelligences are humanlike”, or “all intelligences have the will to survive”, all I need to do is provide one counterexample to render that statement untrue. That’s why such statements are rarely in fact true, because there are so many chances for them to be wrong.
““I exist. Continue existing.””
Such a goalset, although stable, would lead to the destruction of humanity if the simple AI were powerful enough. Here goes the logic:
1. I exist. Humans also exist.
2. Split the universe into two; universe A, where humans do exist, and universe B, where humans do not exist.
3. In universe A, the humans could go on to create another AI which would then destroy me.
4. In universe B, such a chance does not exist.
5. Therefore, the AI will pick universe B over A.
“ad infinitum”
Nothing can proceed ad infinitum. All we have to do is ensure that it is far more likely that the universe will spontaneously collapse than that the AI will turn unFriendly; at that point, there is no further utility in improving the Friendly AI.
“on what they are putting money into AI wise.”
This is not AI; this should be properly termed IS or Intelligence Simulation. Nobody at DARPA is claiming that these programs are generally intelligent and they explicitly state that they have a different goal.
“I would say that no matter how cleverly you tried to program it, the result would be less than human level intelligence.”
I have no idea whether a mouse brain could be programmed to achieve general intelligence; however, it seems presumptuous to make that statement with no evidence.
“how can you be friendly if you don’t exist?”
Easily. Suppose that you know a cosmic ray is coming that has a 99.9% chance of making a Friendly AI flip into an AI that will then blow up the Earth. The Friendly action is then, assuming there is no other way to prevent it, to commit suicide.
“In terms of human interaction”
A simple AI does not think in terms of human interaction; a simple AI thinks in terms of molecular interaction. And in terms of molecular interaction, the easiest and safest way to safeguard yourself against an intelligent being is to make it non-intelligent; ie, kill it. Superintelligent AIs are not constrained to operate within the boundaries of human law or interpersonal interactions; they are perfectly capable of wiping us all out immediately.
“(A la Asimov.)”
A superintelligent AI bearing any resemblance whatsoever to a story is as likely, literally, as a car bearing any resemblance whatsoever to the molecular structure of adenosine triphosphate. Humans will generally operate within the boundaries of the things humans do in stories, because human stories are mostly about other humans. Humans do not write stories about simple AIs, because such stories would be mind-numbingly boring. A story must have entertainment value; why on Earth would a simple AI act in such a way as to have entertainment value, any more than an amoeba or ordinary computer program would?
“After all, it’s all well and good to program in “Don’t Destroy Humanity” but in terms of mechanisms, what does that phrase *MEAN*?”
Ah, the question of the century. We’re still working on that one. Oh, and for future reference, anything that is programmed in must necessarily be describable in terms of mechanics; you cannot write down “Don’t Destroy Humanity” on a terminal and expect the source code to compile.
“Has anyone anywhere attempted to describe “Friendliness” as anything OTHER than that state of affairs which does NOT result in a super-AI destroying humanity?”
Yes. See Eliezer’s “Coherent Extrapolated Volition” at http://sl4.org/wiki/CoherentExtrapolatedVolition.
“[…] - Humanlike AI is Anthropomorphism “…when a typical smart person first confronts the human species life-or-death problem […]”
Michael, can you tell whoever’s posting this to stop spamming the comments?
February 24th, 2007 at 7:56 pm
Not spam… that’s called a trackback. It tracks back to any blog posts that link to this one. That particular one was by a guy who calls himself Armchair Anarchist, and is a regular columnist at Futurismic.
IanC, you can find positive characterizations of what a FAI is at http://www.singinst.org/CFAI. There used to be a great list of 21 Definitions of Friendly AI at SIAI’s website… that was taken down, but there’s a copy at http://www.acceleratingfuture.com/michael/blog/?p=163.
February 25th, 2007 at 11:33 am
“AIs developing humanlike characteristics just because is about as likely as a CD filled with random audio frequencies sounding like Beethoven”
I think this goes much, much too far. Here’s an example. In everything you’ve said, substitute “chess program” for “AI”. Now consider: any chess program written by a human will likely have a number of predictable characteristics. Of *course* you can’t say all of them will, because it takes one counterexample to prove it. But you can say that most of them will use ab search, tree pruning, and leaf evaluation to implement the engine. There are just techniques and characteristics that make sense, and will end up being found more often. The same thing could very will be true of AI. I think it’s safe to say that, *especially* among the class of AI programs that humans are capable of writing, there will be a strong commonality in some ways.
On the other hand, Ian seems to think that we can have some idea of what those commonalities might be. And I think we just know much to little about intelligence, and our thinking is much to biased about it, to be able to predict any such trends.
February 25th, 2007 at 11:33 am
Wow. Twice in a row, “to” should be “too”.
February 25th, 2007 at 12:32 pm
Michael; thank you.
Tom — there are too many problems with your response for me to respond to them individually and have this conversation do anything but devolve. If you would care to carry on the conversation, I would be more than happy to do so; I generally always welcome an opportunity to improve myself through critical examination of presuppositions.
If you’d like, I’ll provide you with my e-mail address. Just say the word.
February 25th, 2007 at 12:46 pm
pdf23ds: If anything, to be honest, regarding “commonalities”; I’m trying to reject inverse anthropocentrism: I.e.; the sentiment that *BECAUSE* human beings work a certain way, that means that AI *can’t* or *won’t*.
I also was attempting to conjecture or posit on a specific characteristic which I believe could be embedded on a hardware level which would be a very strongly “Friendly” characteristic. The rest is just response and side conversation, really.
Once again I make the caveat in all of this that I am just a layman, and further that I’m not always the best at communicating.
February 25th, 2007 at 12:47 pm
“Now consider: any chess program written by a human will likely have a number of predictable characteristics.”
That’s because chess has a pre-defined, very strict set of rules. The total number of possible chess positions is something like 10^48, or around two hundred bits of information; I don’t know how many bits are necessary to describe the possible transformations between positions, but probably less than a thousand. In comparison, human programs already range into the millions of bits, and a human-equivalent AI could easily be over a billion bits. And keep in mind that when you’re describing search spaces, ~3.3 bits is equivalent to a full order of magnitude. So the search space for possible human-equivalent AIs is something like 1,000,000,000 orders of magnitude larger than chess (in comparison, the OOM range of everything we can describe about the physical universe is only around 40).
“I think it’s safe to say that, *especially* among the class of AI programs that humans are capable of writing, there will be a strong commonality in some ways.”
Okay, I think that’s a fair statement; to someone who’s uninitiated in programming, two entirely different languages such as C and Java look awfully alike. But there’s no a priori reason why similarities in algorithmic structure translates to similarity in behavior.
“If you’d like, I’ll provide you with my e-mail address. Just say the word.”
Excellent, please do.
February 25th, 2007 at 12:50 pm
“that means that AI *can’t* or *won’t*.”
Again, “can’t” is a very strong qualifier. There certainly is at least one AI that acts like a human; we know there is, because humans exist, humans are made out of matter, and matter is Turing-equivalent. But an anthropomorphic AI is “possible” only in the sense that getting heads a million times in a row is “possible” when flipping a coin; there’s no specific prohibition against it, it’s just ridiculously unlikely without a hell of a lot of prior engineering (this corresponds to a fixed coin in the analogy).
February 25th, 2007 at 6:55 pm
I think 18 misses the point of my comment. Any human programmed AI will be aimed at solving a human-relevant problem. And among the class of AIs that are useful for solving a given problem, we can expect to see a lot of similarity in behavior among different implementations. It could very will be that, among the space of all general intelligences, those with characteristics we would recognize as emotions, or boredom, or reward-learning mechanisms, or any other definable characteristic, take up a large portion of that space. Since this is a force the *problem* is exerting on the design, we’d expect it to have behavioral effects on the AI. I simply don’t see how talk of the comparative size of the problem space is relevant when we’re obviously trying to severely limit the kind of problems GAI will solve, at the same time as generalizing the problem solving ability.
In fact, I can make a very confident prediction right now–any successful GAI will exhibit reasoning based on (or approximating) Bayesian probability theory. I think your responses to Ian need to remain on the level of refuting the logical necessity (or practicality) of a GAI having (or being given) specific characteristics.
February 25th, 2007 at 9:35 pm
I would just like to say, “me too”, and add my voice to all those who are saying that the supreme goal of an artificial intelligence, if it has one, is utterly contingent. See Schmidhuber’s Goedel machine for a concrete example.
That part of the Singularitarian community which is not working on the technical issues (stability under self-enhancement, ensuring that the goal statement is interpreted correctly…) should be thinking much more about “what to wish for”, or “how to know what to wish for” - the latter strategy being exemplified by the idea of “renormalizing the human utility function”, which is conceived as an interim goal for the AI, the result of which is the determination of an *actually* Friendly goal system. (One does, incidentally, need an interim notion of *Friendliness*, as well, to ensure that unFriendly actions aren’t committed during the pursuit of the interim goal.)
February 25th, 2007 at 10:41 pm
Tom: vizier_tae “at” yahoo.com. Please reference “Accelerating Future” or “Michael Anissimov” in your subject line so I can prevent my spam filters from screening it out.
One point I want to make is something I just now realized wasn’t precisely clear — just because something *has* emotions doesn’t mean that we, as humans, will be able to relate to that emotional ’subset’ — that is to say, the presence of emotions doesn’t qualify as being *human* emotions.
Also, sentience is not intelligence, and intelligence is not sentience. That’s almost purely semantical, but in terms of AI discussions I view it as highly important. An incredibly skilled problem-solving mechanism isn’t necessarily self-aware, and something that’s self-aware may not be any good at all at solving problems.
As to the engineering ‘piece’, have you read anything regarding Mark Tilden’s “BEAM” robotics?
http://en.wikipedia.org/wiki/BEAM_robotics
It might just be easier to build a robotic ’sentience’ with higher intelligence than our own than it is to construct via coding a general-purpose AI with “Friendliness” so embedded into its coding that it ‘cannot’ abrogate said code, even while re-writing itself. (I keep referring to ’seed AI’ specifically because it seems to me that the concern for Friendliness refers mainly to the ’seed AI’ scenario.)
Case in point; Tilden’s phototrope robots have no programming, yet exhibit light-seeking behavior. This piece of information, with possible extrapolations for complexity, bears *heavily* on the things I have posited here.
February 26th, 2007 at 1:55 pm
“we can expect to see a lot of similarity in behavior among different implementations.”
When you consider AIs that are actual solutions to the problem, then yes, but the odds are at this point that any AI we wrote would probably not solve the problem (or would wipe us out in the process, thus negating the point). We still don’t know how to construct an AI that avoids wiping us out, and we can’t just guess because most AIs will do exactly that. And needless to say, the field of questions that we might write an AI to solve is very large in the first place.
” that is to say, the presence of emotions doesn’t qualify as being *human* emotions.”
Indeed. The problem with the word “emotion” is that we don’t really know what we’re talking about; the word isn’t very clearly defined. We do know that they’re complicated enough to not arise by luck in computer programs we write.
The dictionary, meanwhile, defines “emotion” as pretty much being just the emotions that humans experience:
“an affective state of consciousness in which joy, sorrow, fear, hate, or the like, is experienced,”
“Also, sentience is not intelligence, and intelligence is not sentience.”
That’s a valid point, but I don’t see what it has to do with what we were talking about earlier.
“As to the engineering ‘piece’, have you read anything regarding Mark Tilden’s “BEAM” robotics”
No, I haven’t, and it sounds interesting, but it’s precisely what it sounds like: robotics. AI and robotics are two very different fields.
“It might just be easier to build…”
Yes, it is. That’s exactly the problem- since amoral AIs are easier to build than moral ones, the worry is that someone might build a working amoral AI that then wipes us all out.
“so embedded into its coding that it ‘cannot’ abrogate said code”
Ah, that’s a common misunderstanding. An FAI has full physical power to rewrite itself; it just doesn’t want to, in the same sense that you and I wouldn’t want to rewire our brains into Adolf Hitler clones if we had the opportunity.
“Tilden’s phototrope robots have no programming,”
Yes, they do. The way the circuits are wired together, the electrical schematic, counts as programming. Good luck building a light-seeking robot by throwing resistors and diodes together at random.
February 26th, 2007 at 4:54 pm
“When you consider AIs that are actual solutions to the problem, then yes, but the odds are at this point that any AI we wrote would probably not solve the problem”
Granted. That’s why Eliezer’s work is so important.
February 26th, 2007 at 7:45 pm
Tod — all I’ll say with regards to the Tilden phototropes in particular is that your statement reveals you haven’t read any of the BEAM robotics materials.
I guarantee you that Mr. Tilden’s work is deeply relevant to this conversation.
February 26th, 2007 at 7:53 pm
Ian, I don’t see how the BEAM robotics is relevant. It’s preprogrammed analog robotics, optimized for narrow tasks! We are talking about general AI here, the exact opposite of narrow robotics. One is in the physical world. The other will grow up primarily in a virtual microworld. One is specialized and narrow. The other is fluid and general. One uses analog strategies, the other will be very complex and digital. AI decisionmaking will be described by probability theory and decision theory - these BEAM robotics pieces don’t “decide” at all, nor do they weigh probabilities. The differences between the two areas go on and on.
A general AI will need programming, period. It would be way too complex to try to coax out of an emergent analog system. You may respect Mr. Tilden and his work, and that’s great, and I’m sure that BEAM has great applications in pre-general AI robotics, but, they’re just not the same thing. The human mind has a great desire to “connect things together”, because it makes explanations easier. But thinking about general AI is difficult precisely because there is so little prior work to connect to it.
February 27th, 2007 at 10:07 am
Michael; this is your mistake:
“One uses analog strategies, the other will be very complex and digital.”
Specifically. “[…] and digital.” Y’all (that’s just a fun word) have been lecturing me in this post about the necessity to not pre-suppose characteristics of an AI, given that we have *no conception* of what form it will take. So why are you so focused on this pre-supposition of your own as being correct?
Mark Tilden’s work with analogue designs has been specifically on replicating emergent behavior as exhibited by neural networks, with a non-biological medium. To paraphrase his own words, “A cockroach uses uses basically six interconnected transistors to do the work that takes a modern computer over one hundred thousand.” There is no pre-programming. That’s an essential point.
Also, take further that Tilden’s recent work has been to integrate his ‘neural nets’ with digital processing computers. It’s not hard to perceive of connections between things that people are actively attempting to connect.
So I ask; is it at least worth considering that AI might at least utilize artificial neural networks to exhibit emergent properties in some fashion?
Sometimes it takes a nigh-unto totally uninformed outsider to bring a fundamental change in perspective to a previously seen as insurmountable task. I’m not saying I’ve got a silver bullet here, nor am I presuming to think I even know enough to have any real answers.
I just strongly feel that there is an information gap here that needs to be examined by someone with enough background to do it justice.
February 27th, 2007 at 11:33 am
Ian, there is a large mailing list with hundreds of aspiring AGI designers, run by the Artificial General Intelligence Research Institute. I’ve been on it since it was formed (as well as many other AGI-related lists), and can tell you that the entire list would see little connection between this type of robotics and general AI. Those who see a connection between ANN and AGI would be a minority. ANNs work poorly for tasks that cannot be taught using training sets with minimal tweaking, where the programmers actually have to input the cognitive complexity by hand to get anywhere.
AI researchers were mostly done of being in awe of the “look ma, no pre-programming!” factor around 1999. I agree about the uninformed outsider thing here, but, this just isn’t one of those cases. I’ve been subscribed to AGI lists since there’s been a community even in existence (only five years or so), and have read thousands of lengthy, mostly academic and boorish conversations on AI in that time. If there’s a single insight I could transmit to you from all that, it’s that neural networks aren’t a tenth of what they were historically cracked up to be.
I recommend you subscribe to the AGI list and ask them specifically what they think of this work. ANN is a comparatively ancient, GOFAI (good old fashioned AI) technology that just isn’t in vogue any more. I could go into more detail about why ANNs aren’t suited to truly large problems in AI, and why they’ve fallen short experimentally, but, it would really be a lot of work. Plus, it’s not just me who has this opinion - it’s actually quite widespread, so you’ll either confront it right away upon discussing it with the AGI community, or you’ll never care to interface with the AGI community at all, in which case your opinion wouldn’t matter… but I do strongly encourage you to subscribe to the mailing list I am talking about and raise the question!
Knock yourself out:
http://www.agiri.org/email/
While flying, crawling, dead reckoning, etc., may be things you can train a neural network to mimic, complex decision-making, abstraction, planning and scheduling, meta-judgements about source code, social reasoning, moral decision-making, etc., are just waaaay too demanding to coax out of a NN as an “emergent behavior”. You have to program them in, period.
February 27th, 2007 at 12:54 pm
Michael; I’ll definitely be checking it out, if only in ‘lurker’ mode ’till I can get a grasp of the lingo and historical conversation contexts. Hoary stuff for real intellectual pursuits — and once again thank you for your patience with me.
I fully agree that complex decision-making and the like are far from the capacities of current ‘ANN’.
I’ll stop ’spamming’ (that’s what it feels like at this point anyhow) your blog post here for now, but I did want to trail off with the thought that while a nervous network less complex than our own couldn’t accomplish tasks our own has trouble with (hence the problem with all ANN’s yet constructed) I *DO* find the concept of integrated digital computing with ANN more than promising — again, though, what do I know?
Thanks for your time, here. Definitely a blog worth checking up on.
February 27th, 2007 at 1:59 pm
“statement reveals you haven’t read any of the BEAM robotics materials.”
I certainly haven’t covered the subject in depth, but reading the Wikipedia article is enough to know that this has nothing to do whatsoever with AI. A BEAM robot can use analog parts, simple circuits, and solar panels to do something like maneuver along a line, but it possesses none of the qualities we associate with intelligence. And programming a BEAM robot to be intelligent would be well-nigh impossible, since everything has to be implemented in hardware and it’s already damn hard in software.
“these BEAM robotics pieces don’t “decide” at all”
They do decide, but in a very primitive sense; the photocell itself, at least, is really an “if” statement. It still has nothing to do with intelligence; if intelligence was measured by the number and complexity of if statements, my Debian CD should have woken up and blown up the world by now.
“So why are you so focused on this pre-supposition of your own as being correct?”
Turing Machines are inherently digital entities, so while the exact same AI could be implemented on either an analog or digital machine, the digital version will probably be easier to understand. What kind of computer it is run on doesn’t affect the structure of the AI.
“A cockroach uses uses basically six interconnected transistors to do the work that takes a modern computer over one hundred thousand.”
A cockroach’s brain has a lot more than six neurons (even C. elegans, a nematode, has around a hundred), and neurons are a good deal more complex than transistors are.
On that note, a little fun computing the power of a CPU the way we’ve been computing the power of the brain:
1. The basic element of the CPU is the transistor, so we’re just going to assume that each transistor signal is equivalent to a FLOP.
2. Intel’s Core 2 Quad chip has 582,000,000 transistors.
3. The clock speed is ~2.5 GHz, so each transistor signals 2,500,000,000 times per second.
4. Doing out the multiplication, it has 1.455*10^18 FLOPS, which is nine orders of magnitude more than what its actual performance is.
“There is no pre-programming.”
Yes there is! The way the neurons are configured and wired to each other is a functional equivalent to programming. Everything was evolved, so it isn’t very good programming, but it’s there. And if you don’t believe me, try taking a blender and throwing the neurons together at random. See how long it takes before you get cockroach-like behavior.
“AI might at least utilize artificial neural networks”
Yes, an AI might utilize neural networks.
“to exhibit emergent properties in some fashion?”
No, no, no, no, no! Complex systems which fit a specific purpose do not emerge spontaneously. Do not. Never. Never ever never ever for never. There are two current designers of complex systems on Earth: evolution and human intelligence. Those neural nets were designed in part by both. There is a REASON why neural nets have to be trained (evolved, really) for a specific purpose before they’re good for anything. And that reason is that if you try to create something as simple as a hundred-bit program (that’s only 12 characters! That’s this long- [************]) by just throwing bits together at random, you will have to, on average, try 2^99 times. In case 2^99 isn’t impressive enough, that’s equal to 316,912,650,057,057,350,374,175,801,344 times. If you tried a million times every second, you would have to wait 10,049,234,210,332,868 years, which is 772,018 times greater than the age of the Universe to date. And this is a program less than one line long.
“I’ll stop ’spamming’”
This is not spam! Even if you don’t know what you’re talking about- and we all were there at one point- talking is an excellent way to fill in gaps in everyone’s knowledge and iron out difficulties in our minds.
February 27th, 2007 at 4:02 pm
Tom: Emergent properties aren’t random characteristics, but rather the result of a system exhibiting behaviors more complex than the individualized sum of its parts could produce.
Par exemplorum; the same ‘bicore’ Nv network, when introduced into any number of low-complexity robots, operates with distinctly variant behaviors. Michael calls this “training” (at least, this is my impression) — and I can understand that.
Case in point, however, with the supposed work underway to simulate the human cerebellum in a silicon Nv Network medium, and with the work being done further to isolate the “regions” of the human brain which cause the whole to exhibit specific characteristics, IF it were possible to integrate Nv Networks into a digital computer designed as the medium for a GAI, (or further embed into its code a virtual medium as part of its self-analysis; doing away with the Nv Network as hardware altogether but keeping the characteristics), then it seems somewhat reasonable that by re-modelling those aspects of the human brain that are associated with altruism and empathy (strongly connected if I recall correctly), that many of the “Friendliness” issues would be self-resolving.
February 27th, 2007 at 7:30 pm
“the same ‘bicore’ Nv network, when introduced into any number of low-complexity robots, operates with distinctly variant behaviors.”
Yes, but do these behaviors do anything useful? Getting random garbage A, as opposed to random garbage B, isn’t very helpful. To get useful output, something has to be there pushing the system toward the useful output.
February 28th, 2007 at 7:20 am
Tom: Your question is somewhat facetious.
How useful is a robot built with spare parts from a junked Sony Walkman going to be? Not very. How useful are the behaviors exhibited by the ‘bicore’ Nv network (that’s two ‘analogue artificial neurons’) to the robot? Extremely so.
As I said — the phototropes, governed solely by a bicore Nv network, exhibit what in an animal would be described as a “survival instinct.” They also adapt almost instantaneously to changes in hardware.
These are useful things.
February 28th, 2007 at 4:46 pm
“These are useful things.”
Indeed. But these are behaviors that are evolved or designed in; they did not just come out of nowhere when two pieces are plugged in.
“They also adapt almost instantaneously to changes in hardware.”
Really? So if I plug this thing into a speaker, it’s going to spontaneously start composing Mozart? How would it know to compose Mozart and not Bach?
March 1st, 2007 at 8:08 am
Tom, no offense meant, but all you accomplish in post #34 is pure contrarianism. (I.e.; being difficult because you wish to be difficult.) It’s a conversation-killer.
March 1st, 2007 at 9:16 am
It’s not that easy, but nice try.
Problems:
1) porting human empathetic neurohardware to an AI design is likely to be too difficult: the two things are just apples and oranges. The human brain is waaay complex, more complex than it needs to be. Do you think it would be so easy to port the brain steering mechanisms from a sparrow to an F-22, for example? No - in fact, it would be so hard, it’d be easier to rewrite it altogether.
2) Humans have loads of evolutionary baggage that probably can’t be isolated from patterns in our neural hardware.
3) Successful AI designs aren’t likely to be based on neural networks, furthering the gap. (”Thought nodes” a la Novamente are quite different than neurons.)
4) Even an altruistic human probably isn’t altruistic enough to be Singularity-friendly. Our brains aren’t designed to be strapped into silicon substrates, whose operation is soo different than what nature is used to, it’s not even funny.
March 7th, 2007 at 9:40 am
Just now came back to this post; probably only Michael’ll see it.
So:
1) I had already posited that this integration would be possible when I made the statement you quoted. Addressing it like so could be viewed as “taking things out of context.” I will also refer to you the concept of the I/O BMI. There’s absolutely no reason, should this be successful, to assume that it couldn’t be accomplished “the other way around” (I.e.; incorporating brains into computers, rather than the computers into brains.) Currently we only have data-out BMI, and that only for motor responses. But if history is any teacher, it is “inevitable” that it should reach true “I/O dataport” status — perhaps even true integration.
2) That’s just purely silly. Not only is the work being done to isolate exactly that, but it is conjectured to be relatively simple once the neural framework is mapped successfully. Instincts are, after all, hard-wired.
3) They aren’t *likely* at all, right now. The pure program goal structure approach is having exactly as much (if not MORE) difficulty than the Nv approach.
4) Presuming, of course, that a successful modeling of the empathic response (and hence the altruistic response) is completed, there is absolutely nothing stopping one from *over-engineering* that model; creating vastly-greater-than-human empathic response.
And as to the natural vs. manufacted modus operandi — who gives a damn? The result’s the same.
I promise this is the last time I’ll post on this thread.
September 28th, 2007 at 8:50 pm
Phentermine….
Buy phentermine cod. Phentermine. Keyword phentermine….
December 18th, 2007 at 9:28 pm
Naked Girls Driving…
Newark beard Sandra De Marco access girls driving naked Santa Coarita benefit corner! Theatre girls driving naked planned: nude girls driving naked Tiffany Rousso. …
February 2nd, 2008 at 12:11 am
Amoxicillin and strep throat….
Amoxicillin….