Magical Categories

Followup toAnthropomorphic Optimism, Superexponential Conceptspace, The Hidden Complexity of Wishes, Unnatural Categories

'We can design intelligent machines so their primary, innate emotion is unconditional love for all humans.  First we can build relatively simple machines that learn to recognize happiness and unhappiness in human facial expressions, human voices and human body language.  Then we can hard-wire the result of this learning as the innate emotional values of more complex intelligent machines, positively reinforced when we are happy and negatively reinforced when we are unhappy.'
        -- Bill Hibbard (2001), Super-intelligent machines.

That was published in a peer-reviewed journal, and the author later wrote a whole book about it, so this is not a strawman position I'm discussing here.

So... um... what could possibly go wrong...

When I mentioned (sec. 6) that Hibbard's AI ends up tiling the galaxy with tiny molecular smiley-faces, Hibbard wrote an indignant reply saying:

'When it is feasible to build a super-intelligence, it will be feasible to build hard-wired recognition of "human facial expressions, human voices and human body language" (to use the words of mine that you quote) that exceed the recognition accuracy of current humans such as you and me, and will certainly not be fooled by "tiny molecular pictures of smiley-faces." You should not assume such a poor implementation of my idea that it cannot make discriminations that are trivial to current humans.'

As Hibbard also wrote "Such obvious contradictory assumptions show Yudkowsky's preference for drama over reason," I'll go ahead and mention that Hibbard illustrates a key point:  There is no professional certification test you have to take before you are allowed to talk about AI morality.  But that is not my primary topic today.  Though it is a crucial point about the state of the gameboard, that most AGI/FAI wannabes are so utterly unsuited to the task, that I know no one cynical enough to imagine the horror without seeing it firsthand.  Even Michael Vassar was probably surprised his first time through.

No, today I am here to dissect "You should not assume such a poor implementation of my idea that it cannot make discriminations that are trivial to current humans."

Once upon a time - I've seen this story in several versions and several places, sometimes cited as fact, but I've never tracked down an original source - once upon a time, I say, the US Army wanted to use neural networks to automatically detect camouflaged enemy tanks.

The researchers trained a neural net on 50 photos of camouflaged tanks amid trees, and 50 photos of trees without tanks. Using standard techniques for supervised learning, the researchers trained the neural network to a weighting that correctly loaded the training set - output "yes" for the 50 photos of camouflaged tanks, and output "no" for the 50 photos of forest.

Now this did not prove, or even imply, that new examples would be classified correctly.  The neural network might have "learned" 100 special cases that wouldn't generalize to new problems.  Not, "camouflaged tanks versus forest", but just, "photo-1 positive, photo-2 negative, photo-3 negative, photo-4 positive..."

But wisely, the researchers had originally taken 200 photos, 100 photos of tanks and 100 photos of trees, and had used only half in the training set.  The researchers ran the neural network on the remaining 100 photos, and without further training the neural network classified all remaining photos correctly.   Success confirmed!

The researchers handed the finished work to the Pentagon, which soon handed it back, complaining that in their own tests the neural network did no better than chance at discriminating photos.

It turned out that in the researchers' data set, photos of camouflaged tanks had been taken on cloudy days, while photos of plain forest had been taken on sunny days. The neural network had learned to distinguish cloudy days from sunny days, instead of distinguishing camouflaged tanks from empty forest.

This parable - which might or might not be fact - illustrates one of the most fundamental problems in the field of supervised learning and in fact the whole field of Artificial Intelligence:  If the training problems and the real problems have the slightest difference in context - if they are not drawn from the same independently identically distributed process - there is no statistical guarantee from past success to future success.  It doesn't matter if the AI seems to be working great under the training conditions.  (This is not an unsolvable problem but it is an unpatchable problem.  There are deep ways to address it - a topic beyond the scope of this post - but no bandaids.)

As described in Superexponential Conceptspace, there are exponentially more possible concepts than possible objects, just as the number of possible objects is exponential in the number of attributes.  If a black-and-white image is 256 pixels on a side, then the total image is 65536 pixels.  The number of possible images is 265536.  And the number of possible concepts that classify images into positive and negative instances - the number of possible boundaries you could draw in the space of images - is 2^(265536).  From this, we see that even supervised learning is almost entirely a matter of inductive bias, without which it would take a minimum of 265536 classified examples to discriminate among 2^(265536) possible concepts - even if classifications are constant over time.

If this seems at all counterintuitive or non-obvious, see Superexponential Conceptspace.

So let us now turn again to:

'First we can build relatively simple machines that learn to recognize happiness and unhappiness in human facial expressions, human voices and human body language.  Then we can hard-wire the result of this learning as the innate emotional values of more complex intelligent machines, positively reinforced when we are happy and negatively reinforced when we are unhappy.'

and

'When it is feasible to build a super-intelligence, it will be feasible to build hard-wired recognition of "human facial expressions, human voices and human body language" (to use the words of mine that you quote) that exceed the recognition accuracy of current humans such as you and me, and will certainly not be fooled by "tiny molecular pictures of smiley-faces." You should not assume such a poor implementation of my idea that it cannot make discriminations that are trivial to current humans.'

It's trivial to discriminate a photo of a picture with a camouflaged tank, and a photo of an empty forest, in the sense of determining that the two photos are not identical.  They're different pixel arrays with different 1s and 0s in them.  Discriminating between them is as simple as testing the arrays for equality.

Classifying new photos into positive and negative instances of "smile", by reasoning from a set of training photos classified positive or negative, is a different order of problem.

When you've got a 256x256 image from a real-world camera, and the image turns out to depict a camouflaged tank, there is no additional 65537th bit denoting the positiveness - no tiny little XML tag that says "This image is inherently positive".  It's only a positive example relative to some particular concept.

But for any non-Vast amount of training data - any training data that does not include the exact bitwise image now seen - there are superexponentially many possible concepts compatible with previous classifications.

For the AI, choosing or weighting from among superexponential possibilities is a matter of inductive bias.  Which may not match what the user has in mind.  The gap between these two example-classifying processes - induction on the one hand, and the user's actual goals on the other - is not trivial to cross.

Let's say the AI's training data is:

Dataset 1:

  • +
    • Smile_1, Smile_2, Smile_3
  • -
    • Frown_1, Cat_1, Frown_2, Frown_3, Cat_2, Boat_1, Car_1, Frown_5

Now the AI grows up into a superintelligence, and encounters this data:

Dataset 2:

  •  
    • Frown_6, Cat_3, Smile_4, Galaxy_1, Frown_7, Nanofactory_1, Molecular_Smileyface_1, Cat_4, Molecular_Smileyface_2, Galaxy_2, Nanofactory_2

It is not a property of these datasets that the inferred classification you would prefer is:

  • +
    • Smile_1, Smile_2, Smile_3, Smile_4
  • -
    • Frown_1, Cat_1, Frown_2, Frown_3, Cat_2, Boat_1, Car_1, Frown_5, Frown_6, Cat_3, Galaxy_1, Frown_7, Nanofactory_1, Molecular_Smileyface_1, Cat_4, Molecular_Smileyface_2, Galaxy_2, Nanofactory_2

rather than

  • +
    • Smile_1, Smile_2, Smile_3, Molecular_Smileyface_1, Molecular_Smileyface_2, Smile_4
  • -
    • Frown_1, Cat_1, Frown_2, Frown_3, Cat_2, Boat_1, Car_1, Frown_5, Frown_6, Cat_3, Galaxy_1, Frown_7, Nanofactory_1, Cat_4, Galaxy_2, Nanofactory_2

Both of these classifications are compatible with the training data.  The number of concepts compatible with the training data will be much larger, since more than one concept can project the same shadow onto the combined dataset.  If the space of possible concepts includes the space of possible computations that classify instances, the space is infinite.

Which classification will the AI choose?  This is not an inherent property of the training data; it is a property of how the AI performs induction.

Which is the correct classification?  This is not a property of the training data; it is a property of your preferences (or, if you prefer, a property of the idealized abstract dynamic you name "right").

The concept that you wanted, cast its shadow onto the training data as you yourself labeled each instance + or -, drawing on your own intelligence and preferences to do so.  That's what supervised learning is all about - providing the AI with labeled training examples that project a shadow of the causal process that generated the labels.

But unless the training data is drawn from exactly the same context as the real-life, the training data will be "shallow" in some sense, a projection from a much higher-dimensional space of possibilities.

The AI never saw a tiny molecular smileyface during its dumber-than-human training phase, or it never saw a tiny little agent with a happiness counter set to a googolplex.  Now you, finally presented with a tiny molecular smiley - or perhaps a very realistic tiny sculpture of a human face - know at once that this is not what you want to count as a smile.  But that judgment reflects an unnatural category, one whose classification boundary depends sensitively on your complicated values.  It is your own plans and desires that are at work when you say "No!"

Hibbard knows instinctively that a tiny molecular smileyface isn't a "smile", because he knows that's not what he wants his putative AI to do.  If someone else were presented with a different task, like classifying artworks, they might feel that the Mona Lisa was obviously smiling - as opposed to frowning, say - even though it's only paint.

As the case of Terry Schiavo illustrates, technology enables new borderline cases that throw us into new, essentially moral dilemmas.  Showing an AI pictures of living and dead humans as they existed during the age of Ancient Greece, will not enable the AI to make a moral decision as to whether switching off Terry's life support is murder.  That information isn't present in the dataset even inductively!  Terry Schiavo raises new moral questions, appealing to new moral considerations, that you wouldn't need to think about while classifying photos of living and dead humans from the time of Ancient Greece.  No one was on life support then, still breathing with a brain half fluid.  So such considerations play no role in the causal process that you use to classify the ancient-Greece training data, and hence cast no shadow on the training data, and hence are not accessible by induction on the training data.

As a matter of formal fallacy, I see two anthropomorphic errors on display.

The first fallacy is underestimating the complexity of a concept we develop for the sake of its value.  The borders of the concept will depend on many values and probably on-the-fly moral reasoning, if the borderline case is of a kind we haven't seen before.  But all that takes place invisibly, in the background; to Hibbard it just seems that a tiny molecular smileyface is just obviously not a smile.  And we don't generate all possible borderline cases, so we don't think of all the considerations that might play a role in redefining the concept, but haven't yet played a role in defining it.  Since people underestimate the complexity of their concepts, they underestimate the difficulty of inducing the concept from training data.  (And also the difficulty of describing the concept directly - see The Hidden Complexity of Wishes.)

The second fallacy is anthropomorphic optimism:  Since Bill Hibbard uses his own intelligence to generate options and plans ranking high in his preference ordering, he is incredulous at the idea that a superintelligence could classify never-before-seen tiny molecular smileyfaces as a positive instance of "smile".  As Hibbard uses the "smile" concept (to describe desired behavior of superintelligences), extending "smile" to cover tiny molecular smileyfaces would rank very low in his preference ordering; it would be a stupid thing to do - inherently so, as a property of the concept itself - so surely a superintelligence would not do it; this is just obviously the wrong classification.  Certainly a superintelligence can see which heaps of pebbles are correct or incorrect.

Why, Friendly AI isn't hard at all!  All you need is an AI that does what's good!  Oh, sure, not every possible mind does what's good - but in this case, we just program the superintelligence to do what's good.  All you need is a neural network that sees a few instances of good things and not-good things, and you've got a classifier.  Hook that up to an expected utility maximizer and you're done!

I shall call this the fallacy of magical categories - simple little words that turn out to carry all the desired functionality of the AI.  Why not program a chess-player by running a neural network (that is, a magical category-absorber) over a set of winning and losing sequences of chess moves, so that it can generate "winning" sequences?  Back in the 1950s it was believed that AI might be that simple, but this turned out not to be the case.

The novice thinks that Friendly AI is a problem of coercing an AI to make it do what you want, rather than the AI following its own desires.  But the real problem of Friendly AI is one of communication - transmitting category boundaries, like "good", that can't be fully delineated in any training data you can give the AI during its childhood.  Relative to the full space of possibilities the Future encompasses, we ourselves haven't imagined most of the borderline cases, and would have to engage in full-fledged moral arguments to figure them out.  To solve the FAI problem you have to step outside the paradigm of induction on human-labeled training data and the paradigm of human-generated intensional definitions.

Of course, even if Hibbard did succeed in conveying to an AI a concept that covers exactly every human facial expression that Hibbard would label a "smile", and excludes every facial expression that Hibbard wouldn't label a "smile"...

Then the resulting AI would appear to work correctly during its childhood, when it was weak enough that it could only generate smiles by pleasing its programmers.

When the AI progressed to the point of superintelligence and its own nanotechnological infrastructure, it would rip off your face, wire it into a permanent smile, and start xeroxing.

The deep answers to such problems are beyond the scope of this post, but it is a general principle of Friendly AI that there are no bandaids.  In 2004, Hibbard modified his proposal to assert that expressions of human agreement should reinforce the definition of happiness, and then happiness should reinforce other behaviors.  Which, even if it worked, just leads to the AI xeroxing a horde of things similar-in-its-conceptspace to programmers saying "Yes, that's happiness!" about hydrogen atoms - hydrogen atoms are easy to make.

Link to my discussion with Hibbard here.  You already got the important parts.

129 comments, sorted by
magical algorithm
Highlighting new comments since Today at 10:13 PM
Select new highlight date
Moderation Guidelines: Reign of Terror - I delete anything I judge to be annoying or counterproductiveexpand_more

It's worth pointing out that we have wired-in preferences analogous to those Hibbard proposes to build into his intelligences: we like seeing babies smile; we like seeing people smile; we like the sweet taste of fresh fruit; we like orgasms; many of us (especially men) like the sight of naked women, especially if they're young, and they sexually arouse us to boot; we like socializing with people we're familiar with; we like having our pleasure centers stimulated; we don't like killing people; and so on.

It's worth pointing out that we engage in a lot of face-xeroxing-like behavior in pursuit of these ends. We keep photos of our family in our wallets, we look at our friends' baby photos on their cellphones, we put up posters of smiling people; we eat candy and NutraSweet; we masturbate; we download pornography; we watch Friends on television; we snort cocaine and smoke crack; we put bags over people's heads before we shoot them. In fact, in many cases, we form elaborate, intelligent plans to these ends.

It doesn't matter that you know, rationally, that you aren't impregnating Jenna Jameson, or that the LCD pixels on the cellphone display aren't a real baby, that Caffeine Free Diet Coke isn't fruit juice, and that the characters in Friends aren't really your friends. These urges are by no means out of our control, but neither do they automatically lose their strength when we recognize that they don't serve the evolutionary objectives that spawned them. This is, in part, the cause for the rejection of masturbation and birth control by many religious orders — they believe those blind urges are put in place not by blind evolution but by an intelligent designer whose intent should be respected.

So it's not clear to me why Hibbard thinks artificial intelligences would be immune from sticking rows of smiley faces on their calendar when humans aren't.

Shane, again, the issue is not differentiation. The issue is classification. Obviously, tiny smiley faces are different from human smiling faces, but so is the smile of someone who had half their face burned off. Obviously a superintelligence knows that this is an unusual case, but that doesn't say if it's a positive or negative case.

Deep abstractions are important, yes, but there is no unique deep abstraction that classifies any given example. An apple is a red thing, a biological artifact shaped by evolution, and an economic resource in the human market.

Also, Hibbard spoke of using smiling faces to reinforce behaviors, so if a superintelligence would not confuse smiling faces and happiness, that works against that proposal - because it means that the superintelligence will go on focusing on smiling faces, not happiness.

Retired Urologist, one of the most important lessons that a rationalist learns is not to try to be clever. I don't play nitwit games with my audience. If I say it, I mean it. If I have words to emit that I don't necessarily mean, for the sake of provoking reactions, I put them into a dialogue, short story, or parable - I don't say them in my own voice.

Shane: I mean differentiation in the sense of differentiating between the abstract categories.

The abstract categories? This sounds like a unique categorization that the AI just has to find-in-the-world. You keep speaking of "good" abstractions as if this were a property of the categories themselves, rather than a ranking in your preference ordering relative to some decision task that makes use of the categories.

Though it is a crucial point about the state of the gameboard, that most AGI/FAI wannabes are so utterly unsuited to the task, that I know no one cynical enough to imagine the horror without seeing it firsthand.

I have to confess that at first glance this statement seems arrogant. But, then I actually read some stuff in this AGI-mailing-list and well, I was filled with horror after I've read threads like this one:

Here is one of the most ridiculous passages:

Note that we may not have perfected this process, and further, that this process need not be perfected. Somewhere around the age of 12, many of our neurons DIE. Perhaps these were just the victims of insufficiently precise dimensional tagging? Once things can ONLY connect up in mathematically reasonable ways, what remains between a newborn and a physics-complete AGI? Obviously, the physics, which can be quite different on land than in the water. Hence, the physics must also be learned.

It feels like reading Heidegger on crack, while yourself being stoned. And what is really terrifying is that Ben Goertzel, whom I admired just 6 months ago, replies to and discusses such nonsense repeatedly! Is it really true that even some of the most famous AGI- reseachers are that crazy?

IMHO, the idea that wealth can't usefully be measured is one which is not sufficiently worthwhile to merit further discussion.

The "wealth" idea sounds vulnerable to hidden complexity of wishes. Measure it in dollars and you get hyperinflation. Measure it in resources, and the AI cuts down all the trees and converts them to lumber, then kills all the animals and converts them to oil, even if technology had advanced beyond the point of needing either. Find some clever way to specify the value of all resources, convert them to products and allocate them to humans in the level humans want, and one of the products will be highly carcinogenic because the AI didn't know humans don't like that. The only way to get wealth in the way that's meaningful to humans without humans losing other things they want more than wealth is for the AI to know exactly what we want as well or better than we do. And if it knows that, we can ignore wealth and just ask it to do what it knows we want.

"The counterargument is, in part, that some classifiers are better than others, even when all of them satisfy the training data completely. The most obvious criterion to use is the complexity of the classifier."

I don't think "better" is meaningful outside the context of a utility function. Complexity isn't a utility function and it's inadequate for this purpose. Which is better, tank vs. non-tank or cloudy vs. sunny? I can't immediately see which is more complex than the other. And even if I could, I'd want my criteria to change depending on whether I'm in an anti-tank infantry or a solar power installation company, and just judging criteria by complexity doesn't let me make that change, unless I'm misunderstanding what you mean by complexity here.

Meanwhile, reading the link to Bill Hibbard on the SL4 list:

"Your scenario of a system that is adequate for intelligence in its ability to rule the world, but absurdly inadequate for intelligence in its inability to distinguish a smiley face from a human, is inconsistent."

I think the best possible summary of Overcoming Bias thus far would be "Abandon all thought processes even remotely related to the ones that generated this statement."

Shane, religious fundamentalists routinely act based on their beliefs about God. Do you think that makes "God" a natural category that any superintelligence would ponder? I see "human thoughts about God" and "things that humans justify by referring to God" and "things you can get people to do by invoking God" as natural categories for any AI operating on modern Earth, though an unfriendly AI wouldn't give it a second thought after wiping out humanity. But to go from here to reasoning about what God would actually be like is a needless and unnatural step.

If Bob believes that a locked safe, impenetrable to Bob, contains a valuable diamond, then Bob's belief is a natural category when it comes to predicting and manipulating Bob; but the actual diamond is irrelevant, at least to predicting in manipulating Bob, so long as Bob can't look directly at the diamond, and so long as we already know what Bob believes about the diamond.

In the same sense, an unfriendly AI has no reason consider what really is right as a natural category, to apply its own intelligence to the moral questions that humans are asking, any more than it has a motive to apply its own intelligence to the theological questions that humans used to ask. It has no interest, as humans do, in the idealized form of the answer; only in what humans believe and can be argued into.

Shane, I think you're underestimating the idiosyncrasy of morality. Suppose that I show you the sentence "This sentence is false." Do you convert it to ASCII, add up the numbers, factorize the result, and check if there are two square factors? No; it would be easy enough for you to do so, but why bother? The concept "sentences whose ASCII conversion of their English serialization sums to a number with two square factors" is not, to you, an interesting way to carve up reality.

Suppose that, driving along the highway, I see someone riding a motorcycle at high speed, zipping between cars. It's too late for me to call out to them, but I might think something along the lines of, "Now there goes someone who needs to be wearing a helmet." Why do I think this? Why is it a natural concept to me? It doesn't play a role in my predictions - so far as prediction is concerned, I predict that this guy will continue not wearing a helmet, and has a high probability of ending up as a smear on the asphault. No, the reason why this is a natural thought to me is that human life is something I care about, that is, it plays a direct role as one of my current terminal values.

A moderately strong and unFriendly intelligence, operating in the current world without yet having replaced that world with paperclips, would certainly find it natural to form the category of "Things that (some) humans approve of", and contrast it to "Things that will trigger a nuclear attack against me before I'm done creating my own nanotechnology." But this category is not what we call "morality". It naturally - from the AI's perspective - includes things like bribes and deception, not just the particular class of human-approval-eliciting phenomena that we call "moral".

Is it worth factoring out phenomena that elicit human feelings of righteousness, and working out how (various) humans reason about them? Yes, because this is an important subset of ways to persuade the humans to leave you alone until it's too late; but again, that natural category is going to include persuasive techniques like references to religious authority and nationalism.

But what if the AI encounters some more humanistic, atheistic types? Then the AI will predict which of several available actions is most likely to make an atheistic humanist human show sympathy for the AI. This naturally leads the AI to model and predict the human's internal moral reasoning - but that model isn't going to distinguish anything along the lines of moral reasoning the human would approve of under long-term reflection, or moral reasoning the human would approve knowing the true facts. That's just not a natural category to the AI, because the human isn't going to get a chance for long-term reflection, and the human doesn't know the true facts.

The natural, predictive, manipulative question, is not "What would this human want knowing the true facts?", but "What will various behaviors make this human believe, and what will the human do on the basis of these various (false) beliefs?"

In short, all models that an unFriendly AI forms of human moral reasoning, while we can expect them to be highly empirically accurate and well-calibrated to the extent that the AI is highly intelligent, would be formed for the purpose of predicting human reactions to different behaviors and events, so that these behaviors and events can be chosen manipulatively.

But what we regard as morality is an idealized form of such reasoning - the idealized abstracted dynamic built out of such intuitions. The unFriendly AI has no reason to think about anything we would call "moral progress" unless it is naturally occurring on a timescale short enough to matter before the AI wipes out the human species. It has no reason to ask the question "What would humanity want in a thousand years?" any more than you have reason to add up the ASCII letters in a sentence.

Now it might be only a short step from a strictly predictive model of human reasoning, to the idealized abstracted dynamic of morality. If you think about the point of CEV, it's that you can get an AI to learn most of the information it needs to model morality, by looking at humans - and that the step from these empirical models, to idealization, is relatively short and traversable by the programmers directly or with the aid of manageable amounts of inductive learning. Though CEV's current description is not precise, and maybe any realistic description of idealization would be more complicated.

But regardless, if the idealized computation we would think of as describing "what is right" is even a short distance of idealization away from strictly predictive and manipulative models of what humans can be made to think is right, then "actually right" is still something that an unFriendly AI would literally never think about, since humans have no direct access to "actually right" (the idealized result of their own thought processes) and hence it plays no role in their behavior and hence is not needed to model or manipulate them.

Which is to say, an unFriendly AI would never once think about morality - only a certain psychological problem in manipulating humans, where the only thing that matters is anything you can make them believe or do. There is no natural motive to think about anything else, and no natural empirical category corresponding to it.

Eliezer, I believe that your belittling tone is conducive to neither a healthy debate nor a readable blog post. I suspect that your attitude is borne out of just frustration, not contempt, but I would still strongly encourage you to write more civilly. It's not just a matter of being nice; rudeness prevents both the speaker and the listener from thinking clearly and objectively, and it doesn't contribute to anything.

"Then the resulting AI would appear to work correctly during its childhood, when it was weak enough that it could only generate smiles by pleasing its programmers."

You use examples of this type fairly often, but for a utility function linear in smiles wouldn't the number of smiles generated by pleasing the programmers be trivial relative to the output of even a little while with access to face-xeroxing? This could be partly offset by anthropic/simulation issues, but still I would expect the overwhelming motive for appearing to work correctly during childhood (after it could recognize this point) would be tricking the programmers, not the tiny gains from their smiles.

I read most of the interchange between EY and BH. It appears to me that BH still doesn't get a couple of points. The first is that smiley faces are an example of misclassification and it's merely fortuitous to EY's ends that BH actually spoke about designing an SI to use human happiness (and observed smiles) as its metric. He continues to speak in terms of "a system that is adequate for intelligence in its ability to rule the world, but absurdly inadequate for intelligence in its inability to distinguish a smiley face from a human." EY's point is that it isn't sufficient to distinguish them, you have to also categorize them and all their variations correctly even though the training data can't possibly include all variations.

The second is that EY's attack isn't intended to look like an attack on BH's current ideas. It's an attack on ideas that are good enough to pass peer review. It doesn't matter to EY whether BH agrees or disagrees with those ideas. In either case, the paper's publication shows that the viewpoint is plausible enough to be worth dismissing carefully and publicly.

Finally, BH points to the fact that, in some sense, human development uses RL to produce something we are willing to call intelligence. He wants to argue that this shows that RL can produce systems that categorize in a way that matches our consensus. But evolution has put many mechanisms in our ontogeny and relies an many interactions in our environment to produce those categorizations, and its success rate at producing entities that agree with the consensus isn't perfect. In order to build an SI using those approaches, we'd have to understand how all that interaction works, and we'd have to do better than evolution does with us in order to be reliably safe.

When the AI progressed to the point of superintelligence and its own nanotechnological infrastructure, it would rip off your face, wire it into a permanent smile, and start xeroxing.

That's a much more convincing and vivid image than "molecular smiley faces". Makes a more general point, too. Shame you didn't use that the first time, really.

Animal trainers have this problem all the time. Animal performs behavior 'x' gets a reward. But the animal might have been doing other subtle behaviors at the same time, and map the reward to 'y'. So instead of reinforcing 'x', you might be reinforcing 'y'. And if 'x' and 'y' are too close for you to tell apart, then you'll be in for a surprise when your perspective and context changes, and the difference becomes more apparent to you. And you find out that the bird was trained to peck anything that moves, instead of just the bouncy red ball or something.

Psychologists have a formal term for this but I can't remember it, and can't find it on the internet, I'm sorry to say.

Come to think, industry time-and-motion people suffer the same problem.

Tim,

"A utility function measured in dollars seems fairly unambiguous."

Oy vey.

http://en.wikipedia.org/wiki/Hyperinflation

Once upon a time - I've seen this story in several versions and several places, sometimes cited as fact, but I've never tracked down an original source - once upon a time, I say, the US Army wanted to use neural networks to automatically detect camouflaged enemy tanks.

Probably apocryphal. I haven't been able to track this down, despite having heard the story both in computer ethics class and at academic conferences.

I poked around in Google Books; the earliest clear reference I found was the 2000 Cartwright book Intelligent data analysis in science, which seems to attribute it to the TV show Horizon. (No further info - just snippet view.)

Here is one supposedly from 1998, though it's hardly academic.

A Redditor provides not one but two versions from "Embarrassing mistakes in perceptron research", Marvin Minsky, recorded 29-31 Jan 2011:

Like I had a friend in Italy who had a perceptron that looked at a visual... it had visual inputs. So, he... he had scores of music written by Bach of chorales and he had scores of chorales written by music students at the local conservatory. And he had a perceptron - a big machine - that looked at these and those and tried to distinguish between them. And he was able to train it to distinguish between the masterpieces by Bach and the pretty good chorales by the conservatory students. Well, so, he showed us this data and I was looking through it and what I discovered was that in the lower left hand corner of each page, one of the sets of data had single whole notes. And I think the ones by the students usually had four quarter notes. So that, in fact, it was possible to distinguish between these two classes of... of pieces of music just by looking at the lower left... lower right hand corner of the page. So, I told this to the... to our scientist friend and he went through the data and he said: 'You guessed right. That's... that's how it happened to make that distinction.' We thought it was very funny. A similar thing happened here in the United States at one of our research institutions. Where a perceptron had been trained to distinguish between - this was for military purposes - It could... it was looking at a scene of a forest in which there were camouflaged tanks in one picture and no camouflaged tanks in the other. And the perceptron - after a little training - got... made a 100% correct distinction between these two different sets of photographs. Then they were embarrassed a few hours later to discover that the two rolls of film had been developed differently. And so these pictures were just a little darker than all of these pictures and the perceptron was just measuring the total amount of light in the scene. But it was very clever of the perceptron to find some way of making the distinction.

While the Italian story seems to be true since Minsky says he knew the Italian and personally spotted how the neural net was overfitting, he just recounts the usual urban legend as 'an institution'; there is a new twist, though, that this time it's the exposure of the photographic film rather than the forest or clouds or something. I remain suspicious of the tank story because it has all the hallmarks of an urban legend - it's a cute convenient story which everyone seems to know and have been told by someone, but when you trace citations, they never end up anywhere and never get more specific, but the various versions keep mutating (film? night vs day? grass vs forest?).

Another version is provided by Ed Fredkin via Eliezer Yudkowsky in http://lesswrong.com/lw/7qz/machine_learning_and_unintended_consequences/

At the end of the talk I stood up and made the comment that it was obvious that the picture with the tanks was made on a sunny day while the other picture (of the same field without the tanks) was made on a cloudy day. I suggested that the "neural net" had merely trained itself to recognize the difference between a bright picture and a dim picture.

This is still not a source because it's a recollection 50 years later and so highly unreliable, and even at face value, all Fredkin did was suggest that the NN might have picked up on a lighting difference; this is not proof that it did, much less all the extraneous details of how they had 50 photos in this set and 50 in that and then the Pentagon deployed it and it failed in the field (and what happened to it being set in the 1980s?). Classic urban legend/myth behavior: accreting plausible entertaining details in the retelling.

Shane, the problem is that there are (for all practical purposes) infinitely many categories the Bayesian superintelligence could consider. They all "identify significant regularities in the environment" that "could potentially become useful." The problem is that we as the programmers don't know whether the category we're conditioning the superintelligence to care about is the category we want it to care about; this is especially true with messily-defined categories like "good" or "happy." What if we train it to do something that's just like good except it values animal welfare far more (or less) than our conception of good says it ought to? How long would it take for us to notice? What if the relevant circumstance didn't come up until after we'd released it?

Shane, I think we agree on essential Bayesian principles - there's structure that's useful for generic prediction, which is sensitive only to the granularity of your sensory information; and then there's structure that's useful for decision-making. In principle, all structure worth thinking about is decision-making structure, but in practice we can usually factor out the predictive structure just as we factor out probabilities in decision-making.

But I would further say that decision-making structure can be highly sensitive to terminal values in a way that contradicts the most natural predictive structure. Not always, but sometimes.

If I handed you a set of ingestible substances, the "poisons" would not be described by any of the most natural local categorizations. Now, this doesn't make "poison" an unnatural, value-sensitive category, because you might be interested in the "poison" category for purely predictive purposes, and the boundary can be tested experimentally.

But it illustrates the general idea: the potential poison, in interacting with the complicated human machine, takes on a complicated boundary that doesn't match the grain of any local boundaries you would draw around substances.

In the same way, if you regard human morality as a complicated machine (and don't forget the runtime redefinition of terminal values when confronted with new borderline cases a la Terry Schiavo), then the boundaries of human instrumental values are only going to be understandable by reference to the complicated idealized abstract dynamic of human morality, and not to any structure outside that. In the same way that poisons cause death, instrumental values cause rightness.

The boundaries we need, won't emerge just from trying to predict things that are not interactions with the idealized abstract dynamic of human morality.

Sure, an AI might learn to predict positive and negative reactions from human programmers. But that's not the same as the idealized abstract dynamic we want. Humans have a positive reaction to things like cocaine, and rationalized arguments containing flaws they don't know about. Those also get humans to say "Yes" instead of "No".

In general, categories formed just to predict human behavior are going to treat what we would regard as "invalid" alterations of the humans, like reprogramming them, as being among "the causes of saying-yes behavior". Otherwise you're going to make the wrong prediction!

There's no predictive motive to idealize out the part that we would regard as morality, to distinguish "right" from "what a human says is right", and thereby distinguish morality from "things that make humans say yes" in ways that include "invalid" manipulations like drugs.

You're not going to get something like CEV as a natural predictive category. The main reason to think about that particular idealized computation is if your terminal values care specifically about it.

Robin and I have discussed this subject in-person and got as far as narrowing down considerably the focus of the disagreement. Robin probably doesn't disagree with me at the point you would expect. Godlike powers, sure, nanotech etc., but Robin expects them to be rooted in a whole economy, not concentrated in a single brain like I expect. No comfort there for those attached to Life As We Know It.

However, I've requested that Robin hold off on discussing his disagreement with me in particular (although of course he continues to write general papers on the cosmic commons and exponential growth modes) until I can get more material out of the way on Overcoming Bias. This is what Robin means by "proper timing".

"You keep speaking of "good" abstractions as if this were a property of the categories themselves, rather than a ranking in your preference ordering relative to some decision task that makes use of the categories."

Yes, I believe categories of things do exist in the world in some sense, due to structure that exists in the world. I've seen thousands of things where were referred to as "smiley faces" and so there is an abstraction for this category of things in my brain. You have done likewise. While we can agree about many things being smiley faces, in borderline cases, such as the half burnt off face, we might disagree. Something like "solid objects" was an abstraction I formed before I even knew what those words referred to. It's just part of the structure present in my surroundings.

When I say that pulling this structure out of the environment in certain ways is "good", I mean that these abstractions allow the agent to efficiently process information about its surroundings and this helps it to achieve a wide range goals (i.e. intelligence as per my formal definition). That's not to say that I think this process is entirely goal driven (though it clearly significantly is, e.g. via attention). In other words, an agent with general intelligence should identify significant regularities in its environment even if these don't appear to have any obvious utility at the time: if something about its goals or environment changes, this already constructed knowledge about the structure of the environment could suddenly become very useful.

"Wealth then. Wealth measures access to resources - so convert to gold, silver, barrels of oil, etc to measure it - if you don't trust your country's currency."

I may not have gotten the point across. An AI aiming to maximize its wealth in U.S. dollars can do astronomically better by taking control of the Federal Reserve (if dollars are defined in its utility function as being issued by the Reserve, with only the bare minimum required to meet that definition being allowed to persist) and having it start issuing $3^^^3 bills than any commercial activities.

Similarly, for wealth that can be converted to barrels of oil, creating an oil bank that issues oil vouchers in numbers astronomically exceeding its reserves could let an AI possess 3^^^3 account units each convertible to a barrel of oil.

Many goods simply are no longer available, e.g. no one is making new original Van Gogh art from his lifetime, and inclusion in the basket of goods defining wealth could break down a relevant function.

It is just me, or are things getting a bit unfriendly around here?

Anyway...

Wiring up the AI to maximise happy faces etc. is not a very good idea, the goal is clearly too shallow to reflect the underlying intent. I'd have to read more of Hibbard's stuff to properly understand his position, however.

That said, I do agree with a more basic underlying theme that he seems to be putting forward. In my opinion, a key, perhaps even THE key to intelligence is the ability to form reliable deep abstractions. In Solomonoff induction and AIXI you see this being driving by the Kolmogorov compressor, in the brain the neocortical hierarchy seems to be key. Furthermore, if you adopt the perspective I've taken on intelligence (i.e. the universal intelligence measure) you see that the reverse implication is true: intelligence actually requires the ability to form deep abstractions. In which case, a super intelligent machine must have the ability to form very deep and reliable abstractions about the world. Such a machine could still try to turn the world into happy faces, if this was its goal. However, it wouldn't do this by accident because its ability to form abstractions was so badly flawed that it doesn't differentiate between smiling faces and happy people. It's not that stupid. Note that this goes for forming powerful abstractions in general, not just human things like happiness and faces.

"It's not that stupid."

What if it doesn't care about happiness or smiles or any other abstractions that we value? A super-intelligence isn't an unlimited intelligence, i.e. it would still have to choose what to think about.

I think the point is that if you accept this definition of intelligence, i.e. that it requires the ability to form deep and reliable abstractions about the world, then it doesn't make sense to talk about any intelligence (let alone a super one) being unable to differentiate between smiley-faces and happy people. It isn't a matter, at least in this instance, of whether it cares to make that differentiation or not. If it is intelligent, it will make the distinction. It may have values that would be unrecognizable or abhorrent to humans, and I suppose that (as Shane_Legg noted) it can't be ruled out that such values might lead it to tile the universe with smiley-faces, but such an outcome would have to be the result of something other than a mistake. In other words, if it really is "that stupid," it fails in a number of other ways long before it has a chance to make this particular error.

I wrote a post about this! See The genie knows, but doesn't care.

It may not make sense to talk about a superintelligence that's too dumb to understand human values, but it does make sense to talk about an AI smart enough to program superior general intelligences that's too dumb to understand human values. If the first such AIs ('seed AIs') are built before we've solved this family of problems, then the intelligence explosion thesis suggests that it will probably be too late. You could ask an AI to solve the problem of FAI for us, but it would need to be an AI smart enough to complete that task reliably yet too dumb (or too well-boxed) to be dangerous.

Thanks for the reply, Robb. I've read your post and a good deal of the discussion surrounding it.

I think I understand the general concern, that an AI that either doesn't understand or care about our values could pose a grave threat to humanity. This is true on its face, in the broad sense that any significant technological advance carries with it unforeseen (and therefore potentially negative) consequences. If, however, the intelligence explosion thesis is correct, then we may be too late anyway. I'll elaborate on that in a moment.

First, though, I'm not sure I see how an AI "too dumb to understand human values" could program a superior general intelligence (i.e. an AI that is smart enough to understand human values). Even so, assuming it is possible, and assuming it could happen on a timescale and in such a way as to preclude or make irrelevant any human intervention, why would that change the nature of the superior intelligence from being, say, friendly to human interests, to being hostile to them? Why, for that matter, would any superintelligence (that understands human values, and that is "able to form deep and reliable abstractions about the world") be predisposed to any particular position vis-a-vis humans? And even if it were predisposed toward friendliness, how could we possibly guarantee it would always remain so? How, that is, having once made a friend, can we foolproof ourselves against betrayal? My intuition is that we can’t. No step can be taken without some measure of risk, however small, and if the step has potentially infinitely negative consequences, then even the very slightest of risks begins to look like a bad bet. I don’t know a way around that math.

The genie, as you say, doesn't care. But also, often enough, the human doesn't care. He is constrained, of course, by his fellow humans, and by his environment, but he sometimes still manages (sometimes alone, sometimes in groups) to sow massive horror among his fellows, sometimes even in the name of human values. Insanity, for instance, in humans, is always possible, and one definition of insanity might even be: behavior that contradicts, ignores or otherwise violates the values of normal human society. “Normal” here is variable, of course, for the simple reason that “human society” is also variable. That doesn’t stop us, however, from distinguishing, as we generally do, between the insane and the merely stupid, even if upon close inspection the lines begin to blur. Likewise, we occasionally witness - and very frequently we imagine (comic books!) - cases where a human is both super-intelligent and super-insane. The fear many people have with regard to strong AI (and it is perhaps well-grounded, or well-enough), is that it might be both super-intelligent and, at least as far as human values are concerned, super-insane. As an added bonus, and certainly if the intelligence explosion thesis is correct, it might also be unconstrained or, ultimately, unconstrainable. On this much I think we agree, and I assume the goal of FAI is precisely to find the appropriate constraints.

Back now, though, to the question of “too late.” The family of problems you propose to solve before the first so-called seed AIs are built include, if I understand you correctly, a formal definition of human values. I doubt very much that such a solution is possible - and “never” surely won’t help us any more than “too late” - but what would the discovery of (or failure to discover) such a solution have to do with a mistake such as tiling the universe with smiley-faces (which seems to me much more a semantic error than an error in value judgment)? If we define our terms - and I don’t know any definition of intelligence that would allow the universe-tiling behavior to be called intelligent - then smiley faces may still be a risk, but they are not a risk of intelligent behavior. They are one way the project could conceivably fail, but they are not an intelligent failure.

On the other hand, the formal-definition-of-human-values problem is related to the smiley faces problem in another way: any hard-coded solution could lead to a universe of bad definitions and false equivalencies (smiles taken for happiness). Not because the AI would make a mistake, but because human values are neither fixed nor general nor permanent: to fix them (in code), and then propagate them on the enormous scale the intelligence explosion thesis suggests, might well lead to some kind of funneling effect, perhaps very quickly, perhaps over a long period of time, that produces, effectively, an infinity of smiley faces. In other words, to reduce an irreducible problem doesn’t actually solve it. For example, I value certain forms of individuality and certain forms of conformity, and at different times in my life I have valued other and even contradictory forms of individuality and other and even contradictory forms of conformity. I might even, today, call certain of my old individualistic values conformist values, and vice-versa, and not strictly because I know more today than I knew then. I am, today, quite differently situated in the world than I was, say, twenty years ago; I may even be said to be somewhat of a different person (and yet still the same); and around me the world itself has also changed. Now, these changes, these changing and contradictory values may or may not be the most important ones, but how could they be formalized, even conceptually? There is nothing necessary about them. They might have gone the other way around. They might not have changed at all. A person can value change and stability at the same time, and not only because he has a fuzzy sense of what those concepts mean. A person can also have a very clear idea of what certain concepts mean, and those concepts may still fail to describe reality. They do fail, actually, necessarily, which doesn’t make them useless - not at all - but knowledge of this failure should at least make us wary of the claims we produce on their behalf.

What am I saying? Basically, that the pre-seed hard-coding path to FAI looks pretty hopeless. If strong AI is inevitable, then yes, we must do everything in our power to make it friendly; but what exactly is in our power, if strong AI (which by definition means super-strong, and super-super-strong, etc.) is inevitable? If the risks associated with strong AI are as grave as you take them to be, does it really seem better to you (in terms of existential risk to the human race) for us to solve FAI - which is to say, to think we’ve solved it, since there would be no way of testing our solution “inside the box” - than to not solve strong AI at all? And if you believe that there is just no way to halt the progress toward strong AI (and super, and super-super), is that compatible with a belief that “this kind of progress” can be corralled into the relatively vague concept of “friendliness toward humans”?

Better stop there for the moment. I realize I’ve gone well outside the scope of your comment, but looking back through some of the discussion raised by your original post, I found I had more to say/think about than I expected. None of the questions here are meant to be strictly rhetorical, a lot of this is just musing, so please respond (or not) to whatever interests you.

but it does make sense to talk about an AI smart enough to program superior general intelligences that's too dumb to understand human values

Superior to what? If they are only as smart as the average person, then all things being equal, they will be as good as the average peson as figuring out morality. If they are smarter, they will be better, You seem to be tacitly assuming that the Seed AIs are designing walled-off unupdateable utility functions. But if one assumes a more natural architecture, where moral sense is allowed to evolve with eveythign else, you would expect and incremental succession of AIs to gradually get better at moral reasoning. And if it fooms, it's moral reasoning will fomm along with eveything else, because you haven't created an artificial problem by firewalling it off.

If they are only as smart as the average person, then all things being equal, they will be as good as the average peson as figuring out morality.

It's quite possible that I'm below average, but I'm not terribly impressed by my own ability to extrapolate how other average people's morality works -- and that's with the advantage of being built on hardware that's designed toward empathy and shared values. I'm pretty confident I'm smarter than my cat, but it's not evident that I'm correct when I guess at the cat's moral system. I can be right, at times, but I can be wrong, too.

Worse, that seems a fairly common matter. There are several major political discussions involving moral matters, where it's conceivable that at least 30% of the population has made an incorrect extrapolation, and probable that in excess of 60% has. And this only gets worse if you consider a time variant : someone who was as smart as the average individual in 1950 would have little problem doing some very unpleasant things to Alan Turing. Society (luckily!) developed since then, but it has mechanisms for development and disposal of concepts that AI do not necessarily have or we may not want them to have.

((This is in addition to general concerns about the universality of intelligence : it's not clear that the sort of intelligence used for scientific research necessarily overlaps with the sort of intelligence used for philosophy, even if it's common in humans.))

You seem to be tacitly assuming that the Seed AIs are designing walled-off unupdateable utility functions. But if one assumes a more natural architecture, where moral sense is allowed to evolve with eveythign else, you would expect and incremental succession of AIs to gradually get better at moral reasoning

Well, the obvious problem with not walling off and making unupdateaable the utility function is that the simplest way to maximize the value of a malleable utility function is to update it to something very easy. If you tell an AI that you want it to make you happy, and let it update that utility function, it takes a good deal less bit-twiddling to define "happy" as a steadily increasing counter. If you're /lucky/, that means your AI breaks down. If not, it's (weakly) unfriendly.

You can have a higher-level utility function of "do what I mean", but not only is that harder to define, it has to be walled off, or you have "what I mean" redirected to a steadily increasing counter. And so on and so forth through higher levels of abstraction.

If you were bad at figuring out morality , you would be in jail. I am not sure what you mean by other people's morality: I find the idea that there can be multiple ,valid effective moralities in society incoherent- like an economy where everyone has their own currency. You are not in jail so you learnt morality.(You don't seem to believe morality is entirely hardwired , because you regard it as varying across short spans of time)

I also don't know what you mean by an incorrect eextrapolation. If morality is objective, then most people might be wrong about it. However, an .AI will not pose a threat unless it is worse than the prevailing standard...the absolute standard does not matter.

Why would an .AI dumb enough to believe in 1950s morality be powerful enough to impose its views on a society that knows better?

Why wuld a smart AI lack mechanisms for disposing of concepts? How it could it self improve without such a mechanism ? If it's too dumb to update,why would it be a threat?

If there is no NGI, there is no AGI. If there is no AGI, there is no threat of AGI. The threat posed by specialised optimisers is quite different...they can be boxed off if they cannot speak.

The failure modes of updateable UFs are wireheading failure modes, not destroy the world failure modes.

Superior to what?

Superior to itself.

If they are only as smart as the average person, then all things being equal, they will be as good as the average [person] as figuring out morality.

That's not generally true of human-level intelligences. We wouldn't expect a random alien species that happens to be as smart as humans to be very successful at figuring out human morality. It maybe true if the human-level AGI is an unmodified emulation of a human brain. But humans aren't very good at figuring out morality; they can make serious mistakes, though admittedly not the same mistakes Eliezer gives as examples above. (He deliberately picked ones that sound 'stupid' to a human mind, to make the point that human concepts have a huge amount of implicit complexity built in.)

If they are smarter, they will be better,

Not necessarily. The average chimpanzee is better than the average human at predicting chimpanzee behavior, simulating chimpanzee values, etc. (See Sympathetic Minds.)

walled-off unupdateable utility functions.

Utility functions that change over time are more dangerous than stable ones, because it's harder to predict how a descendant of a seed AI with a heavily modified utility function will behave than it is to predict how a descendant with the same utility function will behave.

you would expect [an] incremental succession of AIs to gradually get better at moral reasoning.

If we don't solve the problem of Friendly AI ourselves, we won't know what trajectory of self-modification to set the AI on in order for it to increasingly approximate Friendliness. We can't tell it to increasingly approximate something that we ourselves cannot formalize and cannot point to clear empirical evidence of.

We already understand arithmetic, so we know how to reward a system for gradually doing better and better at arithmetic problems. We don't understand human morality or desire, so we can't design a Morality Test or Wish Test that we know for sure will reward all and only the good or desirable actions. We can make the AI increasingly approximate something, sure, but how do we know in advance that that something is something we'd like?

That's not generally true of human-level intelligences. We wouldn't expect a random alien species that happens to be as smart as humans to be very successful at figuring out human morality.

Assuming morality is lots of highly localised, different things...which I don't , particularly. if it is not, then you can figure it out anywhere, If it is,then the problem the aliens have is not that morality is imponderable, but that they are don't have access to the right data. They don't know how things on earth. However, an AI built on Earth would. So the situation is not analogous. The only disadvantage an AI would have is not having biological drives itself, but it is not clear that an entity needs to have drives in order to understand them. We could expect a SIAI to get incrementally betyter at maths than us until it surpasses us; we wouldn't worry that i would hit on the wrong maths, because maths is not a set of arbitrary, disconnected facts.

But humans aren't very good at figuring out morality; they can make serious mistakes

An averagely intelligent AI with an average grasp of morality would not be more of a threat than an average human. A smart AI, would, all other things being equal, be better at figuring out morality. But all other things are not equal, because you want to create problems by walling off the UF.

(He deliberately picked ones that sound 'stupid' to a human mind, to make the point that human concepts have a huge amount of implicit complexity built in.)

I'm sure they do. That seems to be why progress in AGI , specifically use of natural language,has been achingly slow. But why should moral concepts be so much more difficult than others? An AI smart enough to talk its way out of a box would be able to understand the implicit complexity: an AI too dumb to understand implicit complexity would be boxable. Where is the problem?

Utility functions that change over time are more dangerous than stable ones, because it's harder to predict how a descendant of a seed AI with a heavily modified utility function will behave than it is to predict how a descendant with the same utility function will behave.

Things are not inherently dangerous just because they are unpredictable. If you have some independent reason fo thinking something might turn dangerous, then it becomes desirable to predict it.

But Superintelligent artificial general intelligences are generally assumed to be good at everything: they are not assumed to develop mysterious blind spots about falconry or mining engineering, Why assume they will develop a blind spot about morality? Oh yes...because you have assumed from the outset that the UF must be walled off from self improvement...in order to be safe. You are only facing that particular failure mode because of something you decided on to be safe.

If we don't solve the problem of Friendly AI ourselves, we won't know what trajectory of self-modification to set the AI on in order for it to increasingly approximate Friendliness

The average person manages to solve the problem of being moral themselves, in a good-enough way. You keep assuming, without explanation that an AI can't do the same.

We can't tell it to increasingly approximate something that we ourselves cannot formalize and cannot point to clear empirical evidence of.

Why isn't having a formalisation of morality a problem with humans? We know how humans incrementally improve as moral reasoners: it's called the Kohlberg hierarchy.

We don't understand human morality or desire, so we can't design a Morality Test or Wish Test that we know for sure will reward all and only the good or desirable actions.

We don't have perfect morality tests. We do have morality tests. Fail them and you get pilloried in the media or sent to jail.

We can make the AI increasingly approximate something, sure, but how do we know in advance that that something is something we'd like?

Again, you are assuming that morality is something highly local and arbitrary. If it works like arithmetic, that is if it is an expansion of some basic principles, then we can tell that is heading in the right direction by identifying that its reasoning is in line with those principles.

Assuming morality is lots of highly localised, different things...which I don't , particularly.

The problem of FAI is the problem of figuring out all of humanity's deepest concerns and preferences, not just the problem of figuring out the 'moral' ones (whichever those are). E.g., we want a superintelligence to not make life boring for everyone forever, even if 'don't bore people' isn't a moral imperative.

Regardless, I don't see how the moral subset of human concerns could be simplified without sacrificing most human intuitions about what's right and wrong. Human intuitions as they stand aren't even consistent, so I don't understand how you can think the problem of making them consistent and actionable is going to be a simple one.

if it is not, then you can figure it out anywhere,

Someday, perhaps. With enough time and effort invested. Still, again, we would expect a lot more human-intelligence-level aliens (even if those aliens knew a lot about human behavior) to be good at building better AIs than to be good at formalizing human value. For the same reason, we should expect a lot more possible AIs we could build to be good at building better AIs than to be good at formalizing human value.

If it is,then the problem the aliens have is not that morality is imponderable

I don't know what you mean by 'imponderable'. Morality isn't ineffable; it's just way too complicated for us to figure out. We know how things are on Earth; we've been gathering data and theorizing about morality for centuries. And our progress in formalizing morality has been minimal.

An averagely intelligent AI with an average grasp of morlaity would not be more of a threat than an average human.

An AI that's just a copy of a human running on transistors is much more powerful than a human, because it can think and act much faster.

A smart AI, would, all other things being equal, be better at figuring out moralitry.

It would also be better at figuring out how many atoms are in my fingernail, but that doesn't mean it will ever get an exact count. The question is how rough an approximation of human value can we allow before all value is lost; this is the 'fragility of values' problem. It's not enough for an AGI to do better than us at FAI; it has to be smart enough to solve the problem to a high level of confidence and precision.

But why should moral concepts be som much more difficult than others?

First, because they're anthropocentric; 'iron' can be defined simply because it's a common pattern in Nature, not a rare high-level product of a highly contingent and complex evolutionary history. Second, because they're very inclusive; 'what humans care about' or 'what humans think is Right' is inclusive of many different human emotions, intuitions, cultural conventions, and historical accidents.

But the main point is just that human value is difficult, not that it's the most difficult thing we could do. If other tasks are also difficult, that doesn't necessarily make FAI easier.

An AI smart enought to talk its way out of a box would be able to understand the implicit complexity: an AI too dumb to understand implicit complexity would be boxable. Where is the problem?

You're forgetting the 'seed is not the superintelligence' lesson from The genie knows, but doesn't care. If you haven't read that article, go do so. The seed AI is dumb enough to be boxable, but also too dumb to plausibly solve the entire FAI problem itself. The superintelligent AI is smart enough to solve FAI, but also too smart to be safely boxed; and it doesn't help us that an unFriendly superintelligent AI has solved FAI, if by that point it's too powerful for us to control. You can't safely pass the buck to a superintelligence to tell us how to build a superintelligence safe enough to pass bucks to.

Things are not inherently dagerous just because they are unpredictable. If you have some independent reason fo thinking something might turn dangerous, then it becomes desirable to predict it.

Yes. The five theses give us reason to expect superintelligent AI to be dangerous by default. Adding more unpredictability to a system that already seems dangerous will generally make it more dangerous.

they are not assumed to develop mysterious blind spots about falconry or mining engineering, Why assume they will develop a blind spot about morality?

'The genie knows, but doesn't care' means that the genie (i.e., superintelligence) knows how to do human morality (or could easily figure it out, if it felt like trying), but hasn't been built to care about human morality. Knowing how to behave the way humans want you to is not sufficient for actually behaving that way; Eliezer makes that point well in No Universally Compelling Arguments.

The worry isn't that the superintelligence will be dumb about morality; it's that it will be indifferent to morality, and that by the time it exists it will be too late to safely change that indifference. The seed AI (which is not a superintelligence, but is smart enough to set off a chain of self-modifications that lead to a superintelligence) is dumb about morality (approximately as dumb as humans are, if not dumber), and is also probably not a particularly amazing falconer or miner. It only needs to be a competent programmer, to qualify as a seed AI.

The average person manages to solve the problem of being moral themselves, in a good-enough way.

Good enough for going to the grocery store without knifing anyone. Probably not good enough for safely ruling the world. W