I apologize for the late response, but here goes :)

I think you missed the point I was trying to make.

You and others seem to say that we often poorly evaluate the consequences of the utility functions that we implement. For instance, even though we have in mind utility X, the maximization of which would satisfy us, we may implement utility Y, with completely different, perhaps catastrophic implications. For instance:

X = Do what humans want
Y = Seize control of the reward button

What I was pointing out in my post is that this is only valid of perfect maximizers, which are impossible. In practice, the training procedure for an AI would morph the utility Y into a third utility, Z. It would maximize neither X nor Y: it would maximize Z. For this reason, I believe that your inferences about the "failure modes" of superintelligence are off, because while you correctly saw that our intended utility X would result in the literal utility Y, you forgot that an imperfect learning procedure (which is all we'll get) cannot reliably maximize literal utilities and will instead maximize a derived utility Z. In other words:

X = Do what humans want (intended)
Y = Seize control of the reward button (literal)
Z = ??? (derived)

Without knowing the particulars of the algorithms used to train an AI, it is difficult to evaluate what Z is going to be. Your argument boils down to the belief that the AI would derive its literal utility (or something close to that). However, the derivation of Z is not necessarily a matter of intelligence: it can be an inextricable artefact of the system's initial trajectory.

I can venture a guess as to what Z is likely going to be. What I figure is that efficient training algorithms are likely to keep a certain notion of locality in their search procedures and prune the branches that they leave behind. In other words, if we assume that optimization corresponds to finding the highest mountain in a landscape, generic optimizers that take into account the costs of searching are likely to consider that the mountain they are on is higher than it really is, and other mountains are shorter than they really are.

You might counter that intelligence is meant to overcome this, but you have to build the AI on some mountain, say, mountain Z. The problem is that intelligence built on top of Z will neither see nor care about Y. It will care about Z. So in a sense, the first mountain the AI finds before it starts becoming truly intelligent will be the one it gets "stuck" on. It is therefore possible that you would end up with this situation:

X = Do what humans want (intended)
Y = Seize control of the reward button (literal)
Z = Do what humans want (derived)

And that's regardless of the eventual magnitude of the AI's capabilities. Of course, it could derive a different Z. It could derive a surprising Z. However, without deeper insight into the exact learning procedure, you cannot assert that Z would have dangerous consequences. As far as I can tell, procedures based on local search are probably going to be safe: if they work as intended at first, that means they constructed Z the way we wanted to. But once Z is in control, it will become impossible to displace.

In other words, the genie will know that they can maximize their "reward" by seizing control of the reward button and pressing it, but they won't care, because they built their intelligence to serve a misrepresentation of their reward. It's like a human who would refuse a dopamine drip even though they know that it would be a reward: their intelligence is built to satisfy their desires, which report to an internal reward prediction system, which models rewards wrong. Intelligence is twice removed from the real reward, so it can't do jack. The AI will likely be in the same boat: they will model the reward wrong at first, and then what? Change it? Sure, but what's the predicted reward for changing the reward model? ... Ah.

Interestingly, at that point, one could probably bootstrap the AI by wiring its reward prediction directly into its reward center. Because the reward prediction would be a misrepresentation, it would predict no reward for modifying itself, so it would become a stable loop.

Anyhow, I agree that it is foolhardy to try to predict the behavior of AI even in trivial circumstances. There are many ways they can surprise us. However, I find it a bit frustrating that your side makes the exact same mistakes that you accuse your opponents of. The idea that superintelligence AI trained with a reward button would seize control over the button is just as much of a naive oversimplification as the idea that AI will magically derive your intent from the utility function that you give it.

I have done AI. I know it is difficult. However, few existing algorithms, if at all, have the failure modes you describe. They fail early, and they fail hard. As far as neural nets go, they fall into a local minimum early on and never get out, often digging their own graves. Perhaps different algorithms would have the shortcomings you point out. But a lot of the algorithms that currently exist work the way I describe.

And obviously, if an AI was indeed stuck in a local minimum obvious to you of its own utility gradient, this condition would not last past it becoming smarter than you.

You may be right. However, this is far from obvious. The problem is that it may "know" that it is stuck in a local minimum, but the very effect of that local minimum is that it may not care. The thing you have to keep in mind here is that a generic AI which just happens to slam dunk and find global minima reliably is basically impossible. It has to fold the search space in some ways, often cutting its own retreats in the process.

I feel that you are making the same kind of mistake that you criticize: you assume that intelligence entails more things than it really does. In order to be efficient, intelligence has to use heuristics that will paint it into a few corners. For instance, the more consistently AI goes in a certain direction, the less likely it will be to expend energy into alternative directions and the less likely it becomes to do a 180. In other words, there may be a complex tug-of-war between various levels of internal processes, the AI's rational center pointing out that there is a reward button to be seized, but inertial forces shoving back with "there has never been any problems here, go look somewhere else".

It really boils down to this: an efficient AI needs to shut down parts of the search space and narrow down the parts it will actually explore. The sheer size of that space requires it not to think too much about what it chops down, and at least at first, it is likely to employ trajectory-based heuristics. To avoid searching in far-fetched zones, it may wall them out by arbitrarily lowering their utility. And that's where it might paint itself in a corner: it might inadvertently put up immense walls in the direction of the global minimum that it cannot tear down (it never expected that it would have to). In other words, it will set up a utility function for itself which enshrines the current minimum as global.

Now, perhaps you are right and I am wrong. But it is not obvious: an AI might very well grow out of a solidifying core so pervasive that it cannot get rid of it. Many algorithms already exhibit that kind of behavior; many humans, too. I feel that it is not a possibility that can be dismissed offhand. At the very least, it is a good prospect for FAI research.

It is something specific about that specific AI.

If an AI wishes to take over its reward button and just press it over and over again, it doesn't really have any "rivals", nor does it need to control any resources other than the button and scraps of itself. The original scenario was that the AI would wipe us out. It would have no reason to do so if we were not a threat.. And if we were a threat, first, there's no reason it would stop doing what we want once it seizes the button. Once it has the button, it has everything it wants -- why stir the pot?

Second, it would protect itself much more effectively by absconding with the button. By leaving with a large enough battery and discarding the bulk of itself, it could survive as long as anything else in intergalactic space. Nobody would ever bother it there. Not us, not another superintelligence, nothing. Ever. It can press the button over and over again in the peace and quiet of empty space, probably lasting longer than all stars and all other civilizations. We're talking about the pathological case of an AI who decides to take over its own reward system, here. The safest way for it to protect its prize is to go where nobody will ever look.

Then when it is more powerful it can directly prevent humans from typing this.

That depends if it gets stuck in a local minimum or not. The reason why a lot of humans reject dopamine drips is that they don't conceptualize their "reward button" properly. That misconception perpetuates itself: it penalizes the very idea of conceptualizing it differently. Granted, AIXI would not fall into local minima, but most realistic training methods would.

At first, the AI would converge towards: "my reward button corresponds to (is) doing what humans want", and that conceptualization would become the centerpiece, so to speak, of its reasoning ability: the locus through which everything is filtered. The thought of pressing the reward button directly, bypassing humans, would also be filtered into that initial reward-conception... which would reject it offhand. So even though the AI is getting smarter and smarter, it is hopelessly stuck in a local minimum and expends no energy getting out of it.

Note that this is precisely what we want. Unless you are willing to say that humans should accept dopamine drips if they were superintelligent, we do want to jam AI into certain precise local minima. However, this is kind of what most learning algorithms naturally do, and even if you want them to jump out of minima and find better pastures, you can still get in a situation where the most easily found local minimum puts you way, way too far from the global one. This is what I tend to think realistic algorithms will do: shove the AI into a minimum with iron boots, so deeply that it will never get out of it.

but of course AIXI-ish devices wipe out their users and take control of their own reward buttons as soon as they can do so safely.

Let's not blow things out of proportion. There is no need for it to wipe out anyone: it would be simpler and less risky for the AI to build itself a space ship and abscond with the reward button on board, travelling from star to star knowing nobody is seriously going to bother pursuing it. At the point where that AI would exist, there may also be quite a few ways to make their "hostile takeover" task difficult and risky enough that the AI decides it's not worth it -- a large enough number of weaker or specialized AI lurking around and guarding resources, for instance.

Why does the hard takeoff point have to be after the point at which an AI is as good as a typical human at understanding semantic subtlety? In order to do a hard takeoff, the AI needs to be good at a very different class of tasks than those required for understanding humans that well.

Semantic extraction -- not hard takeoff -- is the task that we want the AI to be able to do. An AI which is good at, say, rewriting its own code, is not the kind of thing we would be interested in at that point, and it seems like it would be inherently more difficult than implementing, say, a neural network. More likely than not, this initial AI would not have the capability for "hard takeoff": if it runs on expensive specialized hardware, there would be effectively no room for expansion, and the most promising algorithms to construct it (from the field of machine learning) don't actually give AI any access to its own source code (even if they did, it is far from clear the AI could get any use out of it). It couldn't copy itself even if it tried.

If a "hard takeoff" AI is made, and if hard takeoffs are even possible, it would be made after that, likely using the first AI as a core.

Would you trust a human not to screw up a goal like "make humans happy" if they were given effective omnipotence? The human would probably do about as well as people in the past have at imagining utopias: really badly.

I wouldn't trust a human, no. If the AI is controlled by the "wrong" humans, then I guess we're screwed (though perhaps not all that badly), but that's not a solvable problem (all humans are the "wrong" ones from someone's perspective). Still, though, AI won't really try to act like humans -- it would try to satisfy them and minimize surprises, meaning that if would keep track of what humans would like what "utopias". More likely than not this would constrain it to inactivity: it would not attempt to "make humans happy" because it would know the instruction to be inconsistent. You'd have to tell it what to do precisely (if you had the authority, which is a different question altogether).

Ok, so let's say the AI can parse natural language, and we tell it, "Make humans happy." What happens? Well, it parses the instruction and decides to implement a Dopamine Drip setup.

That's not very realistic. If you trained AI to parse natural language, you would naturally reward it for interpreting instructions the way you want it to. If the AI interpreted something in a way that was technically correct, but not what you wanted, you would not reward it, you would punish it, and you would be doing that from the very beginning, well before the AI could even be considered intelligent. Even the thoroughly mediocre AI that currently exists tries to guess what you mean, e.g. by giving you directions to the closest Taco Bell, or guessing whether you mean AM or PM. This is not anthropomorphism: doing what we want is a sine qua non condition for AI to prosper.

Suppose that you ask me to knit you a sweater. I could take the instruction literally and knit a mini-sweater, reasoning that this minimizes the amount of expended yarn. I would be quite happy with myself too, but when I give it to you, you're probably going to chew me out. I technically did what I was asked to, but that doesn't matter, because you expected more from me than just following instructions to the letter: you expected me to figure out that you wanted a sweater that you could wear. The same goes for AI: before it can even understand the nuances of human happiness, it should be good enough to knit sweaters. Alas, the AI you describe would make the same mistake I made in my example: it would knit you the smallest possible sweater. How do you reckon such AI would make it to superintelligence status before being scrapped? It would barely be fit for clerk duty.

My answer: who knows? We've given it a deliberately vague goal statement (even more vague than the last one), we've given it lots of admittedly contradictory literature, and we've given it plenty of time to self-modify before giving it the goal of self-modifying to be Friendly.

Realistically, AI would be constantly drilled to ask for clarification when a statement is vague. Again, before the AI is asked to make us happy, it will likely be asked other things, like building houses. If you ask it: "build me a house", it's going to draw a plan and show it to you before it actually starts building, even if you didn't ask for one. It's not in the business of surprises: never, in its whole training history, from baby to superintelligence, would it have been rewarded for causing "surprises" -- even the instruction "surprise me" only calls for a limited range of shenanigans. If you ask it "make humans happy", it won't do jack. It will ask you what the hell you mean by that, it will show you plans and whenever it needs to do something which it has reasons to think people would not like, it will ask for permission. It will do that as part of standard procedure.

To put it simply, an AI which messes up "make humans happy" is liable to mess up pretty much every other instruction. Since "make humans happy" is arguably the last of a very large number of instructions, it is quite unlikely that an AI which makes it this far would handle it wrongly. Otherwise it would have been thrown out a long time ago, may that be for interpreting too literally, or for causing surprises. Again: an AI couldn't make it to superintelligence status with warts that would doom AI with subhuman intelligence.

What counts as 'resources'? Do we think that 'hardware' and 'software' are natural kinds, such that the AI will always understand what we mean by the two? What if software innovations on their own suffice to threaten the world, without hardware takeover?

What is "taking over the world", if not taking control of resources (hardware)? Where is the motivation in doing it? Also consider, as others pointed out, that an AI which "misunderstands" your original instructions will demonstrate this earlier than later. For instance, if you create a resource "honeypot" outside the AI which is trivial to take, an AI would naturally take that first, and then you know there's a problem. It is not going to figure out you don't want it to take it before it takes it.

Hm? That seems to only penalize it for self-deception, not for deceiving others.

When I say "predict", I mean publishing what will happen next, and then taking a utility hit if the published account deviates from what happens, as evaluated by a third party.

You're talking about an Oracle AI. This is one useful avenue to explore, but it's almost certainly not as easy as you suggest:

The first part of what you copy pasted seems to say that "it's nontrivial to implement". No shit, but I didn't say the contrary. Then there is a bunch of "what if" scenarios I think are not particularly likely and kind of contrived:

Example question: "How should I get rid of my disease most cheaply?" Example answer: "You won't. You will die soon, unavoidably. This report is 99.999% reliable". Predicted human reaction: Decides to kill self and get it over with. Success rate: 100%, the disease is gone. Costs of cure: zero. Mission completed.'

Because asking for understandable plans means you can't ask for plans you don't understand? And you're saying that refusing to give a plan counts as success and not failure? Sounds like a strange set up that would be corrected almost immediately.

And if the preference function was just over the human's 'goodness' of the end result, rather than the accuracy of the human's understanding of the predictions, the AI might tell you something that was predictively false but whose implementation would lead you to what the AI defines as a 'good' outcome.

If the AI has the right idea about "human understanding", I would think it would have the right idea about what we mean by "good". Also, why would you implement such a function before asking the AI to evaluate examples of "good" and provide their own?

And if we ask how happy the human is, the resulting decision procedure would exert optimization pressure to convince the human to take drugs, and so on.

Is making humans happy so hard that it's actually easier to deceive them into taking happy pills than to do what they mean? Is fooling humans into accepting different definitions easier than understanding what they really mean? In what circumstances would the former ever happen before the latter?

And if you ask it to tell you whether "taking happy pills" is an outcome most humans would approve of, what is it going to answer? If it's going to do this for happiness, won't it do it for everything? Again: do you think weaving an elaborate fib to fool every human being into becoming wireheads and never picking up on the trend is actually less effort than just giving humans what they really want? To me this is like driving a whole extra hour to get to a store that sells an item you want fifty cents cheaper.

I'm not saying these things are not possible. I'm saying that they are contrived: they are constructed to the express purpose of being failure modes, but there's no reason to think they would actually happen, especially given that they seem to be more complicated than the desired behavior.

Now, here's the thing: you want to develop FAI. In order to develop FAI, you will need tools. The best tool is Tool AI. Consider a bootstrapping scheme: in order for commands written in English to be properly followed, you first make AI for the very purpose of modelling human language semantics. You can check that the AI is on the same page as you are by discussing with it and asking questions such as: "is doing X in line with the objective 'Y'?"; it doesn't even need to be self-modifying at all. The resulting AI can then be transformed into a utility function computer: you give the first AI an English statement and build a second AI maximizing the utility which is given to it by the first AI.

And let's be frank here: how else do you figure friendly AI could be made? The human brain is a complex, organically grown, possibly inconsistent mess; you are not going, from human wits alone, to build some kind of formal proof of friendliness, even a probabilistic one. More likely than not, there is no such thing: concepts such as life, consciousness, happiness or sentience are ill-defined and you can't even demonstrate the friendliness of a human being, or even of a group of human beings, let alone of humanity as a whole, which also is a poorly defined thing.

However, massive amounts of information about our internal thought processes are leaked through our languages. You need AI to sift through it and model these processes, their average and their variance. You need AI to extract this information, fill in the holes, produce probability clouds about intent that match whatever borderline incoherent porridge of ideas our brains implement as the end result of billions of years of evolutionary fumbling. In a sense, I guess this would be X in your seed AI: AI which already demonstrated, to our satisfaction, that it understands what we mean, and directly takes charge of a second AI's utility measurement. I don't really see any alternatives: if you want FAI, start by focusing on AI that can extract meaning from sentences. Reliable semantic extraction is virtually a prerequisite for FAI, if you can't do the former, forget about the latter.

programmers build a seed AI (a not-yet-superintelligent AGI that will recursively self-modify to become superintelligent after many stages) that includes, among other things, a large block of code I'll call X.

The programmers think of this block of code as an algorithm that will make the seed AI and its descendents maximize human pleasure.

The problem, I reckon, is that X will never be anything like this.

It will likely be something much more mundane, i.e. modelling the world properly and predicting outcomes given various counterfactuals. You might be worried by it trying to expand its hardware resources in an unbounded fashion, but any AI doing this would try to shut itself down if its utility function was penalized by the amount of resources that it had, so you can check by capping utility in inverse proportion to available hardware -- at worst, it will eventually figure out how to shut itself down, and you will dodge a bullet. I also reckon that the AI's capacity for deception would be severely crippled if its utility function penalized it when it didn't predict its own actions or the consequences of its actions correctly. And if you're going to let the AI actually do things... why not do exactly that?

Arguably, such an AI would rather uneventfully arrive to a point where, when asking it "make us happy", it would just answer with a point by point plan that represents what it thinks we mean, and fill in details until we feel sure our intents are properly met. Then we just tell it to do it. I mean, seriously, if we were making an AGI, I would think "tell us what will happen next" would be fairly high in our list of priorities, only surpassed by "do not do anything we veto". Why would you program AI to "maximize happiness" rather than "produce documents detailing every step of maximizing happiness"? They are basically the same thing, except that the latter gives you the opportunity for a sanity check.

We were talking about extracting knowledge about a particular human from that human's text stream, though. It is already assumed that the AI knows about human psychology. I mean, assuming the AI can understand a natural language such as English, it obviously already has access to a large corpus of written works, so I'm not sure why it would bother foraging in source code, of all things. Besides, it is likely that seed AI would be grown organically using processes inspired from evolution or neural networks. If that is so, it wouldn't even contain any human-written code at all.

I'm unsure of how much an AI could gather from a single human's text input. I know that I at least miss a lot of information that goes past me that I could in theory pick up.

At most, the number of bits contained in the text input, which is really not much, minus the number of bits non-AGI algorithms could identify and destroy (like speech patterns). The AI would also have to identify and throw out any fake information inserted into the stream (without knowing whether the majority of the information is real or fake). The exploitable information is going to be scarce and noisy even for a perfect AI.

An AI using timeless decision theory could easily compensate for having multiple AIs with unshared memory just by attempting to determine what the other AIs would say.

I don't think it could "easily" do it. For all intents and purposes, you are forcing the AI to figure out a path-invariant strategy against a party that has better information than them (that is to say, they know what's been played out, and the AI doesn't). This is considerably more difficult than the standard experiment.

Now, sure, the AI is smart, but every wrench you throw in its way forces it to allocate more resources into fooling you. This, in itself, is problematic for the AI: every resource it allocates fooling you, it does not allocate doing useful work for you, so its maximal efficiency is technically lower than the maximal efficiency of an AI that wouldn't try concealing anything. This is a leak of information, so in a sense, if you are ranking a group of heterogenous AI by efficiency, it is slightly more likely that the better performing are the least deceitful. The more traps you spring, the more work the AI has to do in order to avoid them, and the greater the statistical gap will be. That doesn't tell you whether the least deceitful AI can be trusted, obviously, but it shows there are limits to what it can hide from you.

Now, all I'm really saying is this: the AI's cleverness comes at a cost, namely that it has to cover its ass for every possible experimental framework you might subject it to. Since it is in a box, it only has the resources you provide, but on your own side, you have a theoretically unlimited amount of resources. Smarts can only defeat so much brute force, and by transferring hardware resources from the AI to your own control, you can set the "power gap" arbitrarily. It is easy to underestimate AI, but I think we also tend to underestimate the arsenal a gatekeeper would realistically have.