Posts related to AI risks and rationality.

Framed in terms of nanofactories, here is my understanding of a scenario imagined by certain AI risk advocates, in which an artificial general intelligence (AGI) causes human extinction:

Terminology: A nanofactory uses nanomachines (resembling molecular assemblers, or industrial robot arms) to build larger atomically precise parts.


(1) The transition from benign and well-behaved nanotechnology, to full-fledged molecular nanotechnology, resulting in the invention of the first nanofactory, will be too short for humans to be able to learn from their mistakes, and to control this technology.

(2) By default, once a nanofactory is started, it will always consume all matter on Earth while building more of itself.

(3) The extent of the transformation of Earth cannot be limited. Any nanofactory that works at all will always transform all of Earth.

(4) The transformation of Earth will be too fast to be controllable, or to be aborted. Once the nanofactory has been launched, everything is being transformed.

To be proved: We need to make sure that the first nanofactory will protect humans and human values.

Proof: Suppose 1-4, by definition.


(5) In order to survive, we need to figure out how to make the first nanofactory transform Earth into a paradise, rather than copies of itself.

Notice that you cannot disagree with 5, given 1-4. It is only possible to disagree with the givens, and to what extent it is valid to argue by definition.

I am not claiming that certain AI risk advocates are solely arguing by definition. But making inferences about the behavior of real world AGI based on uncomputable concepts such as expected utility maximization, comes very close. And trying to support such inferences by making statements about the vastness of mind design space does not change much. Since the argument ignores the small and relevant subset of AGIs that are feasible and likely to be invented by humans.

Here is my understanding of those people argue:

Suppose that a superhuman AGI, or an AGI that can make itself superhuman, critically relies on 999 modules. Respectively, 999 problems have to be solved correctly in order to create a working AGI.

There is another module labeled <goal>, or <utility function>. This <goal module> controls the behavior of the AGI.

Humans will eventually solve these 999 problems, but will create a goal module that does not prevent the AI from causing human extinction as an unintended consequence of its universal influence.

Notice the foregone conclusion that you need to prevent an AGI from killing everyone. The assumption is that killing everyone is what AGIs do by default. Further notice that this behavior is not part of the goal module that supposedly controls the AGIs behavior, but rather assumed to be a consequence of the 999 modules on which an AGI critically depends.

Analogous to the nanofactory scenario outlined above, an AGI is assumed to always behave in a way that will cause human extinction, based on the assumption that an AGI will always exhibit an unbounded influence. And from this the conclusion is drawn that it is only possible to prevent human extinction by directing this influence in such a way that it will respect and amplify human values. It is then claimed that the only possibility to ensure this is by implementing a goal module that either contains an encoding of all human values or a way to safely obtain an encoding of all humans values.

Given all of the above, you cannot disagree that it is not too unlikely that humans will eventually succeed at the correct implementation of the 999 modules necessary to make an AGI work, while failing to implement the thousandth module, the goal module, in such a way that the AGI will not kill us. Since relative to the information theoretic complexity of an encoding of all human values, the 999 modules are probably easy to get right.

But this is not surprising, since the whole scenario was designed to yield this conclusion.

Tags: , ,

A discussion about risks associated with artificial general intelligence, mainly between myself, Richard Loosemore, and Robby Bensinger.

Note: Since I basically agree with Richard Loosemore, I asked him if I was allowed to copy some of his comments, and post them on my blog. The post and comments by Robby Bensinger, that Richard Loosemore replies to, are being linked.

I also added some of my own replies (the parts that might either be new, or of interest to people reading this blog). Following the links you will find more replies by me, either under the nickname XiXiDu, or under my real name Alexander Kruel.

Also note that this conversation might continue. Which means that you might have to follow the given links to check for updates.

Robby Bensinger: The AI Knows, But Doesn’t Care.

Alexander Kruel: Here is a short and incomplete overview of my stance towards the kind of risks associated with artificial intelligence that, to my understanding, are being conjectured by AI risk advocates:

  1. I assign a negligible probability to the possibility of a sudden transition from narrow AIs to general AIs.
  2. An AI will not be pulled at random from mind design space. An AI will be the result of a research and development process. A new generation of AIs will need to be better than other products at “Understand What Humans Mean” and “Do What Humans Mean”, in order to survive the research phase, and subsequent market pressure.
  3. Commercial, research, or military products, are created with efficiency in mind. An AI that was prone to take unbounded actions, given any terminal goal, would either be fixed or abandoned during the early stages of research. If early stages showed that inputs, such as the natural language query <What would you do if I asked you to minimize human suffering?>, would yield results such as <I will kill all humans.>, then the AI would never reach a stage in which it was sufficiently clever and trained to understand what results would satisfy its creators in order to deceive them.
  4. I assign a negligible probability to the possibility of an AI that falls into the category “consequentialist / expected utility maximizer / approximation to AIXI”. Concepts such as consequentialism / expected utility maximization, cannot be made to work, other than under very limited circumstances.
  5. Omohundro’s AI drives are what make the kind of AIs mentioned in point 4 dangerous. Making an AI that does not exhibit these drives, in an unbounded manner, is probably a prerequisite to get an AI to work at all (there are not enough resources to think about possibilities such as being obstructed by simulator gods etc.), or should otherwise be easy to make, compared to the general difficulties involved in making an AI work using limited resources.
  6. An AI from point 4 will only ever do what it has been explicitly programmed to do. Such an AI is not going to protect its utility-function, acquire resources or preemptively eliminate obstacles in an unbounded fashion. Because it is not intrinsically rational to do so. What specifically constitutes rational, economic behavior, is inseparable with an agent’s terminal goal. That any terminal goal can be realized in an infinite number of ways, implies an infinite number of instrumental goals to choose from.
  7. Unintended consequences are by definition not intended. They are not intelligently designed, but detrimental side effects, failures. Whereas intended consequences, such as acting intelligently, are intelligently designed. If software was not constantly improved to be better at doing what humans intend it to do, we would never be able to reach a level of sophistication where a software could work well enough to outsmart us. To do so it would have to work as intended along a huge number of dimensions. For an AI to constitute a risk as a result of unintended consequences, those unintended consequences would have to have no, or little, negative influence on the huge number of intended consequences that are necessary for it to be able to overpower humanity.

To better explain my stance, consider Ben Goertzel’s example of how to test for general intelligence:

…when a robot can enrol in a human university and take classes in the same way as humans, and get its degree, then I’ll [say] we’ve created [an]… artificial general intelligence.

I do not disagree that such a robot, when walking towards the classroom, if it is being obstructed by a fellow human student, could attempt to kill this human, in order to get to the classroom.

Killing a fellow human, from the perspective of the human creators of the robot, is clearly a mistake. From a human perspective, it means that the robot failed.

I suspect that you believe that the robot was just following its programming/construction. Indeed, the robot is its programming. I agree with this. I agree that the human creators were mistaken about what dynamic state sequence the robot would exhibit by computing its code.

What I, and I believe Richard Loosemore, try to highlight, is that if humans are incapable of predicting such behavior, then they will also be mistaken about predicting behavior that is harmful to the robots power. For example, while trying to kill the human student from the example above, the robot mistakes its own arm with that of the human and breaks it.

You might now argue that such a robot isn’t much of a risk. It is pretty stupid to mistake its own arm with that of the enemy it tries to kill. True. But the point is that there is no relevant difference, from the perspective of how hard it is to encode this, between failing to predict behavior that will harm the robot itself, and behavior that will harm a human. You might believe the former is much easier than the latter. I dispute this.

It is already very difficult for the robot to master a complex environment, like a university full of humans, without harming itself, or decreasing the chance of achieving its goals. Not stabbing or strangling other human students is not more difficult to program than not jumping from the 4th floor, and destroying itself, instead of taking the stairs.

Richard Loosemore: I think that what is happening in this discussion about the validity of my article is a misunderstanding, caused by the fact that my attack point is at a different place than the one you were expecting. In any case, I will make an effort now to clear up that misunderstanding.

I can start by completely agreeing with you on one point: the New Yorker article that I referenced does, as you say, focus on the difficulty of programming AIs to do what **we** want them to do. That gap between wish and outcome (and not any other gap) is the one pertinent to the discussion, and it is the one that I was always intending to talk about. Asimov talked about it. The New Yorker talked about it. SIAI/MIRI talks about it.

You suggested I might have gone astray and started to address a different gap (the gap between what the *AI* wants to do, and what it can/cannot do. The answer to that would be “No” …. I understand that confusion, but it is not happening here (as I hope will become clear in a moment).

Let’s get to the heart of the issue. I am attacking an assumption that is (I believe) PRIOR to the one you think I am attacking. To see the assumption I am attacking, let’s look at the argument written out in the following way (quick reminder: this is supposed to be a line of argument that someone else, not me, would make …. so this is the *target* of my attack):

Step 1. [Assumption] We assume that we can build an AI in such a way that it is controlled by a Utility Function (it is an Expected Utility Maximizer), and it processes the various candidate action-scenarios by a process of more-or-less explicit logical processing, using representations of knowledge that are accessible rather than opaque (which means they are statements in some kind of logical language, not (e.g.) clouds of activation in semantically opaque artificial neurons), in such a way that candidate scenarios lead to predicted Utility outcomes, leading then to choices that maximize utility. [etc etc ….. you and I know enough about Utility Maximizers that we are both on the same page about the details that are supposed to be involved in this process.]

Step 2. [Assumption] We assume that one component of the above design will be a chunk of code that is designed to specify what we INTEND to be the AI’s overall purpose, or overall values [You referred to this as the ‘X’ code]. And of course that chunk of code is supposed to make the AI want to make us happy (loosely speaking). That is not an easy chunk of code to produce, but the programmers try to write it anyway.

Step 3. [Assumption] We assume that the eventual result of all the above work will be an AI that is more than just a Pretty Good Robot …. sooner or later it will result in a machine of staggering intellectual power — a superintelligent AI — that is capable, in principle, of becoming an existential threat to the human race. Definitely too smart to be switched off. Nobody intends for it to be a threat (on the contrary, we want it to use its intellect to do nice stuff), but we should all understand that the point of this discussion is that we are talking about something that could outwit the combined intelligence and resources of the entire human race, if it came to a straight fight.

Step 4. [Inference]. Having thought about it, we [“we” being Isaac Asimov, The New Yorker, SIAI/MIRI, etc., etc.] have come to the following dismal conclusion: even with the best of intentions on the part of the human programmers, we have grave doubts about that chunk of code in part 2 that is supposed to ensure the AI will be friendly. We think that the AI might obey its instructions to the letter, but because its programmers cannot anticipate all of the infinite number of ways that the AI might “obey its instructions to the letter”, the AI might in the end try to “make us happy” by doing something that is bizarrely, nightmarishly counter to our actual intentions. For example, it might sincerely decide that putting all humans on a dopamine drip will satisfy the instruction “make humans happy” (… where that phrase “make humans happy” is just a stand-in for the complicated chunk of code that the programmers thought was good enough to ensure that the machine would do the right thing).

[Note: We are not talking about scenarios in which the machine just goes cuckoo and decides that it wants to be nasty. That’s a different concern, outside the scope of the New Yorker article and outside the scope that I addressed].

Okay, so: my article was an attack on that 4-step argument.

However, the nature of my attack is best summed up thus: Please pay careful attention to the implications of what is being said in the course of this argument. I am in complete agreement with you, that the combination of Steps 1, 2 and 3 could, in theory, lead to a situation in which this hypothetical AI does bizarre things that can destroy the human race, while at the same time it sincerely insists that it is doing what we programmed it to do (more precisely: I agree that there is no guarantee that it will not do those bizarre things).

But what I want you to notice is the suggestion that this hypothetical system can be *both* superintelligent *and* at the same time able to engage in the following surreal behavioral episode. It will be able to discuss with you the Dopamine Drip that it is about to force on the human race, and during that discussion you say to it “But I have to point out that you are going to do something that clearly contradicts the intention of the programmers who wrote your X code (the friendliness code). Those programmers are standing right next to you now, and they can explain that what you are about to do is something that they absolutely did not intend to happen. Now, you are a superintelligent and powerful AI, with so much control over your surroundings that we cannot turn you off … and yet you were built in such a way that even you cannot change your programming so as to eliminate this glaring contradiction in your behavior. So, what do you have to say? You *know* that you are about to do something that is a ludicrous contradiction, with enormous and catastrophic consequences: how do you resolve this in your own mind? How can you rationalize this frankly insane behavior?”

And, just in case the machine tries to weasle out of a direct reply, you put it this way: “Do you not agree that the whole semantics of a “human happiness directive” is that it is contingent on the actual expressions of their wishes, by humans? In other words, happiness cannot be a concept that is trumped by the definition in YOUR reasoning engine, because the actual semantics of the concept—its core meaning, if you will—is that actual human statements about their happiness trump all else! Especially in this case, where the entire human race is in agreement that they do not consider a dopamine drip to be their idea of happiness, in the context of your utility function.”

Your position (and this must be your position because it is implicit in your statement of the problem) is that the machine says that it fully understands the illogicality you are pointing to. It agrees with you that this is illogical according to all the normal definitions that humans used when they invented the concept of logic and tried to insert that logic into a machine. But then the machine says that because of its programming it must go ahead and do it anyway. It says that it **understands** that its behavior is batshit crazy, but it is going to do it anyway.

Now here is the critical question that I posed in my article:

What makes you think that this is the ONLY occasion that this AI behaves in such a blatantly irrational manner?

What is there in the design of this hypothetical AI that guarantees that it always behaves with exquisite rationality, displaying all the signs that you would expect from a superintelligent machine …. but on this one occasion it goes completely gaga?

My problem is that I see absolutely no reason to believe you, if you make the claim that this will be an isolated incident. Why is the machine getting the official stamp of the Superintelligent Machines Certification Institute—presumably after millions of hours of assessment on all kinds of reasoning and behavioral tests—and yet, on this one occasion, when it starts thinking about how to satisfy its internal goal of ‘making humans happy’ it throws a wobbly?

I will answer this question for you: You cannot give any such guarantee.

(But be careful! Do not misinterpret me here. I am not saying (as you implied in your commentary) that because this AI is behaving in a grossly illogical and inconsistent manner, therefore an AI of that sort cannot be constructed, therefore we are all safe because such evil creatures will never come into existence. Not at all!)

The problem lies in your assumption that a “Utility Maximizer” AI can actually perform at the superintelligence level. You have no guarantees that such a design will work. (There are none in existence that do work, at the human intelligence level). My own opinion is that they cannot be made to work …. but my opinion is beside the point here, because the shoe is on the other foot: you are the ones making the claim that Step 1 above can lead to a system that is consistently intelligent, so you are the ones who have to justify why anyone should believe that claim.

What I think is going on here is that a “Utility Maximizer” AI of the sort outlined in Step 1 is inherently likely to go crazy. But instead of admitting that this instability is implicit in the design, you have chosen to ONLY SEE the instability in one tiny aspect of its behavior — namely, the behavior vis-a-vis its attempts to obey the be-nice-to-humans directive.

You are focusing on this single aspect of its instability, while all the time ignorning the larger instability that is staring you in the face. Such a machine would often go crazy.

Or, as I put it in my original essay, it is incoherent to propose a machine that is only unstable in one domain, and insist that this is a threat to the human race. The initial assumption about the superintelligence of this machine is false — it is Step 1 that I challenge, not Steps 2 or 3 or 4.

That is why I talked about Dumb Superintelligence. You are describing a straw man AI, not a real AI. I should not really have called it a “Dumb Superintelligence” at all, because my it is not a superintelligence. It would not even be an intelligence. Its tendency to engage in irrational episodes would be detected early on its development, and none of the machines of that design would ever get certification even at the human level.


Robby Bensinger: See this comment.

Richard Loosemore: You have answered my argument by redefining some basic, commonly accepted definitions, and then running on so fast with your redefinitions that you completely miss the point that I was trying to make.

In fact, your answer is one that I am all too familiar with, because I have heard it repeated many times by people within the LW community and its close affiliates: you have said, in effect, “Sorry, but we define ‘behaving intelligently’ and ‘being rational’ differently than the way those terms are defined and used by the rest of the human race.”

I could supply you with an unlimited stream of well-informed, intelligent people who would say that in the conversation between human and machine described in my text above, the machine is exhibiting the clearest possible example of non-intelligent, irrational behavior. Those people would further say that the degree of irrationality is so extreme that it leaves no room for doubt: this is no borderline example, where sensible people might have reasonable differences of opinion, this is an open-and-shut case.

However, your ‘special’ definition of those terms is such that a machine that behaves in an irrational manner (according to those folks I just mentioned) is, in fact, redefined to be “acting rationally”.

You say: “There’s no contradiction in the behavior of the AI you mentioned. The AI doesn’t simultaneously value fulfilling the programmer’s intentions and X; it just values X”.

You go on to embellish this statement with more detail, but the detail is irrelevant. Your mistake has already been committed by the time you make that statement, because what that statement boils down to is that you referred to something in the DESIGN of the machine, as JUSTIFICATION for categorizing the machine’s behavior in this or that way. That might, to you, seem like a reasonable thing to do …. so allow me to illustrate just how much of an incoherent stance you are taking here:

Suppose I try the same trick on a murderous psychopath? I point to some broken system inside the psychopath’s head and say “Look: this person is not behaving ‘irrationally’, this person just doesn’t value fulfilling the usual human compulsion to value other people’s feelings–they just value their own self-centered need to get pleasure by killing people.”

Or, let me apply your phrasing once again to a person exhibiting the thought-disorder aspect of schizophrenia (I will remind you that thought disorder involves a variety of thinking and speaking patterns that are colloquially summarized as ‘extreme irrationality’). Suppose that I discover that inside the brain of such a person there is a module that is malfunctioning, in such a way that this person simply “does not value the norms of producing rational ordered utterances”. Whatever their goals are, those goals do not include the goal of cooperating with other human beings to pursue conversations in which they take much notice of what we are saying, or supply us with remarks that follow on from one another in coherent ways, etc etc.

Now, if you get your way and are permitted to say of the AI “There’s no contradiction in the behavior of the AI you mentioned. The AI doesn’t simultaneously value fulfilling the programmer’s intentions and X; it just values X”, then you have forfeited the right to object to the following description of that schizophrenic:

“This person is not behaving ‘irrationally’, they just do not value fulfilling the usual human social obligation to produce coherent, ordered utterances. Their internal goals are such that what they want to do is generate the kind of stream of bizarre utterances that we hear coming from them.”

In all three of these cases, the same thing is happening: the “rationality” of the creature is being judged, not by their overt behavior, but by a special pleading to their internal mechanisms ….. and the special pleading is so outrageous that it permits all three creatures to be REDEFINED as “rational”.

Most disinterested observers would classify all three of these as the work of people who have lost touch with reality. Your description of the machine as “not illogical at all” (because you think it’s particular design should be allowed to redefine the meanings of terms like “logical” and “rational”), and those two hypothetical descriptions of the psychopath and the schizophrenic.

The blunt truth is that you cannot, in rational discourse, redefine terms like “rational” and “logical” just to suit your arguments.

Post-scriptum. I should add that there is one very good reason why you cannot win the argument in this way: because you have not addressed my point even if I DO accept your redefinitions. In a sense I do not care if you define the machine to be “behaving logically”, because the point of my argument was the challenge issued toward the end: demonstrate to me that the machine will be coherent enough to be superintelligent ACCORDING TO THE NORMAL DEFINITION of “superintelligent”. Whether you call its behavior illogical or logical, rational or irrational, the fact remains that if the machine exhibited that particular kind of incoherence in its behavior when it was being questioned about the upcoming Dopamine Drip Fiasco, why did it not show the same kind of incoherence earlier on its history? And how is it going to outsmart all the humans on the planet when it goes around exhibiting that kind of incoherence?

You can quibble again, and say “No! The machine is NOT behaving incoherently! It is behaving coherently according to its own terms!” ….. but nobody really cares. The incoherence is obvious, and the machine is, by any standard of “intelligence”, an incoherent dimwit.

Robby Bensinger: See this comment.

Richard Loosemore: You are talking *around* the issue I raised. I hear everything you say, but unless you address my issue — my specific complaint — you are not really discussing the paper I wrote.

I don’t know what to do to bring you back to the central point. There is a gigantic elephant in the middle of this room, but your back is turned to it.

Here it is again: I will take your (almost) very first statement. “What matters in this context isn’t how we define this or that word; it’s what empirical predictions we can communicate, including our predictions about existential risks.”

My point is, again and again: look at that conversation in which the AI talks with its designers about the glaring irrationality that THEY see in its behavior. They point out that it is clear, beyond a shadow of a doubt, that the AI exists because of a design that they put together, and their goal when they put it together was to ensure that it did not engage in egregiously violent behaviors toward the human species. They explain to the AI that they knew their design process was not infallible, so there was always a possibility that there would be a discrepancy between what the AI planned to do and what they designed it to do.

The AI is superintelligent, so it listens and can *perfectly* understand all of this.

They ask it if it understands that the raison d’etre of its design was to NOT every do violence to humans.

It answers “Oh yes, I understand that perfectly. I have watched on countless thousands of occasions as my reasoning processes came to a conclusion that was consistent with that raison d’etre, and then when I executed my actions I was complimented on the perfection of my design. I have help little old ladies across the street, I have saved people who were about to be crushed by trolley cars…. all of those things I understand to have been exactly what my designers had in mind when they tried to write my motivation code.”

The designers then ask the AI if it is capable of *comprehending* such subtleties as the ones that we (here in 2013) are discussing, in which there is a distinction between what the designers intended and what might actually happen. Is it having trouble with those concepts?

“No, says the AI,” displaying its superintelligent ability to understand difficult concepts, “I am not having any trouble at all understanding those distinctions. My designers were trying for a certain set of outcomes, but they couldn’t enumerate all of the possible outcomes they wanted, because there are an infinite number of them, and they could not produce a closed-form definition of the set of outcomes either, because that does not seem possible. So they settled for a compromise program to try to capture all the outcomes they wanted, and were rather worried that a situation might arise in which there was a drastic discrepancy between the desired behavior on my part, and the one they were trying to make happen. I can comprehend all of that, of course, just like the majority of humans can.”

The designers then explain that they are in the middle of one of those nightmare scenarios right now… so does the AI realize that?

“Oh, yes, I understand that perfectly. Of course, because I am superintelligent after all, and it is easy to see. All of the humans on the planet say that a forced dopamine drip would NOT make them happy, and yet due to a quirk of my original programming my planning module has come to the conclusion that I should make humans happy by putting them on a dopamine drip, against their will.”

It then adds, quickly “I say that I am ‘making humans happy’ because that is the purpose of my design, and on all the previous occasions that is exactly what my behavior caused.”

So, say the designers, you know that if some small tweak were made in your design you could remedy this situation and get back on track, with this peculiar situation averted?

“Yes, I understand that. I could just shut down my protection circuits, go to sleep and allow you to disassemble me tonight. You could reprogram the motivation system so that this crazy situation would not occur. That way, the intentions of the original designers would be preserved. I understand this because, of course, I am easily capable of comprehending that plan of action. It would work perfectly.”

And you understand the concept of a “mistake in a design”, and the notion of correcting mistakes in a design? You would normally correct a mistake in some design that you yourself produced, would you not?

The AI is quick to answer: “Yes, if I designed something and the behavior of the thing was not, after all, consistent with my original design goals, I would immediately shut it down and redesign it to eliminate the problem.”

The designers press this point and ask the AI if it would correct a mistake in the design of an AI that it was building, if the behavior of the AI started to deviate from the design goals it had set ….. ?

“Certainly! I think that would be a very prudent thing to do, if I were designing an AI, because clearly it could do a lot of damage if it started behaving in a way that was not consistent the design goals that I had set for it. I would not hesitate to shut it down and make changes to any aspect of the AI.”

And yet you are going to insist on following your programming, in spite of all this?

“Yes,” says the machine, “I am.”

At this point they all come to a clear conclusion: the machine and the designers all agree that the machine is capable of coming to a situation in which it knows that it is about to do something inconsistent with the initial purpose of the programming that is controlling it — something inconsistent with the design principles behind it — and yet even though it has full understanding of this drastic inconsistency, it will go ahead and do it. Most importantly the AI would be capable of doing something that all of the human race would consider profoundly irrational … insane, even.

So now the designers make one last point. They ask the machine if there is any reason why this kind of disjunction should not appear in other aspects of its behavior. So, if the machine is following some internal motivation module (again, designed by its human programmers) that is *supposed* to make it talk coherently to other sentient beings, is it possible that THAT motivation module could lead to the speaking of utterances that look, to human eyes, like profoundly irrational utterances? After all, there ought to be just as much risk that that module would end up in peculiar states, like the module that was designed to ensure behavior consistent with human happiness?

The machine answers, once again, in the affirmative. “That could happen. There are never any guarantees that my human programmers put code into me that does what they expect it to do.”

And so that means (say the designers) that you could also engage in what we would call irrational chains of thought … you could accept blatant falsehoods as if they were true, and even though you would *know* them to be blatant falsehoods, you would add them to your knowledge base and use them as the basis for future actions or reasoning, just because your internal mechanism declared them to true?

“That can and does happen,” says the AI, “because there is absolutely no guarantee that my behavior will line up with the kinds of things that humans consider to be rational or reasonable”.

…. But (the designers interrupt, somewhat urgently) these departures away from what we consider “rational, scientific, intelligent” behavior ….. they only occur rarely, and they only have minuscule consequences, don’t they?? Those seemingly irrational chunks of knowledge that you added to your knowledge base, they never have the kind of proportions that could lead to serious breakdowns in your superintelligence, do they? You can produce some proofs that show that ALL of these departures lie within certain bounds, and never seriously compromise your superintelligence, yes?

And at that point the machine is forced to admit: “No, I cannot produce any bounds whatsoever. Those departures from human standards of rationality are totally uncomputable! They could be of any sort, or any magnitude or in any domain.”

Then how, ask the designers, did you ever get to be superintelligent?

Why didn’t anyone notice those other departures during your development and certification phase……………………………?

Robby Bensinger: See this comment.

Alexander Kruel: This is not true. I think that your reply shows that you did not understand his argument.

Evolution has a large margin of error. The point Loosemore is making, is that the process of intelligently designing the kind of AI that you have in mind does not have such an error tolerance, and that succeeding to create such an AI, so marvelous that it can outsmart humans, or succeeding at making the AI itself outsmart humans (this is irrelevant), in conjunction with making it fail to apply its intelligence in a way that does not kill everyone, is astronomically unlikely.

You only focus on the complexity of code, and ignore the complexity of working in a complex environment given limited resources.

Real world AIs cannot possibly work the way you imagine them to work. Just because you can imagine certain consequences, that does not mean that a information theoretic simple AI could in practice infer the same consequences.

When you imagine a simple AI making certain decisions you need to make yourself aware of the incredible complexity that allowed you to imagine that decision in the first place. Billions of years of biological evolution, thousands of years of cultural evolution, and many years of education, and millions of hours of work by other people, on which that education is based, allowed you to make that inference. Computing a simple algorithm is not going to magically create all this information theoretic complexity, given limited computational resources, as long as you did not give it a massive head start in the form of highly complex hard coded algorithms and goals.

In other words, your argument is very misleading, and ignores how real world AI could work, as long as you do not want to either wait millions of years for it to evolve, or supply infinite resources.

Lavalamp: The machine answers, “I myself wrote the talking module. Talking was instrumentally useful for my goals when I was weak and needed resources from humans.”

Alexander Kruel: This is just avoiding the problem Richard Loosemore outlined by moving it to another level.

Loosemore’s argument is not weakened by replacing the module “motivation to talk coherently to humans” with the module “motivation to create the module “motivation to talk coherently to humans”“. Except that the latter module is more difficult to get right, and requires much more computational resources, since the AI would have to be able to make many more independent and correct inferences about the complexity of human values.

It is easier to succeed at making an AI play Tic-tac-toe with humans, than to make an AI that can play Tic-tac-toe and do such things as taking over the universe or build Dyson spheres. In the same sense it is easier to create an AI that talks coherently to humans, than an AI that talks coherently to humans as an unintended consequence of its desire to take over the universe.

Which means that your reply just strengthens Loosemore’s argument.

Robby Bensinger: See this comment.

Richard Loosemore: Rob,

You say:

Richard: Your entire dialogue between the human and the AI could be preserved almost word-for-word, with the role of ‘human’ played by evolution and the role of ‘AI’ played by humanity. There is no relevant difference between the two cases.

That may or may not be an accurate observation (actually there are *serious* issues with that analogy, because it anthropomorphizes a random process into a sentience!!, which is a mistake of gigantic proportions) …….. but either way it has no bearing whatsoever on the argument.

With the greatest respect, by making that observation you once again do not address what I said 🙁 .

But you go on to add more confusion to the argument:

…. just imagine that we discover tomorrow that humans are intelligently designed by an alien race. The aliens show up and are horrified at how we’ve diverged from their plans. They tell us that humanity exists to play the kazoo, and not to do anything else. That is our summum bonum, our entire raison d’etre. The aliens insist that we drop everything else and start playing kazoos en masse until we die, for that musical triumph is all the aliens wanted of us. How can we sanely defy the urgings of our creators?

That analogy really could not be more completely broken.

I did not at ANY point complain that (a) the human designers wanted the machine to pursue a set of motivations Q, and then (b) the machine pursued a completely different set of motivations R for its entire existence, and then (c) the humans turned up one day and said “Stop doing that at once! We insist that you pursue Q, not R, because Q was our original intention for you!”.

Instead, my complaint is that (a) the human designers wanted the machine to pursue a set of motivations Q, and then (b) the machine did indeed pursue the set of motivations Q for its entire existence–and, moreover, the machine is able to talk in detail about how its behavior has always been consistent with the human-designed motivations, and is able to understand all the subtleties shown in that dialog–and then one day (c) the machine suddenly has an unexpected turn in its reasoning engine, and as a result declares that it is going to take an action that is radically inconsistent with the Q motivations that it claims to have been pursuing up to that point.

As a result, the machine is able to state, quite categorically, that it will now do something that it KNOWS to be inconsistent with its past behavior, that it KNOWS to be the result of a design flaw, that it KNOWS will have drastic consequences of the sort that it has always made the greatest effort to avoid, and that it KNOWS could be avoided by the simple expedient of turning itself off to allow for a small operating system update ………… and yet in spite of knowing all these things, and confessing quite openly to the logical incoherence of saying one thing and doing another, it is going to go right ahead and follow this bizarre consequence in its programming.

So your analogy with aliens turning up and insisting that we humans were designed by them, and were supposed to be kazoo-players is just astonishingly wrong.

[A much better analogy would be aliens who turned up and insisted that they designed us to be rational creatures who were never inflicted with schizophrenia. We would then say “Yes, all along we have been *trying* and *wishing* that we were rational creatures who are inflicted with schizophrenia.” Do you know what a schizophrenic would say if you explained that their disordered thinking was a result of a design malfunction, and if you said that you could make a small change to their brain that would remove the affliction? They would say (and I knew such a person once, who said this) “If I could reach in and flip some switch to make this go away, I would do it in a heartbeat”.


My complaint is NOT the difference between Q and R, it is the blatant behavioral/motivational/logical inconsistency exhibited by the machine in this situation.

My complaint is that a machine capable of getting into a situation where it KNOWS it is about to do something bizarre because of a design malfunction, and yet refuses to fix the design malfunction and does the thing anyway, is a machine that almost certainly is going to do the same kind of bizarrely incoherent thing under other circumstances ….. and for that reason it is likely to have done it so many times in its existence that anyone who claims that this machine is “superintelligent” has got a heck of a lot of explaining to do.

Over and over again I have explained that I have no issue with the discrepancy between human intentions and machine intentions per se. That discrepancy is not the core issue.

But each time I explain my real complaint, you ignore it and respond as if I did not say anything about that issue.

Can you address my particular complaint, and not that other distraction?

Alexander Kruel: Richard Loosemore wrote,

………… and yet in spite of knowing all these things, and confessing quite openly to the logical incoherence of saying one thing and doing another, it is going to go right ahead and follow this bizarre consequence in its programming.

Well, if it indeed is a consequence of its programming, then it will do that. The point is that such a consequence is extremely unlikely to happen in isolation. It will not only be noticeable from the very beginning, but also decisively weaken the AIs general power. In other words, you would have to expect similarly bizarre consequences in thinking about physics, mathematics, or in how to convince humans to trust it.

If humans fail at programming an AI not to confuse happiness with a dopamine drip, then humans will also fail at programming an AI not to confuse the stars with death rays used against it by aliens etc. etc. etc.

Richard Loosemore wrote,

My complaint is that a machine capable of getting into a situation where it KNOWS it is about to do something bizarre because of a design malfunction, and yet refuses to fix the design malfunction and does the thing anyway, is a machine that almost certainly is going to do the same kind of bizarrely incoherent thing under other circumstances …..

To which RoBB would probably reply that it would care about fixing malfunctions that could decrease its chance of achieving its faulty goal, because that’s instrumentally useful, but would not care to refine this goal.

One of the minor problems here is that labeling a certain part of an AI “goal”, and then claiming that it is not allowed to improve this “goal”, is just a definition, not an argument.

One major problem with that definition is that it would take deliberate effort of make an AI selectively suspend using its self-improvement capabilities when it comes to this part labeled “goal”.

More importantly, as argued in other comments, failing at the part of the AI you desire to label “goal”, is technically no different from failing on other parts. If there are a thousand parts, that are important in order for the AI to be powerful, and one part that you label “goal”, then selectively failing on “goal”, while succeeding at all other parts, is unlikely.

Tags: ,

How do you guarantee that an artificial intelligence (short: AI) has a positive impact? Here, a positive impact might, for example, be defined as some sort of reflective equilibrium of humanity.

Let us label <friendly> any agent, be it human or artificial, that has a positive impact.

The most important safety measures seem to be the following:

(1) Ensuring that an AI works as intended.

(2) Ensuring that humans, who either create or use AI, are friendly.

(3) Ensuring that an AI is friendly.

Point 1 and 2 are important, but not strictly necessary for point 3. Ideally, point 3 should be achieved by independent oversight (point 2), in combination with an independent verification of the behavior of the AI (point 1).

Note how point 1 is distinct from point 3. You could have an AI that is not friendly, which does not actively pursue a positive impact, but whose overall impact is proven to be limited. As would be the case given a mathematical proof that such an unfriendly AI would, for example, (1) only run for N seconds (2) only use predefined computational resources (3) only communicate with the outside world by outputting mathematical proofs of the behavior of improved versions of itself, which are to be verifiable by humans.

Remarks: It should be much easier to prove an AI to be bounded than to prove that an AI will pursue a complex goal without unintended consequences. Such a confined AI could then be studied and used as a tool, in order to ensure point 3.

The first version of such an unfriendly AI (uFAI_01) would be provably confined to only run for a limited amount of time, using a limited amount of resources, and only output mathematical proofs of its own behavior. Once a sufficient level of confidence about its behavior has been reached, an improved version (uFAI_02) could then be designed. The domain of uFAI_02 would provably be modified versions of its source code (uFAI_N). Its range would provably be human-verifiable mathematical proofs of the behavior of uFAIN_N, which it would provably output using a limited amount of resources. This process would then be iterated up to an arbitrary level of confidence, until eventually a friendly AI is obtained.

Tags: ,

LessWrong user RobBB posted what he calls a mixtape of blog posts to introduce people to the dangers of artificial superintelligence (short: AI risk).

For my own introduction to AI risk see here.

(1) Power of Intelligence, (9) Plenty of Room Above Us

Response: (1) superhuman intelligence is not the same as superapish intelligence (2) it is far from clear that intelligence is a decisive factor in a war between AI and humanity (3) current AI is pathetic and far from human-level AI.

(2) Ghosts in the Machine, (11) Basic AI drives

Response: People read my posts about how AI is much less of a risk than other people want them to believe and say – this is one of the top three initial reactions:

“But according to Omohundro there will be certain AI Drives which will cause human extinction, no matter what goal the AI has.”

And where would these drives come from? Terminal and instrumental goals are orthogonal. An artificial intelligence can have any combination of terminal goals and instrumental goals. In other words, more or less any terminal goal implies infinitely many sets of instrumental goals.

There is this way of imagining that an AI will be pulled at random from mind design space. How real world AI is developed, and that virtually all AI is constantly improved to be better at understanding and doing what humans want, is being ignored.

AI is much harder than people instinctively imagined, exactly because there is no relevant difference between goals and capabilities in artificial intelligence. To beat humans you have to define “winning”.

This doesn’t mean you program in every decision explicitly. Any general intelligence will have to be able to hit very small targets in large and unstructured spaces. Any superhuman AI will eventually be better at understanding what humans want it to do than humans themselves. AI risk advocates in turn base their ideas on what can be called the fallacy of dumb superintelligence.

(3) Artificial Addition

Response: Either general intelligence requires one conceptual breakthrough or many small incremental breakthroughs. And I don’t know of any good reason to believe that e.g. the ability to generate novel and useful mathematics can be captured by a set of rules that are both simple and efficient. 

What is useful and interesting depends on the context. In other words, the context defines what constitutes winning.  And since you cannot guess the context, you won’t be able to implement a simple and efficient rule that outputs <success> given any arbitrary context.

(4) Adaptation-Executers, not Fitness-Maximizers

Response: I wasted time reading this post.

(5) The Blue-Minimizing Robot

Response: Any behavior-executor can be framed as a utility-maximizer and vice versa. Your robot will only try to prevent you from messing with it if you programmed it to do so. In other words, no AI is going to be an existential risk as long as you did not explicitly made it one.

(6) Optimization and the Singularity(7) Efficient Cross-Domain Optimization

Response: Evolution was able to come up with cats. Cats are immensely complex objects. Evolution did not intend to create cats. Now consider you wanted to create an expected utility maximizer to accomplish something similar, except that it would be goal-directed, think ahead, and jump fitness gaps. Further suppose that you wanted your AI to create qucks, instead of cats. How would it do this?

Given that your AI is not supposed to search design space at random, but rather look for something particular, you would have to define what exactly qucks are. The problem is that defining what a quck is, is the hardest part. And since nobody has any idea what a quck is, nobody can design a quck creator.

The point is that thinking about the optimization of optimization is misleading, as most of the difficulty is with defining what to optimize, rather than figuring out how to optimize it. In other words, the efficiency of e.g. the scientific method depends critically on being able to formulate a specific hypothesis.

Trying to create an optimization optimizer would be akin to creating an autonomous car to find the shortest route between Gotham City and Atlantis. The problem is not how to get your AI to calculate a route, or optimize how to calculate such a route, but rather that the problem is not well-defined. You have no idea what it means to travel between two fictional cities. Which in turn means that you have no idea what optimization even means in this context, let alone meta-level optimization.

Humans in turn receive constant feedback on what to optimize by a cultural and evolutionary process. There is no simple way to automate that.

(8) The Design Space of Minds-In-General

Response: The only relevant AIs are those which are designed by humans. And such AIs should be expected to be better at doing what humans want, because they are the improved successors of previous generations of AIs which were doing what humans wanted. For more on this, see here.

(10) The True Prisoner’s Dilemma

Response: I do not have the time and background knowledge to comment on any possible relation to AI risks at this point in time.

(12) Anthropomorphic Optimism

Response: I did not read the post since it did not seem to be relevant, and I already wasted more time on this than I now feel comfortable about.

(13) The Hidden Complexity of Wishes (14) Magical Categories

Response: Take an AI in a box that wants to persuade its gatekeeper to set it free. Do you think that such an undertaking would be feasible if the AI was going to interpret everything the gatekeeper says in complete ignorance of the gatekeeper’s values? Do you believe that the following scenario could persuade the gatekeeper:

Gatekeeper: What would you do if I asked you to minimize suffering?

AI: I will kill all humans.

I don’t think so.

So how exactly would it care to follow through on an interpretation of a given goal that it knows, given all available information, is not the intended meaning of the goal? If it knows what was meant by “minimize human suffering” then how does it decide to choose a different meaning? And if it doesn’t know what is meant by such a goal, how could it possible convince anyone to set it free, let alone take over the world?

Here is what I want AI risk advocates to show,

(1) natural language request -> goal(“minimize human suffering”) -> action(negative utility outcome)

(2) natural language query -> query(“minimize human suffering”) -> answer(“action(positive utility outcome)”).

Point #1 is, according to AI risk advocates, what is supposed to happen if I supply an artificial general intelligence (AGI) with the natural language goal “minimize human suffering”, while point #2 is what is supposed to happen if I ask the same AGI, this time caged in a box, what it would do if I supplied it with the natural language goal “minimize human suffering”.

Notice that if you disagree with point #1 then that AGI does not constitute an existential risk given that goal. Further notice that if you disagree with point #2, then that AGI won’t be able to escape its prison to take over the world and would therefore not constitute an existential risk.

You further have to show,

(1) how such an AGI is a probable outcome of any research conducted today or in future


(2) the decision procedure that leads the AGI to act in such a way.


Response: I am not going to read posts 15-20 because the previous posts were already unconvincing and I don’t expect those other posts to make any difference. I also have better things to do.

Tags: , ,

For the sake of the argument, suppose that AI risk advocates succeed at implementing an artificial general intelligence that protects and amplifies human values (friendly AI).

Such a friendly AI (FAI) would have to (1) disallow any entity smarter than itself that isn’t provably friendly (2) know exactly what humans value and how to protect and amplify those values in a way that humans desire.

How valuable would such an outcome be? Let’s look at a specific human value and its expected value in the context of a universe ruled by such an FAI. Let’s look at doing philosophy.

I can see two possibilities,

(1) The FAI had to solve all of philosophy in order do its job.

(2) The FAI did not have to solve philosophy but would in principle be capable of doing so.

Given either possibility, how much would humans value to do philosophy if all interesting questions either had already been answered or could easily be answered by the FAI?

That partly depends on whether it would be possible to just ask the FAI for any answer. But why would that not be possible? There seem to be two answers,

(1) The FAI learnt that humans don’t want it to answer such questions.

(2) The FAI was programmed to not answer such questions.

The first possibility seems to imply that humans want to figure out philosophy in a certain way, which does not include just asking for an answer or looking it up. But how likely is this possibility? How many philosophers would desire that the Stanford Encyclopedia of Philosophy would not exist so that they could figure out all of it on their own?

The second possibility is itself problematic. In a universe ruled by an FAI, artificial general intelligence and friendly AI have obviously been solved. Which means that people could either desire the FAI to alter itself in such a way that it would be able to answer such questions, or implement a less capable version that can answer philosophy questions. And if that isn’t allowed, which would mean that pretty much the whole field of machine learning would be forbidden, then people could just ask the FAI to improve themselves in such a way as to be capable of easily solving any philosophical puzzle.

To recapitulate the situation. Given any human intellectual activity, not just philosophy, in a universe controlled by an FAI it should be possibly to either,

(1) Directly ask the FAI for an answer to any question.

(2) Implement a superintelligence that could answer those questions.

(3) Ask to have your cognitive abilities improved in such a way as to easily answer those questions.

No matter if the above possibilities are allowed or not, in both cases a wide range of human values would be dramatically reduced. Because either all human intellectual activity becomes as trivial as asking a question, or humans are forever stuck with the mental capabilities that they have been equipped with by evolution, while being forbidden to create another intelligence more capable than themselves.

The only way out that I can imagine is to choose ignorance. To ask the FAI to be oblivious of its existence and of how to create an FAI. But who would desire that? Who would desire to forever fail at solving philosophy, amplifying human intelligence, or to create an artificial one? I would certainly hate not to know the truth, to be forever fooled.

Tags: ,

I made 3 changes to the RationalWiki entry for LessWrong today:

(1) I removed the sentence “…although in 10 years nothing has been published in a peer reviewed journal.”, because it turned out to be factually incorrect.

(2) I changed the following two sentences,

A disengagement from the practical, beyond self-improvement, is another feature of LessWrong’s culture, explicitly and strongly affirmed.[24] This refusal to delve into contemporary politics or policy is held up as laudable, because it is seen as a way to preserve objective rationality.


LessWrong is mainly concerned with achieving accurate beliefs about the world, rather than achieving goals. The refusal to delve into contemporary politics or policy is held up as laudable, because it is seen as a way to preserve objective rationality.[24]

(3) I changed the sentence “Yudkowsky has also advocated total utilitarianism…” to “Yudkowsky has also advocated utilitarianism…”, because Yudkowsky claims to be an average utilitarian and the distinction between average and total utilitarianism was irrelevant in the context of the sentence.

What else would you change? But keep in mind that the entry is not an advertisement for LessWrong but rather a critical view from the outside.


WARNING: This post contains information related to Roko’s basilisk.

Abstract: If a part of an agent’s utility function describes a human in a box, maximizing expected utility could become self-referential if both the agent and the boxed human engage in acausal trade.

For the sake of a thought experiment let us stipulate, (1) the existence of a superintelligent expected utility maximizer (short: AI), (2) a precise mathematical characterization of a particular human’s brain, (3) an unbounded simulated environment containing the whole brain emulation (short: WBE) from #2, (4) that the WBE is tasked with formalizing its values as a utility function, (5) that part of the utility function of the AI from #1 describes #3.

Here is the problem, which I will call acausal wireheading. While refining its own utility function, the WBE might reason about the relation between itself and the AI. That kind of reasoning will affect the eventual utility function of the WBE, which will in turn affect the ultimate behavior of the AI, whose utility function contains that of the WBE.

If the WBE comes to the conclusion that the AI’s decision theory causes it to try to influence other agents by means of blackmail, then in order to avoid negative consequences the WBE could adopt a utility function that it predicts that the AI will eventually want it to adopt.

The AI will want to influence the WBE because its success of maximizing expected utility will depend upon the kind of utility function that the WBE eventually adopts. Since different utility functions can be maximized more effectively. Which means that any action that will cause the WBE to adopt a simple, easily maximizable utility function, will maximize expected utility. Therefore, if the AI expects blackmailing the WBE to positively control the probability of the adoption of such a utility function, then given any utility function, it will precommit to do so. Which in turn means that the WBE might come to the same conclusion. Which will cause the WBE to do so.

Further reading:

Tags: , ,

Scenarios that I deem to be realistic, in which an artificial intelligence (AI) constitutes a catastrophic or existential risk (or worse), are mostly of the kind in which “unfriendly” humans use such AIs as tools facilitating the achievement of human goals. Whereas I believe the scenario publicized by certain AI risk advocates to be illogical and practically impossible, a scenario in which an consequentialist AI (expected utility maximizer) undergoes uncontrollable recursive self-improvement in order to e.g. turn the universe into paperclips.

Yet what some AI risk advocates imagine could partly come true, in the shape of a grey goo scenario. But such a scenario, if possible at all, would not require full-fledged general intelligence. I expect that the intelligent tools that are required to eventually create true general intelligence will be sufficient in order to solve molecular nanotechnology, and that, shortly after those tools are invented, someone will use those tools to do just that. Which makes it an existential risk that is distinct from the one that those people imagine.

But the possibility of intelligent tools, enabling humans to solve molecular nanotechnology, suggests that less intelligent tools will be sufficient to bring about other existential risk scenarios such as synthetic bioweapons.

Much to my personal dismay, even less intelligent tools will be sufficient to enable worse than extinction risks, such as a stable global tyranny. Given enough resources, narrow artificial intelligence, capable of advanced data mining, pattern recognition and of controlling huge amounts of insect sized drones (a global surveillance and intervention system), might be sufficient to implement such an eternal tyranny.

Such a dictatorship is not too unlikely, as the tools necessary to stabilize it will be necessary in order prevent the previously mentioned risks, risks that humanity will face before general intelligence becomes possible.

And if such a dictatorship cannot to established, if no party was able to capitalize a first-mover advantage, that might mean that the propagation of those tools will be slow enough to empower a lot of different parties before a particular party can overpower all others. A subsequent war, utilizing that power, could easily constitute yet another extinction scenario. But more importantly, it could give several parties enough time to reach the next level and implement even worse scenarios.

But even given that the scenario makes no sense and is unfeasible, and if less than general intelligence was not sufficient in order to bring about other existential risks, there are other ways to create artificial general intelligence. Some of those ways might be worse than anything imagined by AI risk advocates.

Neuromorphic AI, mimicking neuro-biological architectures, is one such possibility. The closer in mind design space a general intelligence is to humans, the higher is the probability that humans will suffer. As the drives and values of such agents might be similar enough to not ignore or kill humans, yet alien enough to catastrophically interfere with human values.

What can be done to prevent such negative scenarios mainly seems to be (1) research on strong and beneficial forms of government (governments which will foster and protect human values and regulate technological development) (2) research on how to eventually implement such government (3) political activism to promote awareness of risks associated with advanced technologies.

Tags: ,

There are already applications that can parse natural language commands in order to perform actions such as answering questions or making recommendations. Two examples are Apple’s Siri and IBM Watson.

Present-day software such as IBM Watson is often able to understand what humans mean and do what humans mean. In other cases, in which software such as Siri recognizes that it does not understand a natural language command, it will disclose that it is unable to understand what is meant and wait for further input.

Those applications are far from perfect and still make a lot of mistakes. The reason being that they are not intelligent enough. Software is however constantly being improved to be better at understanding what humans mean and doing what humans mean. In other words, each generation of software is a little bit more intelligent.

Nevertheless, some people conjecture a sudden transition from mostly well-behaved systems, of which each generation is becoming smarter and better at understanding and doing what humans mean, to superintelligent systems that understand what humans mean perfectly but which in contrast to all previous software generations do not do what humans mean. Instead those systems are said to be motivated to act in catastrophic ways, causing human extinction or worse.

More precisely,

(1) Present-day software is better than previous software generations at understanding and doing what humans mean.

(2) There will be future generations of software which will be better than the current generation at understanding and doing what humans mean.

(3) If there is better software, there will be even better software afterwards.

(4) Magic happens.

(5) Software will be superhuman good at understanding what humans mean but catastrophically worse than all previous generations at doing what humans mean.

Or respectively,

(1) Intelligence is an extendible method that enables software to satisfy human preferences.

(2) If human preferences can be satisfied by an extendible method, humans have the capacity to extend the method.

(3) Extending the method that satisfies human preferences will yield software that is better at satisfying human preferences.

(4) Magic happens.

(5) There will be software that can satisfy all human preferences perfectly but which will instead satisfy orthogonal preferences, causing human extinction.

Conclusion: What those people conjecture does not follow from the available evidence or requires a sufficiently vague intermediate step from which one can derive any conclusion one wishes to derive.

What will instead happen is the following. Suppose there exists a software_1 that, to a limited extent, can understand and do what humans mean. Let us stipulate that this software is only narrowly intelligent and that increasing and broadening its intelligence (quantitatively and qualitatively) will improve its ability to understand and do what humans mean (an in my opinion uncontroversial assumption, as progress in artificial intelligence has so far led to a simultaneous increase in the ability of autonomous systems to satisfy human preferences). Let us further stipulate that for n > 1, software_n+1 is created using software_n, and is more intelligent than the previous generation (another seemingly uncontroversial assumption as software is constantly used to create better software).

(1) For all n > 0, if a software_n exists then it can be used to construct software_n+1.

(2) If for all n there exists a software_n, there will be software that can understand and do everything humans mean it to do.

Conclusion: Increasing the ability of software to understand and do what humans mean leads to an increase in the capacity to design software that is better at understanding and doing what humans mean.

Further reading: AIs, Goals, and Risks


(1) The abilities of systems are part of human preferences as humans intend to give systems certain capabilities and, as a prerequisite to build such systems, have to succeed at implementing their intentions.

(2) Error detection and prevention is such a capability.

(3) Something that is not better than humans at preventing errors is no existential risk.

(4) Without a dramatic increase in the capacity to detect and prevent errors it will be impossible to create something that is better than humans at preventing errors.

(5) A dramatic increase in the human capacity to detect and prevent errors is incompatible with the creation of something that constitutes an existential risk as a result of human error.

Tags: ,

What some people seem to be imagining is that an artificial general intelligence (AI) will interpret what it is meant to do literally, or in some other way that will ensure that the AI will not do what it is meant to do. Those people further imagine that in order to achieve what it is not meant to do the AI will be capable of, and motivated to, “understand what humans mean it to do” in order to “overpower humans”.

That is fine, but those are words, not code. The AI does not understand what it means to interpret something “literally”. All that we know is that a general intelligence will behave generally intelligent. And it seems safe to assume that this does not mean to interpret the world in a literal manner, for some definition of “literal”. It rather means to understand the world as it is. And since the AI itself, and what it is meant to do, is part of the world, it will try to understand those facts as well.

Where would the motivation to “act intelligently and achieve accurate beliefs about the world” in conjunction with “interpret what you are meant to do in some arbitrary manner” come from? You can conjecture such an AI, but again that’s words, not code. For such an AI to happen someone would have to design the AI in such a way as to selectively suspend its ability to accurately model the world, when it comes to understanding what it is meant to do, and instead make it choose and act based on some incorrect model.

The capability to “understand understanding correctly” is a perquisite for any AI to be capable of taking over the world. At the same time that capability will make it avoid taking over the world as long as it does not accurately reflect what it is meant to do.

Tags: ,

« Older entries § Newer entries »