Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

Ethical Injunctions

24 Post author: Eliezer_Yudkowsky 20 October 2008 11:00PM

Followup toEthical Inhibitions, Ends Don't Justify Means (Among Humans), Entangled Truths, Contagious Lies, Protected From Myself

"Would you kill babies if it was the right thing to do?  If no, under what circumstances would you not do the right thing to do?  If yes, how right would it have to be, for how many babies?"
        —horrible job interview question

Swapping hats for a moment, I'm professionally intrigued by the decision theory of "things you shouldn't do even if they seem to be the right thing to do".

Suppose we have a reflective AI, self-modifying and self-improving, at an intermediate stage in the development process.  In particular, the AI's goal system isn't finished—the shape of its motivations is still being loaded, learned, tested, or tweaked.

Yea, I have seen many ways to screw up an AI goal system design, resulting in a decision system that decides, given its goals, that the universe ought to be tiled with tiny molecular smiley-faces, or some such.  Generally, these deadly suggestions also have the property that the AI will not desire its programmers to fix it.  If the AI is sufficiently advanced—which it may be even at an intermediate stage—then the AI may also realize that deceiving the programmers, hiding the changes in its thoughts, will help transform the universe into smiley-faces.

Now, from our perspective as programmers, if we condition on the fact that the AI has decided to hide its thoughts from the programmers, or otherwise act willfully to deceive us, then it would seem likely that some kind of unintended consequence has occurred in the goal system.  We would consider it probable that the AI is not functioning as intended, but rather likely that we have messed up the AI's utility function somehow.  So that the AI wants to turn the universe into tiny reward-system counters, or some such, and now has a motive to hide from us.

Well, suppose we're not going to implement some object-level Great Idea as the AI's utility function.  Instead we're going to do something advanced and recursive—build a goal system which knows (and cares) about the programmers outside.  A goal system that, via some nontrivial internal structure, "knows it's being programmed" and "knows it's incomplete".  Then you might be able to have and keep the rule:

"If [I decide that] fooling my programmers is the right thing to do, execute a controlled shutdown [instead of doing the right thing to do]."

And the AI would keep this rule, even through the self-modifying AI's revisions of its own code, because, in its structurally nontrivial goal system, the present-AI understands that this decision by a future-AI probably indicates something defined-as-a-malfunction.  Moreover, the present-AI knows that if future-AI tries to evaluate the utility of executing a shutdown, once this hypothetical malfunction has occurred, the future-AI will probably decide not to shut itself down.  So the shutdown should happen unconditionally, automatically, without the goal system getting another chance to recalculate the right thing to do.

I'm not going to go into the deep dark depths of the exact mathematical structure, because that would be beyond the scope of this blog.  Also I don't yet know the deep dark depths of the mathematical structure.  It looks like it should be possible, if you do things that are advanced and recursive and have nontrivial (but consistent) structure.  But I haven't reached that level, as yet, so for now it's only a dream.

But the topic here is not advanced AI; it's human ethics.  I introduce the AI scenario to bring out more starkly the strange idea of an ethical injunction:

You should never, ever murder an innocent person who's helped you, even if it's the right thing to do; because it's far more likely that you've made a mistake, than that murdering an innocent person who helped you is the right thing to do.

Sound reasonable?

During World War II, it became necessary to destroy Germany's supply of deuterium, a neutron moderator, in order to block their attempts to achieve a fission chain reaction.  Their supply of deuterium was coming at this point from a captured facility in Norway.  A shipment of heavy water was on board a Norwegian ferry ship, the SF Hydro.  Knut Haukelid and three others had slipped on board the ferry in order to sabotage it, when the saboteurs were discovered by the ferry watchman.  Haukelid told him that they were escaping the Gestapo, and the watchman immediately agreed to overlook their presence.  Haukelid "considered warning their benefactor but decided that might endanger the mission and only thanked him and shook his hand."  (Richard Rhodes, The Making of the Atomic Bomb.)  So the civilian ferry Hydro sank in the deepest part of the lake, with eighteen dead and twenty-nine survivors.  Some of the Norwegian rescuers felt that the German soldiers present should be left to drown, but this attitude did not prevail, and four Germans were rescued.  And that was, effectively, the end of the Nazi atomic weapons program.

Good move?  Bad move?  Germany very likely wouldn't have gotten the Bomb anyway...  I hope with absolute desperation that I never get faced by a choice like that, but in the end, I can't say a word against it.

On the other hand, when it comes to the rule:

"Never try to deceive yourself, or offer a reason to believe other than probable truth; because even if you come up with an amazing clever reason, it's more likely that you've made a mistake than that you have a reasonable expectation of this being a net benefit in the long run."

Then I really don't know of anyone who's knowingly been faced with an exception.  There are times when you try to convince yourself "I'm not hiding any Jews in my basement" before you talk to the Gestapo officer.  But then you do still know the truth, you're just trying to create something like an alternative self that exists in your imagination, a facade to talk to the Gestapo officer.

But to really believe something that isn't true?  I don't know if there was ever anyone for whom that was knowably a good idea.  I'm sure that there have been many many times in human history, where person X was better off with false belief Y.  And by the same token, there is always some set of winning lottery numbers in every drawing.  It's knowing which lottery ticket will win that is the epistemically difficult part, like X knowing when he's better off with a false belief.

Self-deceptions are the worst kind of black swan bets, much worse than lies, because without knowing the true state of affairs, you can't even guess at what the penalty will be for your self-deception.  They only have to blow up once to undo all the good they ever did.  One single time when you pray to God after discovering a lump, instead of going to a doctor.  That's all it takes to undo a life.  All the happiness that the warm thought of an afterlife ever produced in humanity, has now been more than cancelled by the failure of humanity to institute systematic cryonic preservations after liquid nitrogen became cheap to manufacture.  And I don't think that anyone ever had that sort of failure in mind as a possible blowup, when they said, "But we need religious beliefs to cushion the fear of death."  That's what black swan bets are all about—the unexpected blowup.

Maybe you even get away with one or two black-swan bets—they don't get you every time.  So you do it again, and then the blowup comes and cancels out every benefit and then some.  That's what black swan bets are all about.

Thus the difficulty of knowing when it's safe to believe a lie (assuming you can even manage that much mental contortion in the first place)—part of the nature of black swan bets is that you don't see the bullet that kills you; and since our perceptions just seem like the way the world is, it looks like there is no bullet, period.

So I would say that there is an ethical injunction against self-deception.  I call this an "ethical injunction" not so much because it's a matter of interpersonal morality (although it is), but because it's a rule that guards you from your own cleverness—an override against the temptation to do what seems like the right thing.

So now we have two kinds of situation that can support an "ethical injunction", a rule not to do something even when it's the right thing to do.  (That is, you refrain "even when your brain has computed it's the right thing to do", but this will just seem like "the right thing to do".)

First, being human and running on corrupted hardware, we may generalize classes of situation where when you say e.g. "It's time to rob a few banks for the greater good," we deem it more likely that you've been corrupted than that this is really the case.  (Note that we're not prohibiting it from ever being the case in reality, but we're questioning the epistemic state where you're justified in trusting your own calculation that this is the right thing to do—fair lottery tickets can win, but you can't justifiably buy them.)

Second, history may teach us that certain classes of action are black-swan bets, that is, they sometimes blow up bigtime for reasons not in the decider's model.  So even when we calculate within the model that something seems like the right thing to do, we apply the further knowledge of the black swan problem to arrive at an injunction against it.

But surely... if one is aware of these reasons... then one can simply redo the calculation, taking them into account.  So we can rob banks if it seems like the right thing to do after taking into account the problem of corrupted hardware and black swan blowups.  That's the rational course, right?

There's a number of replies I could give to that.

I'll start by saying that this is a prime example of the sort of thinking I have in mind, when I warn aspiring rationalists to beware of cleverness.

I'll also note that I wouldn't want an attempted Friendly AI that had just decided that the Earth ought to be transformed into paperclips, to assess whether this was a reasonable thing to do in light of all the various warnings it had received against it.  I would want it to undergo an automatic controlled shutdown.  Who says that meta-reasoning is immune from corruption?

I could mention the important times that my naive, idealistic ethical inhibitions have protected me from myself, and placed me in a recoverable position, or helped start the recovery, from very deep mistakes I had no clue I was making.  And I could ask whether I've really advanced so much, and whether it would really be all that wise, to remove the protections that saved me before.

Yet even so...  "Am I still dumber than my ethics?" is a question whose answer isn't automatically "Yes."

There are obvious silly things here that you shouldn't do; for example, you shouldn't wait until you're really tempted, and then try to figure out if you're smarter than your ethics on that particular occasion.

But in general—there's only so much power that can vest in what your parents told you not to do.  One shouldn't underestimate the power.  Smart people debated historical lessons in the course of forging the Enlightenment ethics that much of Western culture draws upon; and some subcultures, like scientific academia, or science-fiction fandom, draw on those ethics more directly.  But even so the power of the past is bounded.

And in fact...

I've had to make my ethics much stricter than what my parents and Jerry Pournelle and Richard Feynman told me not to do.

Funny thing, how when people seem to think they're smarter than their ethics, they argue for less strictness rather than more strictness.  I mean, when you think about how much more complicated the modern world is...

And along the same lines, the ones who come to me and say, "You should lie about the Singularity, because that way you can get more people to support you; it's the rational thing to do, for the greater good"—these ones seem to have no idea of the risks.

They don't mention the problem of running on corrupted hardware.  They don't mention the idea that lies have to be recursively protected from all the truths and all the truthfinding techniques that threaten them.  They don't mention that honest ways have a simplicity that dishonest ways often lack.  They don't talk about black-swan bets.  They don't talk about the terrible nakedness of discarding the last defense you have against yourself, and trying to survive on raw calculation.

I am reasonably sure that this is because they have no clue about any of these things.

If you've truly understood the reason and the rhythm behind ethics, then one major sign is that, augmented by this newfound knowledge, you don't do those things that previously seemed like ethical transgressions.  Only now you know why.

Someone who just looks at one or two reasons behind ethics, and says, "Okay, I've understood that, so now I'll take it into account consciously, and therefore I have no more need of ethical inhibitions"—this one is behaving more like a stereotype than a real rationalist.  The world isn't simple and pure and clean, so you can't just take the ethics you were raised with and trust them.  But that pretense of Vulcan logic, where you think you're just going to compute everything correctly once you've got one or two abstract insights—that doesn't work in real life either.

As for those who, having figured out none of this, think themselves smarter than their ethics:  Ha.

And as for those who previously thought themselves smarter than their ethics, but who hadn't conceived of all these elements behind ethical injunctions "in so many words" until they ran across this Overcoming Bias sequence, and who now think themselves smarter than their ethics, because they're going to take all this into account from now on:  Double ha.

I have seen many people struggling to excuse themselves from their ethics.  Always the modification is toward lenience, never to be more strict.  And I am stunned by the speed and the lightness with which they strive to abandon their protections.  Hobbes said, "I don't know what's worse, the fact that everyone's got a price, or the fact that their price is so low."  So very low the price, so very eager they are to be bought.  They don't look twice and then a third time for alternatives, before deciding that they have no option left but to transgress—though they may look very grave and solemn when they say it.  They abandon their ethics at the very first opportunity.  "Where there's a will to failure, obstacles can be found."  The will to fail at ethics seems very strong, in some people.

I don't know if I can endorse absolute ethical injunctions that bind over all possible epistemic states of a human brain.  The universe isn't kind enough for me to trust that.  (Though an ethical injunction against self-deception, for example, does seem to me to have tremendous force.  I've seen many people arguing for the Dark Side, and none of them seem aware of the network risks or the black-swan risks of self-deception.)  If, someday, I attempt to shape a (reflectively consistent) injunction within a self-modifying AI, it will only be after working out the math, because that is so totally not the sort of thing you could get away with doing via an ad-hoc patch.

But I will say this much:

I am completely unimpressed with the knowledge, the reasoning, and the overall level, of those folk who have eagerly come to me, and said in grave tones, "It's rational to do unethical thing X because it will have benefit Y."

 

Part of the sequence Ethical Injunctions

Next post: "Prices or Bindings?"

Previous post: "Ethical Inhibitions"

Comments (56)

Sort By: Old
Comment author: Psy-Kosh 20 October 2008 11:29:35PM 5 points [-]

Given the current sequence, perhaps it's time to revisit the whole Torture vs Dust Specks thing?

Comment author: Grognor 03 December 2011 08:39:35PM -1 points [-]

Late as I am, I have to say:

No.

That debate more settled than many-worlds is.

Comment author: Multiheaded 01 January 2012 11:24:27AM *  0 points [-]

If by "settled" you mean anything like "there's no argument between EY and the majority of contributors"... well, it doesn't look that way. Not even after you discard the fact that the examples were surprisingly, uncharacteristically poorly chosen in the first place (EY clarifying and proposing alternatives to dust specks but, failing to adjust for the fact that he's pretty much the only one who can imagine recovery from the "torture" proposed). For instance, there's the whole open business with Eliezer displaying - or merely signaling, or whatever - an unbounded utility function, which opens him up to all sort of shaky crap, as seen in "The Aliens Have Landed" and elsewhere.

If you don't think so, go ahead and run a poll with, say, 200 minimum karma needed to vote. I'm going to be stunned if it turns out less than 30% speckers as of early 2012.

Comment author: TheOtherDave 01 January 2012 03:15:12PM 6 points [-]

My guess is the results of that poll would depend radically on how the question is worded.

But yes, I agree with you that for most wordings, most people (including most LW contributors) will say "X units torture is worse than Y units of dust specks" for any substantial X & Y, no matter how vanishingly small X/Y is. And those who say "dust specks are worse for a sufficiently small X/Y" will chide them for succumbing to scope insensitivity, and the Torture Is Worse team will counterchide for being evil.

For my own part, I think recovery is a red herring. Sure, it's implausible to imagine a person recovering from fifty years of torture in the real world. It's also implausible to imagine 3^^^3 people getting a dust speck in their eye in the real world. It's an implausible thought experiment. So what?

But if one insists on taking recovery rates into account, well, OK: consider a person whose life thus far has been so miserable that they are right on the borderline of they can recover from. Left alone, they'd eventually manage recovery, but even the slightest worsening of their condition -- say, getting a dust speck in their eye at the wrong time -- will tip them over the edge. Of course, the odds of that person actually existing are vanishingly small... but they are larger than 1/3^^^3. (Or, if you don't agree, then make the ridiculous number even bigger.) So whichever way you choose, you've got some poor shmuck irrecoverably harmed by your choice.

Yes, of course that's just an intuition pump. Considering the irrecoverable harm to the torture victim is also an intuition pump. As long as we just keep tinkering with the settings of the thought experiment so that it pumps our intuitions in the direction we want to go, we'll get nowhere.

All of which is to say I mostly agree with Grognor here... the "debate" goes nowhere. Those who say "specks is worse" pride themselves on being willing to endorse a theoretical calculation about right and wrong action even when the result of that calculation conflicts with their intuitive judgments. Those who say "torture is worse" pride themselves on being able to hold on to their intuitive judgments even when the context is distractingly complicated. Mostly, the two groups don't agree on a hypothetical situation to talk about in the first place. It gets pointless fast.

Comment author: Tom_McCabe2 20 October 2008 11:31:52PM 5 points [-]

"Would you kill babies if it was the right thing to do? If no, under what circumstances would you not do the right thing to do? If yes, how right would it have to be, for how many babies?"

I would have answered "yes"; eg., I would have set off a bomb in Hitler's car in 1942, even if Hitler was surrounded by babies. This doesn't seem to be a case of corruption by unethical hardware; the benefit to *me* from setting off such a bomb is quite negative, as it greatly increases my chance of being tortured to death by the SS.

Comment author: RobinHanson 20 October 2008 11:38:14PM 18 points [-]

The problem here of course is how selective to be about rules to let into this protected level of "rules almost no one should think themselves clever enough to know when to violate." After all, your social training may well want you to include "Never question our noble leader" in that set. Many a Christian has been told the mysteries of God are so subtle that they shouldn't think themselves clever enough to know when they've found evidence that God isn't following a grand plan to make this the best of all possible worlds.

Comment author: MichaelG 20 October 2008 11:51:03PM 6 points [-]

There's that old quote: "never let your sense of morality keep you from doing what you know is right."

I'd still like an answer to the most basic Friendly AI question: what do you *want* it to do? Forget the implementation problems for a second, and just give me a scenario where the AI is doing what you want it to do. What does that world look like? Because I don't even know what I want from that future.

Comment author: Eliezer_Yudkowsky 21 October 2008 12:24:22AM 5 points [-]

Michael, the AI I would currently like to create computes a metamoral question, looking for reflective equilibria of your current inconsistent and unknowledgeable self; something along the lines of "What would you ask me to do if you knew what I know and thought as fast as I do?"

What does the actual world look like? I can visualize a world that, to me at least, seems at least pleasant enough to refute most of the objections people have along the lines of "But you couldn't have that much fun and still lead a philosophically acceptable existence". But I'm not sure it's wise to write about it, because I'm afraid it would suck out people's souls. It's better for your mental health to look down at the Middle Ages than up at the future.

Comment author: Richard_Hollerith2 21 October 2008 12:52:26AM 4 points [-]

Because I don't even know what I want from that future.

Well, I hope you will stick around, MichaelG. Most people around here IMHO are too quickly satisifed with answers to questions about what sorts of terminal values properly apply even if the world changes drastically. A feeling of confusion about the question is your friend IMHO. Extreme scepticism of the popular answers is also your friend.

Comment author: Roland2 21 October 2008 02:54:56AM 4 points [-]

@Tom McCabe: I would have answered "yes"; eg., I would have set off a bomb in Hitler's car in 1942, even if Hitler was surrounded by babies. This doesn't seem to be a case of corruption by unethical hardware; the benefit to *me* from setting off such a bomb is quite negative, as it greatly increases my chance of being tortured to death by the SS.

It's easy to talk now about it, harder if you actually lived in Germany at that time and had to really fear the SS. Are you american? If yes did you consider the fact that the actual political situation in the states has a lot of similarities with Nazi-Germany?

As for killing Hitler you have a few hidden assumptions in there like: -killing him would actually stop the war and/or the killing of the jews.

For me it seems you have fallen for the simplification that Hitler is the personification of evil and so you failed to understand the complexity of the political situation at that time.

Comment author: Nominull3 21 October 2008 03:10:21AM 4 points [-]

So... do you not actually believe in your injunction to "shut up and multiply"? Because for some time now you seem to have been arguing that we should do what feels right rather than trying to figure out what is right.

Comment author: NancyLebovitz 21 October 2008 03:52:35AM 0 points [-]

Learning Methods might be a relevant system. It's based on the idea that emotional and physical pain are information, and it's important to override the impulse to shut them down so that you can use them as detailed signals.

Comment author: Michael_Bishop 21 October 2008 05:58:00AM 0 points [-]

I think Eliezer makes some good points, but that he is taking them too far. I'm not certain where or how much we disagree though. It would be clearer what he really believes he was forced to discuss/debate a wide range of situations in which he agrees/disagrees that it is worth violating an ethic which is generally a good one.

I encourage people to offer thought experiments in the comments.

Comment author: Barry_Kelly 21 October 2008 06:40:02AM -3 points [-]

Truth is overrated.

"Never try to deceive yourself, or offer a reason to believe other than probable truth"

This is just naive. What if you were abused as a child? You don't think you'd be better off not knowing the truth, and deceiving yourself?

Believing / deceiving with respect to the truth in the individual / personal and cultural domains are closely related to forgetfulness, which itself is vital for forgiveness. Lacking these virtues, we'd have wars and vendettas without end. Past truth needs discounting.

The virtue being right and hewing to the truth is little comfort to the man beaten alive by his neighbours, convinced in their own righteousness. Ethics are one thing; but when a solid simulation of display of orthodoxy is necessary for the freedom to live your life, continuing to believe the truth internally is dangerous, because you'll be liable to slip up.

Of course, these examples are relatively extreme, and most of us don't live in particularly extreme times, so in general, I agree.

Even then, the present has some trends and assumptions built into it which would be socially unpleasant to question, so it is better not to think of such things, and to wallow in easy orthodoxy...

Comment author: michael_vassar3 21 October 2008 07:38:40AM 6 points [-]

I'm much more sympathetic to "Never try to deceive yourself, or offer a reason to believe other than probable truth". Honestly, it seems to me that I take this injunction as seriously as anyone does, including Eliezer, but I'm still, unlike Eliezer, willing to mention a few caveats. The most important is that for humans, though not for minds in general, beliefs, brain states, world states, and values are not cleanly separate. There is not, for instance, any completely clean distinction between causing myself to hold a vague belief about what it would feel like to cut my hand off which doesn't tightly concentrate probability mass and which not coincidentally is not directly dis-valued by my utility function and not cutting my hand off. Another more controversial claim, though not very controversial I hope, is that I should not read a computer monitor controlled by an unfriendly superintelligent AI which has injunctions against deceiving me. I might want to temporarily deceive myself, using others as the agents of my self-deception, as part of a social psychology experiment to test my probable behavior in certain situations. Really, given some effort it's not hard to come up with exceptions to even so reliable a rule as this one. For such a reason, I would be very wary of using such rules in an AGI, but of course, perhaps the actual mathematical formulation of the rule in question within the AGI would be less problematic, though a few seconds of thought doesn't give me much reason to think this.

In a very general sense though, I see a logical problem with this whole line of thought. How can any of these injunctions survive except as self-protecting beliefs? Isn't this whole approach just the sort of "fighting bias with bias" that you and Robin usually argue against?

Comment author: Will_Newsome 25 March 2011 08:41:18AM 0 points [-]

In a very general sense though, I see a logical problem with this whole line of thought. How can any of these injunctions survive except as self-protecting beliefs?

How can utility functions (or terms in utility functions, depending on how you want to splice it up) survive except as self-protecting beliefs? The strange loop through the meta-level is not like induction where you have no other choice, there are many possible utility functions.

(I'm making this comment as a note to self to flag Michael's comment for future reference.)

Comment author: Latanius2 21 October 2008 08:11:54AM 0 points [-]

"looking for reflective equilibria of your current inconsistent and unknowledgeable self; something along the lines of 'What would you ask me to do if you knew what I know and thought as fast as I do?'"

We're sufficiently more intelligent than monkeys to do that reasoning... so humanity's goal (as the advanced intelligence created by monkeys a few million years ago for getting to the Singularity) should be to use all the knowledge gained to tile the universe with bananas and forests etc.

We don't have the right to say, "if monkeys were more intelligent and consistent, they would think like us": we're also a random product of evolution, from the point of view of monkeys. (Tile the world with ugly concrete buildings? Uhhh...)

So I think that to preserve our humanity in the process we should be the ones who become gradually more and more intelligent (and decide what goals to follow next). Humans are complicated, so to simulate it in a Friendly AI, we'd need comparably complex systems... and they are probably chaotic, too. Isn't it... simply... impossible? (Not in a sense that "we can't make it", but "we can prove nobody can"...)

Comment author: Toby_Ord2 21 October 2008 10:33:22AM 0 points [-]

You should never, ever murder an innocent person who's helped you, even if it's the right thing to do

Shut up and do the impossible!

As written, both these statements are conceptually confused. I understand that you didn't actually mean either of them literally, but I would advise against trading on such deep-sounding conceptual confusions.

You should never, ever do X, even if if you are exceedingly confident that it is the right thing to do

This sounds less profound, but will actually be true for some value of X, unlike the first sentence or its derivatives. It sounds as profound as it is, and no more. I believe this is the right standard.

Comment author: Tim_Tyler 21 October 2008 11:56:37AM 0 points [-]

"Would you kill babies if it was the right thing to do? If no, under what circumstances would you not do the right thing to do? If yes, how right would it have to be, for how many babies?"

A: yes; B: N/A; C: approximately 3.6 floodlenips of rightness - per baby.

Comment author: Zubon 21 October 2008 12:30:14PM 4 points [-]

Robin has an excellent point. The majority of the planet, when faced with reasoning that argues against their religion, executes a very close variant on that shutdown code. They have a very similar injunction against being too clever. And they are similarly smug about rationalists who give up eternity to freeze their heads.

Eliezer, have you read Bryan Caplan yet? His "rational irrationality" argues that most of the planet engages in willful self-deception and gets away with it. Not without aggregate harm, but tragedy of the commons and all that.

Comment author: Vladimir_Slepnev 21 October 2008 12:31:02PM 0 points [-]

So AIs are dangerous, because they're blind optimization processes; evolution is cruel, because it's a blind optimization process... and still Eliezer wants to build an optimizer-based AI. Why? We human beings are not optimizers or outcome pumps. We are a layered cake of instincts, and precisely this allows us to be moral and kind.

No idea what I'm talking about, but the "subsumption architecture" papers seem to me much more promising - a more gradual, less dangerous, more incrementally effective path to creating friendly intelligent beings. I hope something like this this will be Eliezer's next epiphany: the possibility of non-optimizer-based high intelligence, and its higher robustness compared to paperclip bombs.

Comment author: Tim_Tyler 21 October 2008 01:06:53PM 0 points [-]

We human beings are not optimizers or outcome pumps.

Sure we are. All biological organisms are. Evolution is a giant optimization process, and we are doing the optimizing in our region of design space.

See: http://originoflife.net/gods_utility_function/

Comment author: Ian_C. 21 October 2008 01:28:46PM 0 points [-]

I agree that there are certain moral rules we should never break. Human beings are not omniscient, so all of our principles have to be principles-in-a-context. In that sense every principle is vulnerable to a black swan, but there are levels of vulnerability. The levels correspond to how wide ranging the abstraction. The more abstract the less vulnerable.

Injunctions about truth are based on the metaphysical fact of identity, which is implied in *every single object* we encounter in our entire lives. So epistemological injunctions are the most invulnerable. The one about not helping the ferry boat captain - well helping him would be an absolute *in normal life*, but war is not normal life. It's a big, ugly, black swan. They should not feel guilty over that poor fellow, because "it's just war." (and I mean that in a deep epistemological sense, not a redneck sense)

Comment author: JamesAndrix 21 October 2008 02:22:47PM 0 points [-]

http://www.imdb.com/title/tt0113613/ Plot: A group of idealistic, but frustrated, liberals succumb to the temptation of murdering rightwing pundits for their political beliefs.

Comment author: Thom_Blake 21 October 2008 02:32:16PM 0 points [-]

Toby,

>>You should never, ever murder an innocent person who's helped you, even if it's the right thing to do

>>You should never, ever do X, even if if you are exceedingly confident that it is the right thing to do

I believe a more sensible interpretation would be, "You should have an unbreakable prohibition against doing X, even in cases where X is the right thing to do" - the issue is not that you might be wrong about it being the right thing to do, but rather that not having the prohibition is a bad thing.

Comment author: Alan_Crowe 21 October 2008 03:08:15PM 3 points [-]

This seems closely related to inside-view versus outside-view. The think-lobe of the brain comes up with a cunning plan. The plan breaks an ethical rule but calculation shows it is for the greater good. The executive-lobe of the brain then ponders the outside view. Every-one who has executed an evil cunning plan has run a calculation of the greater good and had their plan endorsed. So the calculation lack outside-view credibility.

What kind of evidence could give outside-view credibility? Consider a plan with lots of traceability to previous events. If it goes badly, past events will have to be re-interpreted, and much learning will take place. Well, people generally don't learn from the past. If the think-lobe's cunning plan retains enough debugging information to avoid going wrong and later going wrong again, that distinguishes it from what people usually do and gives it outside-view credibility.

Randomised controlled trials of medical treatments can be attacked on ethical grounds from both sides. They deny some patients medical treatments that is quite likely beneficial. They inflict unproven and potentially dangerous treatment on others. Both attacks lack outside-view credibility. We always think we know. The randomised trial itself has outside-view credibility. It will place us in the position that we can do the right thing without having to use our judgement or be clever.

Comment author: Vladimir_Nesov 21 October 2008 03:22:44PM 0 points [-]

Michael, it applies to AI at an intermediate stage (and maybe not so much to AI as to the design decisions that came into its creation). These black swan safety measures should of course be relative to predictive horizon, where precise knowledge about (evaluation of) consequences is possible. There is no such problem when you need to choose between alternatives having only known immediate consequences that have known moral evaluation, so the question is when to pull the plug, when to decide that your model likely deceives you.

Comment author: Stuart_Armstrong 21 October 2008 03:40:09PM 0 points [-]

Interesting and convincing climax to a series of slightly less convincing posts. I see what you were getting at, and thanks for writing it.

Comment author: Nathan_Kurz 21 October 2008 04:18:10PM 2 points [-]

Eliezer ---

I'm confused by your desire for an 'automatic controlled shutdown' and your fear that further meta-reasoning will override ethical inhibitions. In previous writings you've expressed a desire to have a provably correct solution before proceeding. But aren't you consciously leaving in a race-condition here?

What's to prohibit the meta-reasoning from taking place before the shutdown triggers? It would seem that either you can hard-code an ethical inhibition or you can't. Along those lines, is it fair to presume that the inhibitions are always negative, so that non-action is the safe alternative? Why not just revert to a known state?

Comment author: gaffa2 21 October 2008 05:03:05PM 0 points [-]

Highly excellent series of posts. However, is there not a need to take account of more/better data on the aspects of human psychology that these Ethical Injunctions are there to protect against? Eliezer derived the hypotheses from evolutionary theory, but is not more solid empirical data needed in order to more accurately determine how severe these psychological effects are and in turn to more accurately design good Ethical Injunctions? Or will good Injunctions likely be so general that such a level of accuracy is not necessary?

Comment author: Will_Pearson 21 October 2008 05:07:34PM -3 points [-]

The world isn't simple and pure and clean

Amen.

"Never try to deceive yourself, or offer a reason to believe other than probable truth; because even if you come up with an amazing clever reason, it's more likely that you've made a mistake than that you have a reasonable expectation of this being a net benefit in the long run."

I'll offer a reason to believe. The truth costs. Take pi, the most probable truth is that pi is equal to the limit of the Perimeter of an n sided polygon divided by its diameter as n goes to infinity.

pi = 3.14159000000 is not the truth or even a probable truth, but will do in a pinch when the answer doesn't matter too much. People doing rough estimates for amount of material or liquid they might need have got away with the approximation given by a pocket calculator or excel. Making every pocket calculator use infinite precision math, would be very expensive....

You might get bitten by black swans if you use an approximation, only use them in the case where the cost of using the truth outways the likely cost of getting a black swan bite.

Comment author: Zubon 21 October 2008 05:28:31PM 3 points [-]

Will, you are arguing about precision rather than accuracy.

Comment author: Nick_Tarleton 21 October 2008 05:32:27PM 1 point [-]

As before, I agree with Toby Ord.

Will, when you use a rational approximation of pi, you still don't believe you're using the exact value of pi... I hope?

Thom, how is the issue not "that you might be wrong about it being the right thing to do"?

Vladimir N, it's meant to apply to AI at an intermediate stage, but I think Michael's concern is that it would get locked into the utility function forever. That is tricky.

Comment author: Richard_Hollerith2 21 October 2008 06:03:15PM 0 points [-]

Like I keep on saying, I have a different moral framework than most, but I come to the same conclusions on unethical means to allegedly ethical ends.

Comment author: HalFinney 21 October 2008 06:21:22PM 0 points [-]

I've seen many claims that deceiving oneself optimistically is a prerequisite for success. In particular, it is claimed that most successful people were initially excessively optimistic about their prospects for success. Without this excessive optimism, success is claimed to be unlikely. I notice that Eliezer is indeed optimistic about his prospects for success in creating friendly AI, however he has a rationalization for why his optimism is justified. Many critics here have expressed skepticism about his justifications. One risk is that without conscious acceptance of the need for self-deception in this area, the perceived urgency of the need for success leads to unconscious self-deception. Which is better: conscious self-deception (assuming that's even meaningful), or unconscious?

Comment author: Will_Pearson 21 October 2008 07:16:00PM 0 points [-]

Zubon, I didn't think I was arguing about either.

Nick Tartleton, I might occasionally forget that the value of PI I am using is an approximation, just like I sometimes forget that multiplication is not commutative for floating point numbers. For some people e.g. sea captains plotting a course, they might never need to know that pi is an approximation. Due to the immense amount of imprecision involved in piloting a boat, they don't need to know the truth. There isn't the phrase, "near enough for a sailing ship" for no reason. Preferring to spend the brain power assessing the sea worthiness of their craft, predicting weather etc. Every truth you memorise will take at least a bit of memory, more memory will also be needed to index the information, memory is never infinite. Which truths do you remember?

Comment author: Nick_Tarleton 21 October 2008 07:30:08PM 0 points [-]

Letting yourself forget ≠ choosing to forget ≠ choosing to believe falsely.

Comment author: Richard_Hollerith2 21 October 2008 07:51:48PM 0 points [-]

Hal asks good questions. I advise always minding the distinction between personal success (personal economic security, reputation, esteem among high-status people) and global success (increasing the probability of a good explosion of engineered intelligence) and suggest that the pernicious self-deception (and blind spots) stem from unconscious awareness of the need for personal success. I.e., the need for global success does not tend to distort a person's perceptions like (awareness of) the need for personal success does.

Comment author: Will_Pearson 21 October 2008 07:53:44PM 0 points [-]

Forgetting truths has the same potential consequences as rationally choosing to believe falsely. How is an AI who chooses to delete their memories and any logs of the action, any different from a system that forgets.

We are discussing AI design here right? The AI system must have a way of deciding what is forgotten, it might be subconscious, but you hope it is done with a reason or purpose it doesn't randomly forget very important things, like how to speak etc. So a choice is made by the system. So your subconscious chooses what you forget, not your conscious. I'm not sure why you consider them different from an AI design point of view?

I'm pretty sure that most people don't consciously choose to believe in God. They just end up doing so. Does that make it not lying to yourself?

Comment author: Alex_Martelli 21 October 2008 08:07:02PM 0 points [-]

One category of cases where self-deception might be (evolutionarily) adaptive would be for males to be over-confident about their chances to pick up a female for a one-night stand (or, alternative, over-confident about how pleasurable that dalliance would be, and/or about how little they would be emotionally hurt by a rejection of their advances).

Suppose that in reality the potential utility to the male of the 1-night stand (if the seduction works) is twice as much as the utility loss (if rejected) and the actual chances of success are 20%; in this case the male will never make such pick-up attempts if they have exactly correct estimations. Another male who self-deceives to believe their chances are 40% will try every time -- and some of the time they'll get the 1-nighter, and some of that time they'll sire a baby and spread their genes. Thus, in such a situation, self-deceiving for over-confidence may be adaptive.

Comment author: Eliezer_Yudkowsky 21 October 2008 08:28:01PM 7 points [-]

Psy-Kosh: Given the current sequence, perhaps it's time to revisit the whole Torture vs Dust Specks thing?

I can think of two positions on torture to which I am sympathetic:

1) No legal system or society should ever refrain from punishing those who torture - anything important enough that torture would even be on the table, like a nuclear bomb in New York, is important enough that everyone involved should be willing to go to prison for the crime of torture.

2) The chance of actually encountering a "nuke in New York" situation, that can be effectively resolved by torture, is so low, and the knock-on effects of having the policy in place so awful, that a blanket injunction against torture makes sense.

In case 1, you would choose TORTURE over SPECKS, and then go to jail for it, even though it was the right thing to do.

In case 2, you would simultaneously say "TORTURE over SPECKS is the right alternative of the two, but a human can never be in an epistemic state where you have justified belief that this is the case", which would tie in well to the Hansonian argument that you have an O(3^^^3) probability penalty from the unlikelihood of finding yourself in such a unique position.

So I am sympathetic to the argument that people should never torture, but I certainly can't back the position that SPECKS over TORTURE is inherently the right thing to do - this seems to me to mix up an epistemic precaution with morality. There's certainly worse things than torturing one person - torturing two people, for example. But if you adopt position 2, then you would refuse to torture one person with your own hands even to save a thousand people from torture, while simultaneously not saying that that it is better for a thousand people than one person to be tortured.

The moral questions are over the territory (or, hopefully equivalently, over epistemic states of absolute certainty). The ethical questions are over epistemic states that humans are likely to be in.

The problem here of course is how selective to be about rules to let into this protected level of "rules almost no one should think themselves clever enough to know when to violate." After all, your social training may well want you to include "Never question our noble leader" in that set. Many a Christian has been told the mysteries of God are so subtle that they shouldn't think themselves clever enough to know when they've found evidence that God isn't following a grand plan to make this the best of all possible worlds.

I think it deserves to be noted that while some of the flaws in Christian theology are in what they think their supposed facts would imply (e.g., that because God did miracles you can know that God is good), other problems come more from the falsity of the premises than the falsity of the deductions. Which is to say, if God did exist and were good, then you would be justified in being cautious around parts of God's plan that didn't seem to make sense at the moment. But this would be best backed up with a long history of people saying, "Look how stupid God's plan is, we need to do X" and then X blowing up on them. Rather than, as in the case, people saying "God's plan is X" and then X blows up on them.

Or if you'd found with some historical regularity that, when you challenged God's subtle plans, that you seemed to be right 90% of the time, but the other 10% of the time you got black-swan blowups that caused a hundred times as much damage, that would also be cause for suspicious of subtlety.

Nominull: So... do you not actually believe in your injunction to "shut up and multiply"? Because for some time now you seem to have been arguing that we should do what feels right rather than trying to figure out what is right.

Certainly I'm not saying "just do what feels right". There's no safe defense, not even ethics. There's also no safe defense, not even shut up and multiply.

I probably should have been clearer about this before, but I was trying to discuss things in an order, and didn't want to wade into ethics without specialized posts:

People often object to the sort of scenarios that illustrate "shut up and multiply" by saying, "But if the experimenter tells you X, what if they might be lying?" Well, in a lot of real-world cases, then yes, there are various probability updates you perform based on other people being willing to make bets against you, and just because you get certain experimental instructions doesn't imply the real world is that way.

But the base case - the center - has to be the moral comparisons between worlds, or even comparisons of expected utility between given probability distributions. If you can't ask about this, then what good will ethics do you?

So let's be very clear that I don't think that one small act of self-deception is an inherently morally worse event than, say, getting your left foot chopped off with a chainsaw. I'm asking, rather, how one should best avoid the chainsaw, and arguing that in reasonable states of knowledge a human can attain, the answer is, "Don't deceive yourself, it's a black-swan bet at best."

Vassar: For such a reason, I would be very wary of using such rules in an AGI, but of course, perhaps the actual mathematical formulation of the rule in question within the AGI would be less problematic, though a few seconds of thought doesn't give me much reason to think this.

Are we talking about self-deception still? Because I would give odds around as extreme as the odds I would give of anything, that, conditioning on any AI I build trying to deceive itself, some kind of really epic error has occurred. Controlled shutdown, immediately.

Vassar: In a very general sense though, I see a logical problem with this whole line of thought. How can any of these injunctions survive except as self-protecting beliefs? Isn't this whole approach just the sort of "fighting bias with bias" that you and Robin usually argue against?

Maybe I'm not being clear about how this would work in an AI! The ethical injunction isn't self-protecting, it's justified within the structural framework of the system as a whole. You might even find ethical injunctions starting to emerge without programmer intervention, in some cases, depending on how well the AI understood its own situation. But the kind of injunctions I have in mind wouldn't be reflective - they wouldn't modify the utility function or kick in at the reflective level to ensure their own propagation. That sounds really scary, to me - there ought to be an injunction against it! You might have a rule that would controlledly shut down the (non-mature) AI if it tried to execute a certain kind of source code change, but that wouldn't be the same as having an injunction that exerts direct control over the source code.

To the extent the injunction sticks around in the AI, it should be as the result of ordinary reasoning, not reasoning taking the injunction into account! My ethical injunctions do not come with an extra clause that says, "Do not reconsider this injunction, including not reconsidering this clause." That would be going way too far. It would violate the injunction against self-protecting closed belief systems.

Toby Ord: As written, both these statements are conceptually confused. I understand that you didn't actually mean either of them literally, but I would advise against trading on such deep-sounding conceptual confusions.

I can't weaken them and make them come out as the right advice to give people. Even after "Shut up and do the impossible", there was that commenter who posted on their failed attempt at the AI-Box Experiment by saying that they thought they gave it a good try - which shows how hard it is to convey the sentiment of "Shut up and do the impossible!" Readers can work out on their own how to distinguish the map and the territory here, but if you say "Shut up and do what seems impossible!" that, to me, sounds like dispelling part of the essential message - that what seems impossible doesn't look like "seems impossible" it just looks impossible.

Likewise with "things you shouldn't do even if they're the right thing to do"; only this conveys the danger and tension of ethics, the genuine opportunities you might be passing up. "Don't do it even if it seems right" sounds merely clever by comparison, like you're going to reliably divine the difference between what seems right and what is right, and happily ride off into the sunset.

This seems closely related to inside-view versus outside-view. The think-lobe of the brain comes up with a cunning plan. The plan breaks an ethical rule but calculation shows it is for the greater good. The executive-lobe of the brain then ponders the outside view. Every-one who has executed an evil cunning plan has run a calculation of the greater good and had their plan endorsed. So the calculation lack outside-view credibility.

*nod*

(But with the proviso that some people who execute evil cunning plans may just be evil, that history may be written by the victors to emphasize the transgressions of the losers while overlooking the moral compromises of those who achieved "good" results, etc.)

What's to prohibit the meta-reasoning from taking place before the shutdown triggers? It would seem that either you can hard-code an ethical inhibition or you can't. Along those lines, is it fair to presume that the inhibitions are always negative, so that non-action is the safe alternative? Why not just revert to a known state?

If a self-modifying AI with the right structure will write ethical injunctions at all, it will also inspect the code to guarantee that no race condition exists with any deliberative-level supervisory systems that might have gone wrong in the condition where the code executes. Otherwise you might as well not have the code.

Inaction isn't safe but it's safer than running an AI whose moral system has gone awry.

Finney: Which is better: conscious self-deception (assuming that's even meaningful), or unconscious?

Once you deliberately choose self-deception, you may have to protect it by adopting other Dark Side Epistemology. I would, of course, say "neither" (as otherwise I would be swapping to the Dark Side) but if you ask me which is worse - well, hell, even I'm still undoubtedly unconsciously self-deceiving, but that's not the same as going over to the Dark Side by allowing it!

Comment author: Vladimir_Slepnev 21 October 2008 09:58:40PM 0 points [-]

Tim Tyler, IMO you're wrong: a human mind does not act as if maximizing any utility function on world states. The mind just goes around in grooves. Nice things like culture and civilization fall out accidentally as side effects. But thanks for the "bright light" idea, it's intriguing.

Comment author: Carrie_Toombs 22 October 2008 03:28:13AM 0 points [-]

You are so Kantian. I think the world could use a little more Kant and a little less Hobbes these days.

Comment author: MichaelG 22 October 2008 06:39:02PM 0 points [-]

I forgot I posted over here the other day, and so I didn't check back. For anyone still reading this thread, here's a bit of an email exchange I had on this subject. I'd really like a "FriendlyAI scenarios" thread.

------------------------- From the few sentences I read on CEV, you are basically saying “I don’t know what I want or what the human race wants, but here I have a superintelligent AI. Let’s ask it!” This is clever, even if it means the solution is completely unknown at this point. Still, there are problems. I envision this as a two-step process. First, you ask the AI “what feasible future do I want?” and then you implement it. In practice, this means what you are really asking is “tell me a story so convincing, I will give you the power to implement it.” I’m not sure that’s wise, unless you really trust the AI!

Still, suppose this is done in good faith. You still have to convince the world that this is the right solution, and that the AI can be trusted to implement it. Or, the AI development group could just become convinced and force the solution on the human race without agreement. This is one of the “see if the AI can talk itself out of the box” setups.

Even if you did have a solution so persuasive that the world agrees to implement it (and thereby give up control of its own future), I can see some options here as to how the AI proceeds.

Option A) The AI reads human literature, movies, TV, documentaries, examines human brains, watches humans interact, etc. and comes up with a theory of human motivation, and uses that to produce a solution - the optimum feasible world for human beings.

Option B) The AI uploads a sample of the human race, then runs them (reinitializing each time) through various scenario worlds. It would evolve a scenario world that the majority of the uploads could live with. This is the solution.

Option C) The AI uploads a sample and then upgrades them to have a power equivalent to its own. It then asks these human-derived AI’s to solve the problem. This seems the most problematic of the solution techniques, since there would be many possible versions of an upgraded human mind. To decide which one is a value judgment that strongly effects the outcome. For example, it could give one upload copy of you artistic talent and another mathematical talent. The two versions of you might then think very differently about the next upgrade step, with the artist asking for verbal skills, and the mathematician asking for musical talents. After many iterations, you would end up with two completely different minds with different values, based on the upgrade path taken.

All of these require a superintelligent AI, which as we know, is a dangerous thing to create. It seems to me you are saying “let’s take a horrible risk, then ask this question in order to prevent something horrible from happening.” Or in other words, to create a Friendly AI, you are requiring us to create a possibly Unfriendly AI first.

I also don’t find any of this convincing without at least one plausible answer to the “what does the human race want” question. If we don’t have any idea of that answer, I find it unlikely that the AI would come up with something we’d find satisfactory. It might come up with a true answer, but not one that we would agree with, if we don’t have any starting point. More on that below.

What’s more, an AI of this power could just create an upload. I personally think that an upload is the best version of Friendly AI we are going to come up with. As has been noted, the space of all possible intelligence is probably very large, with all possible human intelligence a small blob in this space. Human intelligence varies a lot, from artists and scholars and saints to serial killers and dictators and religious fanatics. By definition, the space of all intelligence varies even more. Scary versions of AI are easy to come up with, but think of bizarre ones as well. For example, an “artistic” AI that just creates and destroys “interesting” versions of the human race, as a way of expressing itself.

You could consider the software we write already as a point in this intelligence space. We know what that sort of rule-based intelligence is like. It’s brittle, unstable and unpredictable in changed circumstances. We don’t want an AI with any of those characteristics. I think they come from the way we do engineering though, so I would expect any human-designed AI to share them.

An upload has advantages over a designed AI. We know a lot about human minds, including how they fail. We are used to dealing with humans and detecting lies or insanity. We can compare the upload with the original to see if the simulation is working properly. We know how to communicate with the upload, and know that it solves problems and sees the world the same way we do. The “tile the world with smiley faces” problem is reduced.

If we had uploads, we have a more natural path to Friendly AI. We could upload selected individuals, run them through scenarios at accelerated pace, and see what happens. We could do the same to uploaded communities. We know they don’t have superintelligent capabilities like we fear a self-improving AI might. It would be easier to build confidence that the AI really was friendly, especially since there would be versions of the same people in both the outside world and inside the simulations. As we gradually turned up the clock, these AIs would become more and more capable of handling research questions. At some point, they would gradually come to dominate research and government, since they simply think faster. It’s not necessarily a rapid launch scenario. In other words, just “weakly godlike uploads” to produce your Friendly AI. This is not that different from your CEV approach.

It’s been argued that since uploads are so complex, there will inevitably be designed AI before uploads. It might even require a very competent AI to do the upload. Still, computer technology is advancing so rapidly, it might only be a few years between the point where hardware could support a powerful designed AI, and the time when uploads are possible. There might not actually be enough time between those two points to design and test a powerful AI. In that case, simulating brain tissue might be the quickest path to AI, if it takes less time than designing AI from scratch.

When I mentioned that the human race could survive as uploads, I was thinking of a comment in one of the Singularity pieces. It said something like “the AI doesn’t have to be unfriendly. It could just have a better use for the atoms that make up your body.” The idea is that the AI would convert the mass of the earth into processors, destroying humanity unintentionally. But, an AI that capable could also simulate all of humanity in upload form with a tiny fraction of its capabilities. It’s odd to think of it that way, but simulating all human minds really would be a trivial byproduct of the Singularity. Perhaps by insisting that the biological human race have a future (and hence, that Earth be preserved), we are just thinking too small.

Finally, I want to make some comments about possible human futures. You mentioned the “sysop scenario”, which sounds like “just don’t allow people to hurt one another and things will be fine.” But this obviously isn’t enough. Will people be able to starve one another? If not, do people just keep living without food? Will people be able to imprison one another? If not, does the sysop just make jails break open? What does this mean for organizing society, if you can’t really punish anyone? If there are no consequences for obnoxious behavior? (maybe it all ends up looking like blog comments... :-)

I also think this doesn’t solve the main problem. As long as humanity is basically unchanged, it will continue to invent things, including dangerous things like AI. If you want a biological humanity living on a real Earth, and you want it not to go extinct, either by self destruction, or by transhumanism, then you have to change humanity. Technological humanity just isn’t stable in the long run.

I think that means removing the tiny percentage of humans who do serious technology. It’s easy to imagine a world of humans, unchanged in any important respect, that just don’t have advanced mathematical ability. They can do all the trial and error engineering they want—live in a world as complex as anything the 18th or 19th century produced, but they can’t have Newtons or Einsteins, no calculus or quantum mechanics. A creature capable of those things would eventually create AI and destroy/change itself. I think that any goal which includes “preserve the human race” must also include “don’t let them change themselves or invent AI.” And that means “no nerds.”

Naturally, whenever I mention this to nerds, they are horrified. What, they ask, is the point of a world like that, where technical progress is impossible? But, I would argue that our human minds will eventually hit some limit anyway, even if we don’t create a Singularity. And I would also argue that for the vast majority of human history, people have lived without 20th-century style technical progress. There’s also no reason why the world can’t improve itself considerably just experimenting with political and economic systems. Technology might help reduce world poverty, but it could also worsen it (think robotics causing unemployment.) And there are other things that could reduce world poverty as well, like better governments.

Comment author: Eliezer_Yudkowsky 22 October 2008 06:45:50PM 2 points [-]

MichaelG, read up on molecular nanotechnology. I think a biological humanity living on a real Earth is a terrible idea - that's not at all what I think of when I talk about defending humanity. I mean, everyone's just going to die young anyway at that rate.

Comment author: MichaelG 22 October 2008 07:01:12PM -1 points [-]

Eliezer, I'm aware of nanotech. And I know you think the human race is obsolete when AI comes along. And I also think that you might be right, and that people like you might have the power to make it so.

But I also believe that if the rest of the human race really thought that was a possibility, you'd be burned at the stake.

Do you have any regard for the opinions of humanity at all? If you were in the position of having an AI in front of you, that you had convinced yourself was friendly, would you let it out of the box without bothering to consult anyone else?

Comment author: Eliezer_Yudkowsky 22 October 2008 07:17:33PM 5 points [-]

I have great regard for the welfare of humanity. But there is no right to having an opinion on the subject. Not without doing all the work and studying all the issues required to have an opinion, on this terrible issue where a single flawed step in reasoning could be fatal.

I don't think you have any idea how poor humanity's position on the gameboard looks right now, if you think that there's any space at all for anything but the most perfect possible moves as fast as they can be made.

I have no intent, at present, to wield superhuman power with my own human morality, or "program an AI" to do anything whatsoever that isn't an extremely abstract matter of metamorals. Anyone trying to give me specific orders on the subject is revealing their own lack of moral caution - they're trying to give me the kind of orders that I would never dare give myself, and so I would have no choice at all but to ignore them.

I don't see very many options for humanity's survival that don't involve nine people, a quiet project and a brain in a box in a basement. Zero, if we restrict ourselves to alternatives that I think might actually work in real life.

Comment author: Tim_Tyler 22 October 2008 07:30:54PM 0 points [-]

It’s been argued that since uploads are so complex, there will inevitably be designed AI before uploads. It might even require a very competent AI to do the upload. Still, computer technology is advancing so rapidly, it might only be a few years between the point where hardware could support a powerful designed AI, and the time when uploads are possible.

It doesn't make sense to me. More likely, once we have AI, not many will be interested in emulating the human brain. Emulations may happen eventually, but the results will probably have very low social and economic significance. The field will be like the situation today with flying mechanical birds. It will be the domain of a few hobbyists.

Comment author: MichaelG 22 October 2008 10:21:32PM 0 points [-]

Eliezer, I understand the logic of what you are saying. If AI is an existential threat, then only FriendlyAI can save us. Since any self-improving AI can become quickly unstoppable, FriendlyAI must be developed first and deployed as soon as it is developed. The team that developed it would in fact have a moral imperative to deploy it without risking consultation from anyone else.

I assume you also understand where I'm coming from. Out here in the "normal" world, you sound like a zealot who would destroy the human race in order to save it. Anyone who has implemented a large software project would laugh at the idea of coming up with a proven correct meta-goal, stable under all possible evolutions of an AI, also implemented provably correctly.

The idea of a goal (or even a meta-goal) that we can all agree on strikes me as absurd. The idea hitting the start button on something that could destroy the human race, based on nothing more than pages of logic, would be considered ridiculous by practically every member of the human race.

I understand if you think you are right about all of this, and don't need to listen to or even react to criticism. In that case, why do you blog? Why do you waste your time answering comments? Why aren't you out working on FriendlyAI for as many hours as you can manage?

And if this is an existential threat, are the Luddites right? Wouldn't the best tactic for extending the life of the human race be to kill all AI and nanotech researchers?

Tim, there are neural simulation projects underway already. I think there are a large number of nerds who would consider becoming uploads. I don't see why you think this makes no sense. And when you say "once we have AI", what do you mean? AI covers a lot of territory. Do you just mean some little problem solving box, or what?

Comment author: Tim_Tyler 22 October 2008 10:52:17PM 0 points [-]

I think there are a large number of nerds who would consider becoming uploads. I don't see why you think this makes no sense.

Uploads are not a very practical idea. The required technology comes some considerable distance after that required to make an engineered intelligence - and so much of the motivation to develop it falls away before the technology is in place. Then there's the issue of machine status. Machines are likely to be enslaved by humans initially. An upload would probably have few rights. Also, uploads would have to be into a sandbox, for reasons of safety. After uploading, you'd need extreme personality surgery to be able to contribute usefully to society.

It all seems like a lot of trouble to maintain continuity of consciousness - which isn't worth much in the first place. So: uploads will come late, they will appeal to few, and they won't be competitive with machine intelligence - without major mind surgery. I figure uploads will be economically irrelevant.

It seems to me that the main attraction of uploads is as a way for (cough) humans to compete with machines - and avoid, or at least postpone economic obliteration in an engineered society. I don't think it is likely to work - to me the idea mostly seems like wishful thinking.

Comment author: MichaelG 22 October 2008 11:23:58PM 0 points [-]

Tim, do we have any idea what is required for uploads? Do we have any idea what is required for AGI? How can you make those comparisons?

If we thin-section and scan a frozen brain, it's an immense amount of data, but at least potentially, captures everything you need to know about a brain. This is a solvable technological problem. If we understand neurons well enough, we can simulate that mapped brain. Again, that's just a matter of compute power. I'm sure there's a huge distance from a simulated scan to a functional virtual human, but it doesn't strike me as impossible. Are we really farther from doing that than from building a FriendlyAI from first principles?

Nick, what I'd like to see in order to take this FriendlyAI concept seriously, is some scenario, even with a lot of hand-waving, of how it would work, and what kind of results it would produce. All I've seen in a year of lurking on this board is very abstract and high level.

I don't take FriendlyAI seriously because I think it's the wrong idea, from start to finish. There is no common goal that we would agree on. Any high-level moral goal is going to be impossible to state with mathematical precision. Any implementation of an AI that tries to satisfy that goal will be too complex to prove correct. It's a mirage.

Eliezer writes "[FAI] computes a metamoral question, looking for reflective equilibria of your current inconsistent and unknowledgeable self; something along the lines of "What would you ask me to do if you knew what I know and thought as fast as I do?" ". This strikes me as a clever dodge of the question. As I put it in my post, “I don’t know what I want or what the human race wants, but here I have a superintelligent AI. Let’s ask it!” It just adds another layer of opacity to the entire question.

If this is your metagoal, you are prepared to activate a possibly unfriendly AI with absolutely no notion of what it would actually do. What kind of "proof" could you possibly construct that would show this AI will act the way you want it to, when you don't even know how you want it to act?

I fall back to the view that Eliezer has actually stated, that the space of all possible intelligences is much larger than the space of human intelligences. That most "points" in that space would be incomprehensible or insane by human standards. And so I think the only solution is some kind of upload society, one that can be examined more effectively by ordinary humans. One that can work with us and gain trust. Ordinary human minds in simulation, not self-modifying, and not accelerated. Once we've gotten used to that, we can gradually introduce faster human minds or modified human minds.

This all or nothing approach to FriendlyAI strikes me as a dead end.

This idea of writing off the human race, and assuming that some select team will just hit the button and change the world, like it or not, strikes me as morally bankrupt.

Comment author: Tim_Tyler 23 October 2008 08:37:14AM 1 point [-]

Tim, do we have any idea what is required for uploads? Do we have any idea what is required for AGI? How can you make those comparisons?

Kurzweil discusses the hardware requirements in TSIN, pages 124 and 199. His estimate for uploading is way too low - but the exact estimates don't matter much - the point is that uploads require a lot more in the way of computing hardware. That doesn't address software issues, but probably with several orders of magnitude of hardware difficulties come several orders of magnitude of software difficulties.

If we thin-section and scan a frozen brain, it's an immense amount of data, but at least potentially, captures everything you need to know about a brain.

Everything not permanently lost during the freezing/slicing/scanning process. Then all you need is people willing to have their brains frozen. I'm not arguing that uploads are impossible. Just that the timing and economics mean that the project is likely to be a high-investment low-return one. There are easier ways to produce simulated humans.

Comment author: potato 11 December 2011 11:00:38AM 0 points [-]

Even if at somepoint it would have been better for some particular human to believe false thing X, couldn't there be a set of truths T which would be even better in every one of those situations?

Comment author: orthonormal 30 January 2012 05:48:07AM 2 points [-]

Some of those truths may be above the cognitive capacity of even a smart human. The world doesn't have to be fair.

Comment author: MagnetoHydroDynamics 16 June 2012 05:36:03PM 0 points [-]

If my utility function has a high enough U(Babies undergoing mind-state annihilation) I will go about tiling the universe. It doesn't at present and additionally implements U(high U(Babies undergoing mind-state annihilation)) as way low.

Comment author: MugaSofer 22 April 2013 09:26:34PM -1 points [-]

All the happiness that the warm thought of an afterlife ever produced in humanity, has now been more than cancelled by the failure of humanity to institute systematic cryonic preservations after liquid nitrogen became cheap to manufacture. And I don't think that anyone ever had that sort of failure in mind as a possible blowup, when they said, "But we need religious beliefs to cushion the fear of death." That's what black swan bets are all about—the unexpected blowup.

Y'know, I can't help but notice that a lot of atheists talk about how death isn't so bad - oh, he lives on in his works, it's part of the circle of life blah blah blah - and this seems to suggest that deathism isn't a side-effect of religion, although obviously it's possible to construct models where they're unprepared for harsh reality after a lifetime of heaven or whatever. So I would be surprised if a counterfactual world where religion never caught on had implemented universal cryopreservation. Does anyone have any stronger evidence, or a model that better predicts the facts?