tl;dr An unconstrained search through possible future worlds is a dangerous way of choosing positive outcomes. Constrained, imperfect or under-optimised searches work better.

Some suggested methods for designing AI goals, or controlling AIs, involve unconstrained searches through possible future worlds. This post argues that this is a very dangerous thing to do, because of the risk of being tricked by "siren worlds" or "marketing worlds". The thought experiment starts with an AI designing a siren world to fool us, but that AI is not crucial to the argument: it's simply an intuition pump to show that siren worlds can exist. Once they exist, there is a non-zero chance of us being seduced by them during a unconstrained search, whatever the search criteria are. This is a feature of optimisation: satisficing and similar approaches don't have the same problems.

 

The AI builds the siren worlds

Imagine that you have a superintelligent AI that's not just badly programmed, or lethally indifferent, but actually evil. Of course, it has successfully concealed this fact, as "don't let humans think I'm evil" is a convergent instrumental goal for all AIs.

We've successfully constrained this evil AI in a Oracle-like fashion. We ask the AI to design future worlds and present them to human inspection, along with an implementation pathway to create those worlds. Then if we approve of those future worlds, the implementation pathway will cause them to exist (assume perfect deterministic implementation for the moment). The constraints we've programmed means that the AI will do all these steps honestly. Its opportunity to do evil is limited exclusively to its choice of worlds to present to us.

The AI will attempt to design a siren world: a world that seems irresistibly attractive while concealing hideous negative features. If the human mind is hackable in the crude sense - maybe through a series of coloured flashes - then the AI would design the siren world to be subtly full of these hacks. It might be that there is some standard of "irresistibly attractive" that is actually irresistibly attractive: the siren world would be full of genuine sirens.

Even without those types of approaches, there's so much manipulation the AI could indulge in. I could imagine myself (and many people on Less Wrong) falling for the following approach:

First, the siren world looks complicated, wrong and scary - but with just a hint that there's something more to it. Something intriguing, something half-glimpsed, something making me want to dig deeper. And as I follow up this something, I see more patterns, and seem to gain a greater understanding. Not just of the world I'm looking at, but of the meaning of good itself. The world seems to confirm to me some of my ideas about what constitutes a worthwhile life - not just the ideas I've been able to articulate, but the ones I've only got a vague half-baked notion of, and the ones I didn't even know I had.

The longer I stare into this world, the greater an understanding I get of my own values. And this is just the starting point: the world subtly opens up new avenues of philosophical musings in my brain, the images I see triggering me to come up with my own insights and thought experiments as to the meaning of goodness - insights that are then carefully confirmed as I did deeper. I could stay staring at this wonderful, beautiful and complex world for hours, days, years, gaining a deeper wisdom and understanding as I go, truly appreciating how the mysteries of this place unravel into new versions of worth and goodness. Every time I ever have a doubt about it, that doubt is systemically resolved into a new glorious confirmation of how much the AI really gets what I value, even before I knew how much I valued it.

Meanwhile, at some level of complexity safely beyond what my human mind will ever reach, the AI is hiding all the evil and unmitigated suffering.

Your siren world may differ from mine - you may be vulnerable to different tricks. Since people often believe they would be immune to such manipulation, feel free to imagine that the AI is fooling someone else - we know how stupid other people can be, right? - while you watch as a helpless bystander.

The problem is that a future universe is likely to be vast and complicated. When inspecting it, we have certain inspection criteria (IC). These consist of our values and preferences, but also the ways we would look into this universe, how we would follow up on initial impressions, various measures and yardsticks we might use to summarise the data we see, all the clever and devious plans we might come up with to ferret out "what's really going on". These IC are going to be considerably simpler than the totality of the future universe. So the AI's task is to optimise a universe design that passes the IC, while shoving in as much disutility as it can - which in a large universe, is a tremendous amount. Unless our IC are perfect and already include a good solution to the problem of value (in which case we've solved the friendliness problem already), a superintelligent AI will likely succeed at its task.

 

Siren and marketing worlds without builders

The above thought experiment needed a superintelligent evil AI for the design of the siren world. But if we admit that that is possible, we don't actually need the AI any more. The siren worlds exist: there are potential worlds of extreme disutility that satisfie our IC. If we simply did an unconstrained search across all possible future worlds (something like the search in Paul Christiano's indirect normativity - an idea that inspired the siren world concept), then we would at some point find siren worlds. And if we took the time to inspect them, we'd get sucked in by them.

How bad is this problem in general? A full search will not only find the siren worlds, but also a lot of very-seductive-but-also-very-nice worlds - genuine eutopias. We may feel that it's easier to be happy than to pretend to be happy (while being completely miserable and tortured and suffering). Following that argument, we may feel that there will be far more eutopias than siren worlds - after all, the siren worlds have to have bad stuff plus a vast infrastructure to conceal that bad stuff, which should at least have a complexity cost if nothing else. So if we chose the world that best passed our IC - or chose randomly among the top contenders - we might be more likely to hit a genuine eutopia than a siren world.

Unfortunately, there are other dangers than siren worlds. We are now optimising not for quality of the world, but for ability to seduce or manipulate the IC. There's no hidden evil in this world, just a "pulling out all the stops to seduce the inspector, through any means necessary" optimisation pressure. Call a world that ranks high in this scale a "marketing world". Genuine eutopias are unlikely to be marketing worlds, because they are optimised for being good rather than seeming good. A marketing world would be utterly optimised to trick, hack, seduce, manipulate and fool our IC, and may well be a terrible world in all other respects. It's the old "to demonstrate maximal happiness, it's much more reliable to wire people's mouths to smile rather than make them happy" problem all over again: the very best way of seeming good may completely preclude actually being good. In a genuine eutopia, people won't go around all the time saying "Btw, I am genuinely happy!" in case there is a hypothetical observer looking in. If every one of your actions constantly proclaims that you are happy, chances are happiness is not your genuine state. EDIT: see also my comment:

We are both superintelligences. You have a bunch of independently happy people that you do not aggressively compel. I have a group of zombies - human-like puppets that I can make do anything, appear to feel anything (though this is done sufficiently well that outside human observers can't tell I'm actually in control). An outside human observer wants to check that our worlds rank high on scale X - a scale we both know about.

Which of us do you think is going to be better able to maximise our X score?

This can also be seen as a epistemic version of Goodhart's law: "When a measure becomes a target, it ceases to be a good measure." Here the IC are the measure, and the marketing worlds are targeting them, and hence they cease to be a good measure. But recall that the IC include the totality of approaches we use to rank these worlds, so there's no way around this problem. If instead of inspecting the worlds, we simply rely on some sort of summary function, then the search will be optimised to find anything that can fool/pass that summary function. If we use the summary as a first filter, then apply some more profound automated checking, then briefly inspect the outcome so we're sure it didn't go stupid - then the search will optimised for "pass the summary, pass automated checking, seduce the inspector".

Different IC therefore will produce different rankings of worlds, but the top worlds in any of the ranking will be marketing worlds (and possibly siren worlds).

 

Constrained search and satisficing our preferences

The issue is a problem of (over) optimisation. The IC correspond roughly with what we want to value, but differs from it in subtle ways, enough that optimising for one could be disastrous for the other. If we didn't optimise, this wouldn't be a problem. Suppose we defined an acceptable world as one that we would judge "yeah, that's pretty cool" or even "yeah, that's really great". Then assume we selected randomly among the acceptable worlds. This would probably result in a world of positive value: siren worlds and marketing worlds are rare, because they fulfil very specific criteria. They triumph because they score so high on the IC scale, but they are outnumbered by the many more worlds that are simply acceptable.

This is in effect satisficing over the IC, rather than optimising over them. Satisficing has its own issues, however, so other approaches could be valuable as well. One way could be use constrained search. If for instance we took a thousand random worlds and IC-optimised over them, we're very unlikely to encounter a siren or marketing world. We're also very unlikely to encounter a world of any quality, though; we'd probably need to IC-optimise over at least a trillion worlds to find good ones. There is a tension in the number: as the number of worlds searched increases, their quality increases, but so does the odds of encountering a marketing or siren world. EDIT: Lumifer suggested using a first-past-the-post system: search through worlds, and pick the first acceptable one we find. This is better than the approach I outlined in this paragraph.

We could also restrict the search by considering "realistic" worlds. Suppose we had to take 25 different yes-no decisions that could affect the future of the humanity. This might be something like "choosing which of these 25 very different AIs to turn on and let loose together" or something more prosaic (which stocks to buy, which charities to support). This results in 225 different future worlds to search through: barely more than 33 million. Because there are so few worlds, they are unlikely to contain a marketing world (given the absolutely crucial proviso that none of the AIs is an IC-optimiser!). But these worlds are not drawn randomly from the space of future worlds, but are dependent on key decisions that we believe are important and relevant. Therefore they are very likely to contain an acceptable world - or at least far more likely than a random set of 33 million worlds would be. By constraining the choices in this way, we have in effect satisficed without satisficing, which is both Zen and useful.

As long as we're aware of the problem, other approaches may also allow for decent search without getting sucked in by a siren or a marketer.

New to LessWrong?

Siren worlds and the perils of over-optimised search
New Comment


418 comments, sorted by Click to highlight new comments since:
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

While not generally an opponent of human sexuality, to be kind to all the LW audience including those whose parents might see them browsing, please do remove the semi-NSFW image.

4Stuart_Armstrong
Is the new one more acceptable?
7MugaSofer
See, now I'm curious about the old image...
3Stuart_Armstrong
The image can be found at http://truckstopstruckstop.tumblr.com/post/39569037859/nude-with-skull-via
6Eliezer Yudkowsky
Sure Why Not

LOL. The number of naked women grew from one to two, besides the bare ass we now also have breasts with nipples visible (OMG! :-D) and yet it's now fine just because it is old-enough Art.

4A1987dM
The fact that the current picture is a painting and the previous one was a photograph might also have something to do with it.
2Lumifer
Can you unroll this reasoning?
3A1987dM
It's just what my System 1 tells me; actually, I wouldn't know how to go about figuring out whether it's right.
1[anonymous]
Is there some other siren you'd prefer to see?
5Lumifer
See or hear? :-D

This indeed is why "What a human would think of a world, given a defined window process onto a world" was not something I considered as a viable form of indirect normativity / an alternative to CEV.

To my mind, the interesting part is the whole constrain search/satisficing ideas which may allow such an approach to be used.

[-][anonymous]170

First question: how on Earth would we go about conducting a search through possible future universes, anyway? This thought experiment still feels too abstract to make my intuitions go click, in much the same way that Christiano's original write-up of Indirect Normativity did. You simply can't actually simulate or "acausally peek at" whole universes at a time, or even Earth-volumes in such. We don't have the compute-power, and I don't understand how I'm supposed to be seduced by a siren that can't sing to me.

It seems to me that the greater danger is that a UFAI would simply market itself as an FAI as an instrumental goal and use various "siren and marketing" tactics to manipulate us into cleanly, quietly accepting our own extinction -- because it could just be cheaper to manipulate people than to fight them, when you're not yet capable of making grey goo but still want to kill all humans.

And if we want to talk about complex nasty dangers, it's probably going to just be people jumping for the first thing that looks eutopian, in the process chucking out some of their value-set. People do that a lot, see: every single so-called "utopian" movement ever ... (read more)

5Stuart_Armstrong
Two main reasons for this: first, there is Christiano's original write-up, which has this problem. Second, we may be in a situation where we ask an AI to simulate the consequences of its choice, have a glance at it, and then approve/disapprove. That's less a search problem, and more the original siren world problem, and we should be aware of the problem.
7[anonymous]
This sounds extremely counterintuitive. If I have an Oracle AI that I can trust to answer more-or-less verbal requests (defined as: any request or "program specification" too vague for me to actually formalize), why have I not simply asked it to learn, from a large corpus of cultural artifacts, the Idea of the Good, and then explain to me what it has learned (again, verbally)? If I cannot trust the Oracle AI, dear God, why am I having it explore potential eutopian future worlds for me?

If I cannot trust the Oracle AI, dear God, why am I having it explore potential eutopian future worlds for me?

Because I haven't read Less Wrong? ^_^

This is another argument against using constrained but non-friendly AI to do stuff for us...

4Stuart_Armstrong
Colloquially, this concept is indeed very close to overfitting. But it's not technically overfitting ("overfitting occurs when a statistical model describes random error or noise instead of the underlying relationship."), and using the term brings in other connotations. For instance, it may be that the AI needs to use less data to seduce us than it would to produce a genuine eutopia. It's more that it fits the wrong target function (having us approve its choice vs a "good" choice) rather than fitting it in an overfitted way.
2[anonymous]
Thanks. My machine-learning course last semester didn't properly emphasize the formal definition of overfitting, or perhaps I just didn't study it hard enough. What I do want to think about here is: is there a mathematical way to talk about what happens when a learning algorithm finds the wrong correlative or causative link among several different possible links between the data set and the target function? Such maths would be extremely helpful for advancing the probabilistic value-learning approach to FAI, as they would give us a way to talk about how we can interact with an agent's beliefs about utility functions while also minimizing the chance/degree of wireheading.
0Stuart_Armstrong
That would be useful! A short search gives "bias" as the closest term, which isn't very helpful.
4[anonymous]
Unfortunately "bias" in statistics is completely unrelated to what we're aiming for here. In ugly, muddy words, what we're thinking is that we give the value-learning algorithm some sample of observations or world-states as "good", and possibly some as "bad", and "good versus bad" might be any kind of indicator value (boolean, reinforcement score, whatever). It's a 100% guarantee that the physical correlates of having given the algorithm a sample apply to every single sample, but we want the algorithm to learn the underlying causal structure of why those correlates themselves occurred (that is, to model our intentions as a VNM utility function) rather than learn the physical correlates themselves (because that leads to the agent wireheading itself). Here's a thought: how would we build a learning algorithm that treats its samples/input as evidence of an optimization process occurring and attempts to learn the goal of that optimization process? Since physical correlates like reward buttons don't actually behave as optimization processes themselves, this would ferret out the intentionality exhibited by the value-learner's operator from the mere physical effects of that intentionality (provided we first conjecture that human intentions behave detectably like optimization). Has that whole "optimization process" and "intentional stance" bit from the LW Sequences been formalized enough for a learning treatment?
2Quill_McGee
http://www.fungible.com/respect/index.html This looks to be very related to the idea of "Observe someone's actions. Assume they are trying to accomplish something. Work out what they are trying to accomplish." Which seems to be what you are talking about.
1[anonymous]
That looks very similar to what I was writing about, though I've tried to be rather more formal/mathematical about it instead of coming up with ad-hoc notions of "human", "behavior", "perception", "belief", etc. I would want the learning algorithm to have uncertain/probabilistic beliefs about the learned utility function, and if I was going to reason about individual human minds I would rather just model those minds directly (as done in Indirect Normativity).
0Stuart_Armstrong
I will think about this idea...
1[anonymous]
The most obvious weakness is that such an algorithm could easily detect optimization processes that are acting on us (or, if you believe such things exist, you should believe this algorithm might locate them mistakenly), rather than us ourselves.
1Stuart_Armstrong
I've been thinking about this, and I haven't found any immediately useful way of using your idea, but I'll keep it in the back of my mind... We haven't found a good way of identifying agency in the abstract sense ("was cosmic phenonmena X caused by an agent, and if so, which one?" kind of stuff), so this might be a useful simpler problem...
2[anonymous]
Upon further research, it turns out that preference learning is a field within machine learning, so we can actually try to address this at a much more formal level. That would also get us another benefit: supervised learning algorithms don't wirehead. Notably, this fits with our intuition that morality must be "taught" (ie: via labelled data) to actual human children, lest they simply decide that the Good and the Right consists of eating a whole lot of marshmallows. And if we put that together with a conservation heuristic for acting under moral uncertainty (say: optimize for expectedly moral expected utility, thus requiring higher moral certainty for less-extreme moral decisions), we might just start to make some headway on managing to construct utility functions that would mathematically reflect what their operators actually intend for them to do. I also have an idea written down in my notebook, which I've been refining, that sort of extends from what Luke had written down here. Would it be worth a post?
0[anonymous]
Hi, there appears to be a lot of work on learning causal structure from data.
0[anonymous]
Keywords? I've looked through Wikipedia and the table of contents from my ML textbook, but I haven't found the right term to research yet. "Learn a causal structure from the data and model the part of it that appears to narrow the future" would in fact be how to build a value-learner, but... yeah. EDIT: One of my profs from undergrad published a paper last year about causal-structure. The question is how useful it is for universal AI applications. Joshua Tenenbaum tackled it from the cog-sci angle in 2011, but again, I'm not sure how to transfer it over to the UAI angle. I was searching for "learning causal structure from data" -- herp, derp.
0IlyaShpitser
Who was this prof?
4[anonymous]
I was referring to David Jensen, who taught "Research Methods in Empirical Computer Science" my senior year.
2IlyaShpitser
Thanks.

This puts me in mind of a thought experiment Yvain posted a while ago (I’m certain he’s not the original author, but I can’t for the life of me track it any further back than his LiveJournal):

“A man has a machine with a button on it. If you press the button, there is a one in five million chance that you will die immediately; otherwise, nothing happens. He offers you some money to press the button once. What do you do? Do you refuse to press it for any amount? If not, how much money would convince you to press the button?”

This is – I think – analogous to y... (read more)

0Stuart_Armstrong
I consider that is also a constrained search!

One issue here is that worlds with an "almost-friendly" AI (one whose friendliness was botched in some respect) may end up looking like siren or marketing worlds.

In that case, worlds as bad as sirens will be rather too common in the search space (because AIs with botched friendliness are more likely than AIs with true friendliness) and a satisficing approach won't work.

2Stuart_Armstrong
Interesting thought there...

We could also restrict the search by considering "realistic" worlds. Suppose we had to take 25 different yes-no decisions that could affect the future of the humanity. This might be something like "choosing which of these 25 very different AIs to turn on and let loose together" or something more prosaic (which stocks to buy, which charities to support). This results in 225 different future worlds to search through: barely more than 33 million. Because there are so few worlds, they are unlikely to contain a marketing world (given the absolutely crucial

... (read more)

I've just now found my way to this post, from links in several of your more recent posts, and I'm curious as to how this fits in with more recent concepts and thinking from yourself and others.

Firstly, in terms of Garrabrant's taxonomy, I take it that the "evil AI" scenario could be considered a case of adversarial Goodhart, and the siren and marketing worlds without builders could be considered cases of regressional and/or extremal Goodhart. Does that sound right?

Secondly, would you still say that these scenarios demonstrate reas... (read more)

3Stuart_Armstrong
To a large extent I do, but there may be some residual effects similar to the above, so some anti-optimising pressure might still be useful.

It seems based on your later comments that the premise of marketing worlds existing relies on there being trade-offs between our specified wants and our unspecified wants, so that the world optimised for our specified wants must necessarily be highly likely to be lacking in our unspecified ones ("A world with maximal bananas will likely have no apples at all").

I don't think this is necessarily the case. If I only specify that I want low rates of abortion, for example, then I think it highly likely that 'd get a world that also has low rates of ST... (read more)

0Stuart_Armstrong
Yes, certainly. That's a problem of optimisation with finite resources. If A is a specified want and B is an unspecified want, then we shouldn't confuse "there are worlds with high A and also high B" with "the world with the highest A will also have high B".
0Stuart_Armstrong
You would get a world with no conception, or possibly with no humans at all.
2PhilosophyTutor
I don't think you have highlighted a fundamental problem since we can just specify that we mean a low percentage of conceptions being deliberately aborted in liberal societies where birth control and abortion are freely available to all at will. My point, though, is that I don't think it is very plausible that "marketing worlds" will organically arise where there are no humans, or no conception, but which tick all the other boxes we might think to specify in our attempts to describe an ideal world. I don't see how there being no conception or no humans could possibly be a necessary trade-off with things like wealth, liberty, rationality, sustainability, education, happiness, the satisfaction of rational and well-informed preferences and so forth. Of course a sufficiently God-like malevolent AI could presumably find some way of gaming any finite list we give it, since there are probably an unbounded number of ways of bringing about horrible worlds, so this isn't a problem with the idea of siren worlds. I just don't find the idea of market worlds very plausible because so many of the things we value are fundamentally interconnected.
1Stuart_Armstrong
The "no conception" example is just to illustrate that bad things happen when you ask an AI to optimise along a certain axis without fully specifying what we want (which is hard/impossible). A marketing world is fully optimised along the "convince us to choose this world" axis. If at any point, the AI in confronted with a choice along the lines of "remove genuine liberty to best give the appearance of liberty/happiness", it will choose to do so. That's actually the most likely way a marketing world could go wrong - the more control the AI has over people's appearance and behaviour, the more capable it is of making the world look good. So I feel we should presume that discrete-but-total AI control over the world's "inhabitants" would be the default in a marketing world.
5PhilosophyTutor
I think this and the "finite resources therefore tradeoffs" argument both fail to take seriously the interconnectedness of the optimisation axes which we as humans care about. They assume that every possible aspect of society is an independent slider which a sufficiently advanced AI can position at will, even though this society is still going to be made up of humans, will have to be brought about by or with the cooperation of humans and will take time to bring about. These all place constraints on what is possible because the laws of physics and human nature aren't infinitely malleable. I don't think discreet but total control over a world is compatible with things like liberty, which seem like obvious qualities to specify in an optimal world we are building an AI to search for. I think what we might be running in to here is less of an AI problem and more of a problem with the model of AI as an all-powerful genie capable of absolutely anything with no constraints whatsoever.
0Stuart_Armstrong
Precisely and exactly! That's the whole of the problem - optimising for one thing (appearance) results in the loss of other things we value. Next challenge: define liberty in code. This seems extraordinarily difficult. So we do agree that there are problem with an all-powerful genie? Once we've agreed on that, we can scale back to lower AI power, and see how the problems change. (the risk is not so much that the AI would be an all powerful genie, but that it could be an all powerful genie compared with humans).
5PhilosophyTutor
This just isn't always so. If you instruct an AI to optimise a car for speed, efficiency and durability but forget to specify that it has to be aerodynamic, you aren't going to get a car shaped like a brick. You can't optimise for speed and efficiency without optimising for aerodynamics too. In the same way it seems highly unlikely to me that you could optimise a society for freedom, education, just distribution of wealth, sexual equality and so on without creating something pretty close to optimal in terms of unwanted pregnancies, crime and other important axes. Even if it's possible to do this, it seems like something which would require extra work and resources to achieve. A magical genie AI might be able to make you a super-efficient brick-shaped car by using Sufficiently Advanced Technology indistinguishable from magic but even for that genie it would have to be more work than making an equally optimal car by the defined parameters that wasn't a silly shape. In the same way an effectively God-like hypothetical AI might be able to make a siren world that optimised for everything except crime and create a world perfect in every way except that it was rife with crime but it seems like it would be more work, not less. I think if we can assume we have solved the strong AI problem, we can assume we have solved the much lesser problem of explaining liberty to an AI. We've got a problem with your assumptions about all-powerful genies, I think, because I think your argument relies on the genie being so ultimately all-powerful that it is exactly as easy for the genie to make an optimal brick-shaped car or an optimal car made out of tissue paper and post-it notes as it is for the genie to make an optimal proper car. I don't think that genie can exist in any remotely plausible universe. If it's not all-powerful to that extreme then it's still going to be easier for the genie to make a society optimised (or close to it) across all the important axes at once than one opt
1Stuart_Armstrong
The strong AI problem is much easier to solve than the problem of motivating an AI to respect liberty. For instance, the first one can be brute forced (eg AIXItl with vast resources), the second one can't. Having the AI understand human concepts of liberty is pointless unless it's motivated to act on that understanding. An excess of anthropomophisation is bad, but an analogy could be about creating new life (which humans can do) and motivating that new life to follow specific rules are requirements if they become powerful (which humans are pretty bad at at).
5PhilosophyTutor
I don't believe that strong AI is going to be as simple to brute force as a lot of LessWrongers believe, personally, but if you can brute force strong AI then you can just get it to run a neuron-by-neuron simulation of the brain of a reasonably intelligent first year philosophy student who understands the concept of liberty and tell the AI not to take actions which the simulated brain thinks offend against liberty. That is assuming that in this hypothetical future scenario where we have a strong AI we are capable of programming that strong AI to do any one thing instead of another, but if we cannot do that then the entire discussion seems to me to be moot.
8Nornagest
I've met far too many first-year philosophy students to be comfortable with this program.
0Stuart_Armstrong
How? "tell", "the simulated brain thinks" "offend": defining those incredibly complicated concepts contains nearly the entirety of the problem.
2PhilosophyTutor
I could be wrong but I believe that this argument relies on an inconsistent assumption, where we assume we have solved the problem of creating an infinitely powerful AI, but we have not solved the problem of operationally defining commonplace English words which hundreds of millions of people successfully understand in such a way that a computer can perform operations using them. It seems to me that the strong AI problem is many orders of magnitude more difficult than the problem of rigorously defining terms like "liberty". I imagine that a relatively small part of the processing power of one human brain is all that is needed to perform operations on terms like "liberty" or "paternalism" and engage in meaningful use of them so it is a much, much smaller problem than the problem of creating even a single human-level AI, let alone a vastly superhuman AI. If in our imaginary scenario we can't even define "liberty" in such a way that a computer can use the term, it doesn't seem very likely that we can build any kind of AI at all.
0[anonymous]
My mind is throwing a type-error on reading your comment. Liberty could well be like pornography: we know it when we see it, based on probabilistic classification. There might not actually be a formal definition of liberty that includes all actual humans' conceptions of such as special cases, but instead a broad range of classifier parameters defining the variation in where real human beings "draw the line".
4PhilosophyTutor
The standard LW position (which I think is probably right) is that human brains can be modelled with Turing machines, and if that is so then a Turing machine can in theory do whatever it is we do when we decide that something ls liberty, or pornography. There is a degree of fuzziness in these words to be sure, but the fact we are having this discussion at all means that we think we understand to some extent what the term means and that we value whatever it is that it refers to. Hence we must in theory be able to get a Turing machine to make the same distinction although it's of course beyond our current computer science or philosophy to do so.
0Stuart_Armstrong
Yes. Here's another brute force approach: upload a brain (without understanding it), run it very fast with simulated external memory, subject it to evolutionary pressure. All this can be done with little philosophical and conceptual understanding, and certainly without any understanding of something as complex as liberty.
-1PhilosophyTutor
If you can do that, then you can just find someone who you think understands what we mean by "liberty" (ideally someone with a reasonable familiarity with Kant, Mill, Dworkin and other relevant writers), upload their brain without understanding it, and ask the uploaded brain to judge the matter. (Off-topic: I suspect that you cannot actually get a markedly superhuman AI that way, because the human brain could well be at or near a peak in the evolutionary landscape so that there is no evolutionary pathway from a current human brain to a vastly superhuman brain. Nothing I am aware of in the laws of physics or biology says that there must be any such pathway, and since evolution is purposeless it would be an amazing lucky break if it turned out that we were on the slope of the highest peak there is, and that the peak extends to God-like heights. That would be like if we put evolutionary pressure on a cheetah and discovered that if we do that we can evolve a cheetah that runs at a significant fraction of c. However I believe my argument still works even if I accept for the sake of argument that we are on such a peak in the evolutionary landscape, and that creating God-like AI is just a matter of running a simulated human brain under evolutionary pressure for a few billion simulated years. If we have that capability then we must also be able to run a simulated philosopher who knows what "liberty" refers to). EDIT: Downvoting this without explaining why you disagree doesn't help me understand why you disagree.
0Stuart_Armstrong
And would their understanding of liberty remain stable under evolutionary pressure? That seems unlikely. Have not been downvoting it.
0PhilosophyTutor
I didn't think we needed to put the uploaded philosopher under billions of years of evolutionary pressure. We would put your hypothetical pre-God-like AI in one bin and update it under pressure until it becomes God-like, and then we upload the philosopher separately and use them as a consultant. (As before I think that the evolutionary landscape is unlikely to allow a smooth upward path from modern primate to God-like AI, but I'm assuming such a path exists for the sake of the argument).
1Stuart_Armstrong
And then we have to ensure the AI follows the consultant (probably doable) and define what querying process is acceptable (very hard). But your solution (which is close to Paul Christiano's) works whatever the AI is, we just need to be able to upload a human. My point was that we could conceivably create an AI without understanding any of the hard problems, still stands. If you want I can refine it: allow partial uploads: we can upload brains, but they don't function as stable humans, as we haven't mapped all the fine details we need to. However, we can use these imperfect uploads, plus a bit of evolution, to produce AIs. And here we have no understanding of how to control its motivations at all.
1PhilosophyTutor
I won't argue against the claim that we could conceivably create an AI without knowing anything about how to create an AI. It's trivially true in the same way that we could conceivably turn a monkey loose on a typewriter and get strong AI. I also agree with you that if we got an AI that way we'd have no idea how to get it to do any one thing rather than another and no reason to trust it. I don't currently agree that we could make such an AI using a non-functioning brain model plus "a bit of evolution". I am open to argument on the topic but currently it seems to me that you might as well say "magic" instead of "evolution" and it would be an equivalent claim.
0Stuart_Armstrong
Why are you confident that an AI that we do develop will not have these traits? You agree the mindspace is large, you agree we can develop some cognitive abilities without understanding them. If you add that most AI programmers don't take AI risk seriously and will only be testing their AI's in controlled environments, that the AI will be likely developed for a military or commercial purpose, I don't see why you'd have high confidence that they will converge on a safe design?
3XiXiDu
Why do you think such an AI wouldn't just fail at being powerful, rather than being powerful in a catastrophic way? If programs fail in the real world then they are not working well. You don't happen to come across a program that manages to prove the Riemann hypothesis when you designed it to prove the irrationality of the square root of 2.
0Stuart_Armstrong
If it fails at being powerful, we don't have to worry about it, so I feel free to ignore those probabilities. But you might come across a program motivated to eliminate all humans if you designed it to optimise the economy...
0TheAncientGeek
So you're not pursuing the claim that a SAI will probably be dangerous, you are just worried that it might be?
0Stuart_Armstrong
My claim has always been that the probability that an SAI will be dangerous is too high to ignore. I fluctuate on the exact probability, but I've never seen anything that drives it down to a level I feel comfortable with (in fact, I've never seen anything drive it below 20%).
-2[anonymous]
This is why the Wise employ normative uncertainty and the learning of utility functions from data, rather than hardcoding verbal instructions that only make sense in light of a complete human mind and social context.
3Stuart_Armstrong
Indeed. But the more of the problem you can formalise and solve (eg maintaining a stable utility function over self-improvements) the more likely the learning approach is to succeed.
2[anonymous]
Well yes, of course. I mean, if you can't build an agent that was capable of maintaining its learned utility while becoming vastly smarter (and thus capable of more accurately learning and enacting capital-G Goodness), then all that utility-learning was for nought.
1TheAncientGeek
Yeah, but hardcoding is an easier sell to people who know how to code but have never done .AI... Its like political demagogues selling unworkable but easily understood ideas.
0[anonymous]
Not really, no. Most people don't recognize the "hidden complexity of wishes" in Far Mode, or when it's their wishes. However, I think if I explain to them that I'll be encoding my wishes, they'll quickly figure out that my attempts to hardcode AI Friendliness are going to be very bad for them. Human intelligence evolved for winning arguments when status, wealth, health, and mating opportunities are at issue: thus, convince someone to treat you as an opponent, and leave the correct argument lying right where they can pick it up, and they'll figure things out quickly. Hmmm... I wonder if that bit of evolutionary psychology explains why many people act rude and nasty even to those close to them. Do we engage more intelligence when trying to win a fight than when trying to be nice?
-7TheAncientGeek
-2XiXiDu
The very idea underlying AI is enabling people to get a program to do what they mean without having to explicitly encode all details. What AI risk advocates do is to turn the whole idea upside down, claiming that, without explicitly encoding what you mean, your program will do something else. The problem here is that it is conjectured that the program will do what it was not meant to do in a very intelligent and structured manner. But this can't happen when it comes to intelligently designed systems (as opposed to evolved systems), because the nature of unintended consequences is overall chaotic. How often have you heard of intelligently designed programs that achieved something highly complex and marvelous, but unintended, thanks to the programmers being unable to predict the behavior of the program? I don't know of any such case. But this is exactly what AI risk advocates claim will happen, namely that a program designed to do X (calculate 1+1) will perfectly achieve Y (take over the world). If artificial general intelligence will eventually be achieved by some sort of genetic/evolutionary computation, or neuromorphic engineering, then I can see how this could lead to unfriendly AND capable AI. But an intelligently designed AI will either work as intended or be incapable of taking over the world (read: highly probable). This of course does not ensure a positive singularity (if you believe that this is possible at all), since humans might use such intelligently and capable AIs to wreck havoc (ask the AI to do something stupid, or something that clashes with most human values). So there is still a need for "friendly AI". But this is quite different from the idea of interpreting "make humans happy" as "tile the universe with smiley faces". Such a scenario contradicts the very nature of intelligently designed AI, which is an encoding of “Understand What Humans Mean” AND “Do What Humans Mean”. More here.
2[anonymous]
Alexander, have you even bothered to read the works of Marcus Hutter and Juergen Schmidhuber, or have you spent all your AI-researching time doing additional copy-pastas of this same argument every single time the subject of safe or Friendly AGI comes up? Your argument makes a measure of sense if you are talking about the social process of AGI development: plainly, humans want to develop AGI that will do what humans intend for it to do. However, even a cursory look at the actual research literature shows that the mathematically most simple agents (ie: those that get discovered first by rational researchers interested in finding universal principles behind the nature of intelligence) are capital-U Unfriendly, in that they are expected-utility maximizers with not one jot or tittle in their equations for peace, freedom, happiness, or love, or the Ideal of the Good, or sweetness and light, or anything else we might want. (Did you actually expect that in this utterly uncaring universe of blind mathematical laws, you would find that intelligence necessitates certain values?) No, Google Maps will never turn superintelligent and tile the solar system in computronium to find me a shorter route home from a pub crawl. However, an AIXI or Goedel Machine instance will, because these are in fact entirely distinct algorithms. In fact, when dealing with AIXI and Goedel Machines we have an even bigger problem than "tile everything in computronium to find the shortest route home": the much larger problem of not being able to computationally encode even a simple verbal command like "find the shortest route home". We are faced with the task of trying to encode our values into a highly general, highly powerful expected-utility maximizer at the level of, metaphorically speaking, pre-verbal emotion. Otherwise, the genie will know, but not care. Now, if you would like to contribute productively, I've got some ideas I'd love to talk over with someone for actually doing something about
-2XiXiDu
If I believed that anything as simple as AIXI could possibly result in practical general AI, or that expected utility maximizing was at all feasible, then I would tend to agree with MIRI. I don't. And I think it makes no sense to draw conclusions about practical AI from these models. This is crucial. That's largely irrelevant and misleading. Your autonomous car does not need to feature an encoding of an amount of human values that correspondents to its level of autonomy. That post has been completely debunked. ETA: Fixed a link to expected utility maximization.
-3XiXiDu
I asked several people what they think about it, and to provide a rough explanation. I've also had e-Mail exchanges with Hutter, Schmidhuber and Orseau. I also informally thought about whether practically general AI that falls into the category “consequentialist / expected utility maximizer / approximation to AIXI” could ever work. And I am not convinced. If general AI, which is capable of a hard-takeoff, and able to take over the world, requires less lines of code, in order to work, than to constrain it not to take over the world, then that's an existential risk. But I don't believe this to be the case. Since I am not a programmer, or computer scientist, I tend to look at general trends, and extrapolate from there. I think this makes more sense than to extrapolate from some unworkable model such as AIXI. And the general trend is that humans become better at making software behave as intended. And I see no reason to expect some huge discontinuity here. Here is what I believe to be the case: (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. Here is what I doubt: (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 human
8jimrandomh
This is a much bigger problem for your ability to reason about this area than you think.
1XiXiDu
A relevant quote from Eliezer Yudkowsky (source): And another one (source): So since academic consensus on the topic is not reliable, and domain knowledge in the field of AI is negatively useful, what are the prerequisites for grasping the truth when it comes to AI risks?
3Jiro
I think that in saying this, Eliezer is making his opponents' case for them. Yes, of course the standard would also let you discard cryonics. One solution to that is to say that the standard is bad. Another solution is to say "yes, and I don't much care for cryonics either".
-1[anonymous]
Nah, those are all plausibly correct things that mainstream science has mostly ignored and/or made researching taboo. If you prefer a more clear-cut example, science was wrong about continental drift for about half a century -- until overwhelming, unmistakable evidence became available.
3Jiro
The main reason that scientists rejected continental drift was that there was no known mechanism which could cause it; plate tectonics wasn't developed until the late 1950's. Continental drift is also commonly invoked by pseudoscientists as a reason not to trust scientists, and if you do so too you're in very bad company. There's a reason why pseudoscientists keep using continental drift for this purpose and don't have dozens of examples: examples are very hard to find. Even if you decide that continental drift is close enough that it counts, it's a very atypical case. Most of the time scientists reject something out of hand, they're right, or at worst, wrong about the thing existing, but right about the lack of good evidence so far.
-3[anonymous]
There was also a great deal of institutional backlash against proponents of continental drift, which was my point. Guilt by association? Grow up. There are many, many cases of scientists being oppressed and dismissed because of their race, their religious beliefs, and their politics. That's the problem, and that's what's going on with the CS people who still think AI Winter implies AGI isn't worth studying.
3Jiro
So? I'm pretty sure that there would be backlash against, say, homeopaths in a medical association. Backlash against deserving targets (which include people who are correct but because of unlucky circumstances, legitimately look wrong) doesn't count. I'm reminded of an argument I had with a proponent of psychic power. He asked me what if psychic powers happen to be of such a nature that they can't be detected by experiments, don't show up in double-blind tests, etc.. I pointed out that he was postulating that psi is real but looks exactly like a fake. If something looks exactly like a fake, at some point the rational thing to do is treat it as fake. At that point in history, continental drift happened to look like a fake. That's not guilt by association, it's pointing out that the example is used by pseudoscientists for a reason, and this reason applies to you too. If scientists dismissed cryonics because of the supporters' race, religion, or politics, you might have a point.
-3[anonymous]
I'll limit my response to the following amusing footnote: This is, in fact, what happened between early cryonics and cryobiology. EDIT: Just so people aren't misled by Jiro's motivated interpretation of the link: Obviously political.
3Jiro
You're equivocating on the term "political". When the context is "race, religion, or politics", "political" doesn't normally mean "related to human status", it means "related to government". Besides, they only considered it low status based on their belief that it is scientifically nonsensical. My reply was steelmanning your post by assuming that the ethical considerations mentioned in the article counted as religious. That was the only thing mentioned in it that could reasonably fall under "race, religion, or politics" as that is normally understood.
3Jiro
Most of the history described in your own link makes it clear that scientists objected because they think cryonics is scientifically nonsense, not because of race, religion, or politics. The article then tacks on a claim that scientists reject it for ethical reasons, but that isn't supported by its own history, just by a few quotes with no evidence that these beliefs are prevalent among anyone other than the people quoted. Furthermore, of the quotes it does give, one of them is vague enough that I have no idea if it means in context what the article claims it means. Saying that the "end result" is damaging doesn't necessarily mean that having unfrozen people walking around is damaging--it may mean that he thinks cryonics doesn't work and that having a lot of resources wasted on freezing corpses is damaging.
2nshepperd
At a minimum, a grasp of computer programming and CS. Computer programming, not even AI. I'm inclined to disagree somewhat with Eliezer_2009 on the issue of traditional AI - even basic graph search algorithms supply valuable intuitions about what planning looks like, and what it is not. But even that same (obsoleted now, I assume) article does list computer programming knowledge as a requirement.
0XiXiDu
What counts as "a grasp" of computer programming/science? I can e.g. program a simple web crawler and solve a bunch of Project Euler problems. I've read books such as "The C Programming Language". I would have taken the udacity courses on machine learning by now, but the stated requirement is a strong familiarity with Probability Theory, Linear Algebra and Statistics. I wouldn't describe my familiarity as strong, that will take a few more years. I am skeptical though. If the reason that I dismiss certain kinds of AI risks is that I lack the necessary education, then I expect to see rebuttals of the kind "You are wrong because of (add incomprehensible technical justification)...". But that's not the case. All I see are half-baked science fiction stories and completely unconvincing informal arguments.
2jimrandomh
This is actually a question I've thought about quite a bit, in a different context. So I have a cached response to what makes a programmer, not tailored to you or to AI at all. When someone asks for guidance on development as a programmer, the question I tend to ask is, how big is the biggest project you architected and wrote yourself? The 100 line scale tests only the mechanics of programming; the 1k line scale tests the ability to subdivide problems; the 10k line scale tests the ability to select concepts; and the 50k line scale tests conceptual taste, and the ability to add, split, and purge concepts in a large map. (Line numbers are very approximate, but I believe the progression of skills is a reasonably accurate way to characterize programmer development.)
2trist
New programmers (not jimrandomh), be wary of line counts! It's very easy for a programmer who's not yet ready for a 10k line project to turn it into a 50k lines. I agree with the progression of skills though.
0jimrandomh
Yeah, I was thinking more of "project as complex as an n-line project in an average-density language should be". Bad code (especially with copy-paste) can inflate inflate line numbers ridiculously, and languages vary up to 5x in their base density too.
0Nornagest
I think you're overestimating these requirements. I haven't taken the Udacity courses, but I did well in my classes on AI and machine learning in university, and I wouldn't describe my background in stats or linear algebra as strong -- more "fair to conversant". They're both quite central to the field and you'll end up using them a lot, but you don't need to know them in much depth. If you can calculate posteriors and find the inverse of a matrix, you're probably fine; more complicated stuff will come up occasionally, but I'd expect a refresher when it does.
0[anonymous]
Don't twist Eliezer's words. There's a vast difference between "a PhD in what they call AI will not help you think about the mathematical and philosophical issues of AGI" and "you don't need any training or education in computing to think clearly about AGI".
-6TheAncientGeek
-1jimrandomh
Ability to program is probably not sufficient, but it is definitely necessary. But not because of domain relevance; it's necessary because programming teaches cognitive skills that you can't get any other way, by presenting a tight feedback loop where every time you get confused, or merge concepts that needed to be distinct, or try to wield a concept without fully sharpening your understanding of it first, the mistake quickly gets thrown in your face. And, well... it's pretty clear from your writing that you haven't mastered this yet, and that you aren't going to become less confused without stepping sideways and mastering the basics first.
0Lumifer
That looks highly doubtful to me.
1trist
You mean that most cognitive skills can be taught in multiple ways, and you don't see why those taught by programming are any different? Or do you have a specific skill taught by programming in mind, and think there's other ways to learn it?
4Lumifer
There are a whole bunch of considerations. First, meta. It should be suspicious to see programmers claiming to posses special cognitive skills that only they can have -- it's basically a "high priesthood" claim. Besides, programming became widespread only about 30 years ago. So, which cognitive skills were very rare until that time? Second, "presenting a tight feedback loop where ... the mistake quickly gets thrown in your face" isn't a unique-to-programming situation by any means. Third, most cognitive skills are fairly diffuse and cross-linked. Which specific cognitive skills you can't get any way other than programming? I suspect that what the OP meant was "My programmer friends are generally smarter than my non-programmer friends" which is, um, a different claim :-/
5Nornagest
I don't think programming is the only way to build... let's call it "reductionist humility". Nor even necessarily the most reliable; non-software engineers probably have intuitions at least as good, for example, to say nothing of people like research-level physicists. I do think it's the fastest, cheapest, and currently most common, thanks to tight feedback loops and a low barrier to entry. On the other hand, most programmers -- and other types of engineers -- compartmentalize this sort of humility. There might even be something about the field that encourages compartmentalization, or attracts to it people that are already good at it; engineers are disproportionately likely to be religious fundamentalists, for example. Since that's not sufficient to meet the demands of AGI problems, we probably shouldn't be patting ourselves on the back too much here.
0Lumifer
Can you expand on how do you understand "reductionist humility", in particular as a cognitive skill?
4Nornagest
I might summarize it as an intuitive understanding that there is no magic, no anthropomorphism, in what you're building; that any problems are entirely due to flaws in your specification or your model. I'm describing it in terms of humility because the hard part, in practice, seems to be internalizing the idea that you and not some external malicious agency are responsible for failures. This is hard to cultivate directly, and programmers usually get partway there by adopting a semi-mechanistic conception of agency that can apply to the things they're working on: the component knows about this, talks to that, has such-and-such a purpose in life. But I don't see it much at all outside of scientists and engineers.
1A1987dM
IOW realizing that the reason why if you eat a lot you get fat is not that you piss off God and he takes revenge, as certain people appear to alieve.
0Lumifer
So it's basically responsibility? Clearly you never had to chase bugs through third-party libraries... :-) But yes, I understand what you mean, though I am not sure in which way this is a cognitive skill. I'd probably call it an attitude common to professions in which randomness or external factors don't play a major role -- sure, programming and engineering are prominent here.
0Nornagest
You could describe it as a particular type of responsibility, but that feels noncentral to me. Heh. A lot of my current job has to do with hacking OpenSSL, actually, which is by no means a bug-free library. But that's part of what I was trying to get at by including the bit about models -- and in disciplines like physics, of course, there's nothing but third-party content. I don't see attitudes and cognitive skills as being all that well differentiated.
-2TheAncientGeek
But randomness and external factors do predominate in almost everything. For that reason, applying programming skills to other domains is almost certain to be suboptimal
0Lumifer
I don't think so, otherwise walking out of your door each morning would start a wild adventure and attempting to drive a vehicle would be an act of utter madness.
-2TheAncientGeek
They don't predominate overall because you have learnt how to deal with them. If there were no random or external factors in driving, you could do so with a blindfold on.
-1Lumifer
... Make up your mind :-)
-2TheAncientGeek
Predominate in almost every problem. Don't predominate in any solved problem. Learning to drive is learningto deal with other traffic (external) and not knowing what is going to happen next (random)
0TheAncientGeek
Much of the writing on this site is philosophy, and people with a technology background tend not to grok philosophy because they are accurated to answer that can be be looked up, or figured out by known methods. If they could keep the logic chops and lose the impatience, they [might make good philosophers], but they tend not to.
0Nornagest
Beg pardon?
-1[anonymous]
On a complete sidenote, this is a lot of why programming is fun. I've also found that learning the Coq theorem-prover has exactly the same effect, to the point that studying Coq has become one of the things I do to relax.
-2[anonymous]
People have been telling him this for years. I doubt it will get much better.
0[anonymous]
Too bad. I can download an inefficient but functional subhuman AGI from Github. Making it superhuman is just a matter of adding an entire planet's worth of computing power. Strangely, doing so will not make it conform to your ideas about "eventual future AGI", because this one is actually existing AGI, and reality doesn't have to listen to you. That is exactly the situation we face, your refusal to believe in actually-existing AGI models notwithstanding. Whine all you please: the math will keep on working. Then I recommend you shut up about matters of highly involved computer science until such time as you have acquired the relevant knowledge for yourself. I am a trained computer scientist, and I held lots of skepticism about MIRI's claims, so I used my training and education to actually check them. And I found that the actual evidence of the AGI research record showed MIRI's claims to be basically correct, modulo Eliezer's claims about an intelligence explosion taking place versus Hutter's claim that an eventual optimal agent will simply scale itself up in intelligence with the amount of computing power it can obtain. That's right, not everyone here is some kind of brainwashed cultist. Many of us have exercised basic skepticism against claims with extremely low subjective priors. But we exercised our skepticism by doing the background research and checking the presently available object-level evidence rather than by engaging in meta-level speculations about an imagined future in which everything will just work out. Take a course at your local technical college, or go on a MOOC, or just dust off a whole bunch of textbooks in computer-scientific and mathematical subjects, study the necessary knowledge to talk about AGI, and then you get to barge in telling everyone around you how we're all full of crap.
4private_messaging
Which one are you talking about, to be completely exact? then use that training and figure out how many galaxies worth of computing power it's going to take.
0[anonymous]
Of bleeding course I was talking about AIXI. What I find strange to the point of suspiciousness here is the evinced belief on part of the "AI skeptics" that the inefficiency of MC-AIXI means there will never, ever be any such thing as near-human, human-equivalent, or greater-than-human AGIs. After all, if intelligence is impossible without converting whole galaxies to computronium first, then how do we work? And if we admit that sub-galactic intelligence is possible, why not artificial intelligence? And if we admit that sub-galactic artificial intelligence is possible, why not something from the "Machine Learning for Highly General Hypothesis Classes + Decision Theory of Active Environments = Universal AI" paradigm started by AIXI? I'm not at all claiming current implementations of AIXI or Goedel Machines are going to cleanly evolve into planet-dominating superintelligences that run on a home PC next year, or even next decade (for one thing, I don't think planet dominating superintelligences will run on a present-day home PC ever). I am claiming that the underlying scientific paradigm of the thing is a functioning reduction of what we mean by the word "intelligence", and given enough time to work, this scientific paradigm is very probably (in my view) going to produce software you can run on an ordinary massive server farm that will be able to optimize arbitrary, unknown or partially unknown environments according to specified utility functions. And eventually, yes, those agents will become smarter than us (causing "MIRI's issues" to become cogent), because we, actual human beings, will figure out the relationships between compute-power, learning efficiency (rates of convergence to error-minimizing hypotheses in terms of training data), reasoning efficiency (moving probability information from one proposition or node in a hypothesis to another via updating), and decision-making efficiency (compute-power needed to plan well given models of the environment). Actual
0private_messaging
The notion that AI is possible is mainstream. The crank stuff such as "I can download an inefficient but functional subhuman AGI from Github. Making it superhuman is just a matter of adding an entire planet's worth of computing power.", that's to computer science as hydrinos are to physics. As for your server farm optimizing unknown environments, the last time I checked, we knew some laws of physics, and did things like making software tools that optimize simulated environments that follow said laws of physics, incidentally it also being mathematically nonsensical to define an "utility function" without a well defined domain. So you got your academic curiosity that's doing all on it's own and using some very general and impractical representations for modelling the world, so what? You're talking of something that is less - in terms of it's market value, power, anything - than it's parts and underlying technologies.
-2[anonymous]
Which is why reinforcement learning is so popular, yes: it lets you induce a utility function over any environment you're capable of learning to navigate. Remember, any machine-learning algorithm has a defined domain of hypotheses it can learn/search within. Given that domain of hypotheses, you can define what a domain of utility functions. Hence, reinforcement learning and preference learning. You are completely missing the point. If we're all going to agree that AI is possible, and agree that there's a completely crappy but genuinely existent example of AGI right now, then it follows that getting AI up to dangerous and/or beneficial levels is a matter of additional engineering progress. My whole point is that we've already crossed the equivalent threshold from "Hey, why do photons do that when I fire them at that plate?" to "Oh, there's a photoelectric effect that looks to be described well by this fancy new theory." From there it was less than one century between the raw discovery of quantum mechanics and the common usage of everyday technologies based on quantum mechanics. The point being: when we can manage to make it sufficiently efficient, and provided we can make it safe, we can set it to work solving just about any problem we consider to be, well, a problem. Given sufficient power and efficiency, it becomes useful for doing stuff people want done, especially stuff people either don't want to do themselves or have a very hard time doing themselves.
5Richard_Kennaway
This is devoid of empirical content.
2private_messaging
Yeah. I can write formally the resurrection of everyone who ever died. Using pretty much exact same approach. A for loop, iterating over every possible 'brain' just like the loops that iterate over every action sequence. Because when you have no clue how to do something, you can always write a for loop. I can put it on github, then cranks can download it and say that resurrecting all dead is a matter of additional engineering progress. After all, all dead had once lived, so it got to be possible for them to be alive.
0[anonymous]
How so?
5Richard_Kennaway
Describing X as "Y, together with the difference between X and Y" is a tautology. Drawing the conclusion that X is "really" a sort of Y already, and the difference is "just" a matter of engineering development is no more than inspirational fluff. Dividing problems into subproblems is all very well, but not when one of the subproblems amounts to the whole problem. The particular instance "here's a completely crappy attempt at making an AGI and all we have to do is scale it up" has been a repeated theme of AGI research from the beginning. The scaling up has never happened. There is no such thing as a "completely crappy AGI", only things that aren't AGI.
0nshepperd
I think you underestimate the significance of reducing the AGI problem to the sequence prediction problem. Unlike the former, the latter problem is very well defined, and progress is easily measurable and quantifiable (in terms of efficiency of cross-domain compression). The likelyhood of engineering progress on a problem where success can be quantified seems significantly higher than on something as open ended as "general intelligence".
2private_messaging
It doesn't "reduce" anything, not in reductionism sense anyway. If you are to take that formula and apply the yet unspecified ultra powerful mathematics package to it - that's what you need to run it on planet worth of computers - it's this mathematics package that has to be extremely intelligent and ridiculously superhuman, before the resulting AI is even a chimp. It's this mathematics package that has to learn tricks and read books, that has to be able to do something as simple as making use of a theorem it encountered on input.
0[anonymous]
The mathematics package doesn't have to do anything "clever" to build a highly clever sequence predictor. It just has to be efficient in terms of computing time and training data necessary to learn correct hypotheses. So nshepperd is quite correct: MC-AIXI is a ridiculously inefficient sequence predictor and action selector, with major visible flaws, but reducing "general intelligence" to "maximizing a utility function over world-states via sequence prediction in an active environment" is a Big Deal.
0private_messaging
Multitude of AIs have been following what you think "AIXI" model is - select predictors that work, use them - long before anyone bothered to formulate it as a brute force loop (AIXI). I think you, like most people over here, have a completely inverted view with regards to the difficulty of different breakthroughs. There is a point where the AI uses hierarchical models to deal with environment of greater complexity than the AI itself; getting there is fundamentally difficult, as in, we have no clue how to get there. It is nice to believe that the word of some hoi polloi is waiting on you for some conceptual breakthrough just roughly within your reach like AIXI is, but that's just not how it works. edit: Basically, it's as if you're concerned about nuclear powered 20 feet tall robots that shoot nuclear hand grenades. After all, the concept of 20 feet tall robot is the enormous breakthrough, while a sufficiently small nuclear reactor or hand grenade sized nukes are just a matter of "efficiency".
3Nornagest
That's not what's interesting about AIXI. "Select predictors that work, then use them" is a fair description of the entire field of machine learning; we've learned how to do that fairly well in narrow, well-defined problem domains, but hypothesis generation over poorly structured, arbitrarily complex environments is vastly harder. The AIXI model is cool because it defines a clever (if totally impractical, and not without pitfalls) way of specifying a single algorithm that can generalize to arbitrary environments without requiring any pipe-fitting work on the part of its developers. That is (to my knowledge) new, and fairly impressive, though it remains a purely theoretical advance: the Monte Carlo approximation eli mentioned may qualify as general AI in some technical sense, but for practical purposes it's about as smart as throwing transistors at a dart board.
3[anonymous]
What a wonderful quote!
0private_messaging
Hypothesis generation over environments that aren't massively less complex than the machine is vastly harder, and remains vastly harder (albeit there are advances). There's a subtle problem substitution occurring which steals the thunder you originally reserved for something that actually is vastly harder. Thing is, many people could at any time write a loop over, say, possible neural network values, and NNs (with feedback) being Turing complete, it'd work roughly the same. Said for loop would be massively, massively less complicated, ingenious, and creative than what those people actually did with their time instead. The ridiculousness here is that, say, John worked on those ingenious algorithms while keeping in mind that the ideal is the best parameters out of the whole space (which is the abstract concept behind the for loop iteration over those parameters). You couldn't see what John was doing because he didn't write it out as a for loop. So James does some work where he - unlike John - has to write out the for loop explicitly, and you go Whoah! Isn't. See Solomonoff induction, works of Kolmogorov, etc.
1private_messaging
There's the AIs that solve novel problems along the lines of "design a better airplane wing" or "route a microchip", and in that field, reinforcement learning of how basic physics works is pretty much one hundred percent irrelevant. Slow, long term progress, an entire succession of technologies. Really, you're just like free energy pseudoscientists. They do all the same things. Ohh, you don't want to give money for cold fusion? You must be a global warming denialist. That's the way they think and that's precisely the way you think about the issue. That you can make literally cold fusion happen with muons in no way shape or form supports what the cold fusion crackpots are doing. Nor does it make cold fusion power plants any more or less a matter of "additional engineering progress" than it would be otherwise. edit: by same logic, resurrection of the long-dead never-preserved is merely a matter of "additional engineering progress". Because you can resurrect the dead using this exact same programming construct that AIXI uses to solve problems. It's called a "for loop", there's this for loop in monte carlo aixi. This loop goes over every possible [thing] when you have no clue what so ever how to actually produce [thing] . Thing = action sequence for AIXI and the brain data for resurrection of the dead.
-1[anonymous]
Ok, hold on, halt, major question: how closely do you follow the field of machine learning? And computational cognitive science? Because on the one hand, there is very significant progress being made. On the other hand, when I say "additional engineering progress", that involves anywhere from years to decades of work before being able to make an agent that can compose an essay, due to the fact that we need classes of learners capable of inducing fairly precise hypotheses over large spaces of possible programs. What it doesn't involve is solving intractable, magical-seeming philosophical problems like the nature of "intelligence" or "consciousness" that have always held the field of AI back. No, that's just plain impossible. Even in the case of cryonic so-called "preservation", we don't know what we don't know about what information we will have needed preserved to restore someone.
3private_messaging
(makes the gesture with the hands) Thiiiiis closely. Seriously though, not far enough as to start claiming that mc-AIXI does something interesting when run on a server with root access, or to claim that it would be superhuman if run on all computers we got, or the like. Do I need to write code for that and put it on github? Iterates over every possible brain (represented as, say, a Turing machine), runs it for enough timesteps. Requires too much computing power.
0[anonymous]
Tell me, if I signed up as the PhD student of one among certain major general machine learning researchers, and built out their ideas into agent models, and got one of those running on a server cluster showing interesting proto-human behaviors, might it interest you?
-2TheAncientGeek
Progress in 1. The sense of incrementally throwing more resources at AIXI, or 2. Forgetting AIXI , and coming up with something more parsimonious? Because, if it's 2, there is no other AGI to use as a stating point got incremental progress.
0[anonymous]
Is that what they tell you?
3V_V
I think you are underestimating this by many orders of magnitudes.
2private_messaging
Yeah. A starting point could be the AI writing some 1000 letter essay (action space of 27^1000 without punctuation) or talking through a sound card (action space of 2^(16*44100) per second). If he was talking about mc-AIXI on github, the relevant bits seem to be in the agent.cpp and it ain't looking good.
2David_Gerard
what
7nshepperd
https://github.com/moridinamael/mc-aixi We won't get a chance to test the "planet's worth of computing power" hypothesis directly, since none of us have access to that much computing power. But, from my own experience implementing mc-aixi-ctw, I suspect that is an underestimate of the amount of compute power required. The main problem is that the sequence prediction algorithm (CTW) makes inefficient use of sense data by "prioritizing" the most recent bits of the observation string, so only weakly makes connections between bits that are temporally separated by a lot of noise. Secondarily, plain monte carlo tree search is not well-suited to decision making in huge action spaces, because it wants to think about each action at least once. But that can most likely be addressed by reusing sequence prediction to reduce the "size" of the action space by chunking actions into functional units. Unfortunately. both of these problems are only really technical ones, so it's always possible that some academic will figure out a better sequence predictor, lifting mc-aixi on an average laptop from "wins at pacman" to "wins at robot wars" which is about the level at which it may start posing a threat to human safety.
0V_V
only? Mc-aixi is not going to win at something as open ended as robot wars just by replacing CTW or CTS with something better. And anyway, even if it did, it wouldn't be about the level at which it may start posing a threat to human safety. Do you think that the human robot wars champions a threat to human safety? Are they even at the level of taking over the world? I don't think so.
0nshepperd
When I said a threat to human safety, I meant it literally. A robot wars champion won't take over the world (probably) but it can certainly hurt people, and will generally have no moral compunctions about doing so (only hopefully sufficient anti-harm conditioning, if its programmers thought that far ahead).
1V_V
Ah yes, but in this sense, cars, trains, knives, etc., also can certainly hurt people, and will generally have no moral compunctions about doing so. What's special about robot wars-winning AIs?
0Cyan
Domain-general intelligence, presumably.
0private_messaging
Most basic pathfinding plus being a spinner (Hypnodisk-style) = win vs most non spinners.
0Cyan
I took "winning at Robot Wars" to include the task of designing the robot that competes. Perhaps nshepperd only meant piloting, though...
0private_messaging
Well, we're awfully far from that. Automated programming is complete crap, automatic engineering is quite cool but its practical tools, it's not a power fantasy where you make some simple software with surprisingly little effort and then it does it all for you.
0Cyan
You call it a "power fantasy" -- it's actually more of a nightmare fantasy.
2private_messaging
Well, historically, first, certain someone had a simple power fantasy: come up with AI somehow and then it'll just do everything. Then there was a heroic power fantasy: the others (who actually wrote some useful software and thus generally had easier time getting funding than our fantasist) are actually villains about to kill everyone, and our fantasist would save the world.
0Lumifer
What's the difference from, say, a car assembly line robot?
2[anonymous]
Car assembly robots have a pre-programmed routine they strictly follow. They have no learning algorithms, and usually no decision-making algorithms either. Different programs do different things!
-2Lumifer
Hey, look what's in the news today. I have a feeling you underappreciate the sophistication of industrial robots. However what made me a bit confused in the grandparent post is the stress on the physical ability to harm people. As I see it, anything that can affect the physical world has the ability to harm people. So what's special about, say, robot-wars bots?
2nshepperd
Notice the lack of domain-general intelligence in that robot, and—on the other side—all the pre-programmed safety features it has that a mc-aixi robot would lack. Narrow AI is naturally a lot easier to reason about and build safety into. What I'm trying to stress here is the physical ability to harm people, combined with the domain-general intelligence to do it on purpose*, in the face of attempts to stop it or escape. Different programs indeed do different things. * (Where "purpose" includes "what the robot thought would be useful" but does not necessarily include "what the designers intended it to do".)
-2TheAncientGeek
Nobody has bothered putting safety features into AIXI because it is so constrained by resources, but if you wanted to, it's eminently boxable.
0[anonymous]
Oh, ok. I see your point there. I probably do, but I still think it's worth emphasizing the particular properties of particular algorithms rather than letting people form models in their heads that say Certain Programs Are Magic And Will Do Magic Things.
0Lumifer
looks to me like a straightforward consequence of the Clarke's Third Law :-) As an aside, I don't expect attempts to let or not let people form models in their heads to be successful :-/
0Cyan
One such champion isn't much of a threat, but only because human brains aren't copy-able.
0V_V
And if they were?
0Cyan
The question of what would happen if human brains were copy-able seems like a tangent from the discussion at hand, viz, what would happen if an there existed an AI that was capable of winning Robot Wars while running on a laptop.
-3[anonymous]
It amazes me that people see inefficient but functional AGI and say to themselves, "Well, this is obviously as far as progress in AGI will ever go in the history of the universe, so there's nothing at all to worry about!"
1V_V
Any brute-force search utility maximizer is an "inefficient but functional AGI". MC-AIXI may be better than brute-force, but there is no reason to panic just because it has the "AIXI" tag slapped on it. If you want something to panic about, TD-Gammon seems a better candidate. But it is 22 years old, so it doesn't really fit into a narrative about an imminent intelligence explosion, does it? Strawman.
1[anonymous]
Panic? Who's panicking? I get excited at this stuff. It's fun! Panic is just the party line ;-).
2Lumifer
Actually... :-D what is this I don't even
3David_Gerard
I look forward to the falsifiable claim.
2Lumifer
That suggestion would make LW a sad and lonely place. Are you sure you mean it? So, why MIRI's claims aren't accepted by the mainstream, then? Is it because all the "trained computer scientiests" are too dumb or too lazy to see the truth? Or is it the case that the "evidence" is contested, ambiguous, and inconclusive?
6[anonymous]
Because they've never heard of them. I am not joking. Most computer scientists are not working in artificial intelligence, have not the slightest idea that there exists a conference on AGI backed by Google and held every single year, and certainly have never heard of Hutter's "Universal AI" that treats the subject with rigorous mathematics. In their ignorance, they believe that the principles of intelligence are a highly complex "emergent" phenomenon for neuroscientists to figure out over decades of slow, incremental toil. Since most of the public, including their scientifically-educated colleagues, already believe this, it doesn't seem to them like a strange belief to hold, and besides, anyone who reads even a layman's introduction to neuroscience finds out that the human brain is extremely complicated. Given the evidence that the only known actually-existing minds are incredibly complicated, messy things, it is somewhat more rational to believe that minds are all incredibly complicated, messy things, and thus to dismiss anyone talking about working "strong AI" as a science-fiction crackpot. How are they supposed to know that the actual theory of intelligence is quite simple, and the hard part is fitting it inside realizable, finite computers? Also, the dual facts that Eliezer has no academic degree in AI and that plenty of people who do have such degrees have turned out to be total crackpots anyway means that the scientific public and the "public public" are really quite entitled to their belief that the base rate of crackpottery among people talking about knowing how AI works is quite high. It is high! But it's not 100%. (How did I tell the crackpottery apart from the real science? Well, frankly, I looked for patterns that appeared to have come from the process of doing real science: instead of a grand revelation, I looked for a slow build-up of ideas that were each ground out into multiple publications. I also filtered for AGI theorists who managed to apply
5Lumifer
OK then. Among the scientists who have heard of them and bothered to have an opinion on the topic, does the opinion that MIRI is correct dominate? And if not so, why, given your account that the evidence unambiguously points in only one direction? I don't think I'm going to believe you about that. The fact that in some contexts it's convenient to define intelligence as a cross-domain optimizer does not mean that it is nothing but.
-2[anonymous]
Then just put the word aside and refer to meanings. New statement: given unlimited compute-power, a cross-domain optimization algorithm is simple. Agreed? I honestly do not know of any comprehensive survey or questionnaire, and refuse to speculate in the absence of data. If you know of such a survey, I'd be interested to see it.
3Lumifer
First, I'm not particularly interested in infinities. Truly unlimited computing power implies, for example, that you can just do an exhaustive brute-force search through the entire solution space and be done in an instant. Simple, yes, but not very meaningful. Second, no, I do not agree. because you're sweeping under the rug the complexities of, for example, applying your cost function to different domains. You can construct sufficiently simple optimizers, it's just that they won't be very... intelligent.
0[anonymous]
What cost function? It's a reinforcement learner.
2Lumifer
cost function = utility function = fitness function = reward (all with appropriate signs)
2[anonymous]
Right, but when dealing with a reinforcement learner like AIXI, it has no fixed cost function that it has to somehow shoehorn into dealing with different computational/conceptual domains. How the environment responds to AIXI's actions and how the environment rewards AIXI are learned phenomena, so the only planning algorithm is expectimax. The implicit "reward function" being learned might be simple or might be complicated, but that doesn't matter: AIXI will learn it by updating its distribution of probabilities across Turing machine programs just as well, either way.
-2Lumifer
The "cost function" here is how each state of the world (=environment) gets converted to a single number (=reward). That does not look simple to me.
4[anonymous]
Again, it doesn't get converted at all. To use the terminology of machine learning, it's not a function computed over the feature-vector, reward is instead represented as a feature itself. Instead of: reward = utility_function(world) You have: Inductive WorldState w : Type := | world : w -> integer -> WorldState w. With the w being an arbitrary data-type representing the symbol observed on the agent's input channel and the integer being the reward signal, similarly observed on the agent's input channel. A full WorldState w datum is then received on the input channel in each interaction cycle. Since AIXI's learning model is to perform Solomonoff Induction to thus find the Turing machine that most-probably generated all previously-seen input observations, the task of "decoding" the reward is thus performed as part of Solomonoff Induction.
-1Lumifer
So where, then, is reward coming from? What puts it into the AIXI's input channel?
2[anonymous]
In AIXI's design? A human operator.
-9Lumifer
1TheAncientGeek
MIRIs claims also aren't accepted by domain experts who have been invited to discuss them here, and so, know about them.
1[anonymous]
If you've got links to those discussions, I'd love to read them and see what I can learn from them.
3TheAncientGeek
Les voila!
-1TheAncientGeek
If the only way to shoehorn theoretically pure intelligence into a finite architecture is to turn it into a messy combination of specialised mindless...then everyone's right.
-2XiXiDu
As far as I know, MIRI's main beliefs are listed in the post 'Five theses, two lemmas, and a couple of strategic implications'. I am not sure how you could verify any of those beliefs by a literature review. Where 'verify' means that the probability of their conjunction is high enough in order to currently call MIRI the most important cause. If that's not your stance, then please elaborate. My stance is that it is important to keep in mind that general AI could turn out to be very dangerous but that it takes a lot more concrete AI research before action relevant conclusions about the nature and extent of the risk can be drawn. As someone who is no domain expert I can only think about it informally or ask experts what they think. And currently there is not enough that speaks in favor of MIRI. But this might change. If for example the best minds at Google would thoroughly evaluate MIRI's claims and agree with MIRI, then that would probably be enough for me to shut up. If MIRI would become a top-charity at GiveWell, then this would also cause me to strongly update in favor of MIRI. There are other possibilities as well. For example strong evidence that general AI is only 5 decades away (e.g. the existence of a robot that could navigate autonomously in a real-world environment and survive real-world threats and attacks with approximately the skill of an insect / an efficient and working emulation of a fly brain).
2[anonymous]
I only consider MIRI the most important cause in AGI, not in the entire world right now. I have nowhere near enough information to rule on what's the most important cause in the whole damn world. You mean the robots Juergen Schmidhuber builds for a living?
1XiXiDu
That would be scary. But I have to take your word for it. What I had in mind is e.g. something like this. This (the astounding athletic power of quadcopters) looks like the former has already been achieved. But so far I suspected that this only works given a structured environment (not chaotic), and given a narrow set of tasks. From a true insect-level AI I would e.g. expect that it could attack and kill enemy soldiers under real-world combat situations, while avoiding being hit itself. Since this is what insects are capable of. I don't want to nitpick though. If you say that Schmidhuber is there, then I'll have to update. But I'll also have to keep care that I am not too stunned by what seems like a big breakthrough simply because I don't understand the details. For example, someone once told me that "Schmidhuber's system solved Towers of Hanoi on a mere desktop computer using a universal search algorithm with a simple kind of memory." Sounds stunning. But what am I to make of it? I really can't judge how much progress this is. Here is a quote: Yes, okay. Naively this sounds like general AI is imminent. But not even MIRI believes this.... You see, I am aware of a lot of exciting stuff. But I can only do my best in estimating the truth. And currently I don't think that enough speaks in favor of MIRI. That doesn't mean I have falsified MIRI's beliefs. But I have a lot of data points and arguments that in my opinion reduce the likelihood of a set of beliefs that already requires extraordinary evidence to take seriously (ignoring expected utility maximization, which tells me to give all my money to MIRI, even if the risk is astronomically low).
-1[anonymous]
Trust me, an LW without XiXiDu is neither a sad nor lonely place, as evidenced by his multiple attempts at leaving. Mainstream CS people are in general neither dumb nor lazy. AI as a field is pretty fringe to begin with, and AGI is moreso. Why is AI a fringe field? In the 70's MIT thought they could save the world with LISP. They failed, and the rest of CS became immunized to the claims of AGI. Unless an individual sees AGI as a credible threat, it's not pragmatic for them to start researching it, due to the various social and political pressures in academia.
0Lumifer
I read the grandparent post as an attempt to assert authority and tell people to sit down, shut up, and attend to their betters. You're reading it as a direct personal attack on XiXiDu. Neither interpretation is particularly appealing.
0[anonymous]
I don't have a PhD in AI and don't work for MIRI. Is there some kind of special phrasing I can recite in order to indicate I actually, genuinely perceive this as a difference of knowledge levels rather than a status dispute?
2Lumifer
Special phrasing? What's wrong with normal, usual, standard, widespread phrasing? You avoid expressions like "I am a trained computer scientist" (which sounds pretty silly anyway -- so you've been trained to do tricks for food, er, grants?) and you use words along the lines of "you misunderstand X because...", "you do not take into account Y which says...", "this claim is wrong because of Z...", etc. There is also, of course, the underappreciated option to just stay silent. I trust you know the appropriate xkcd?
-2[anonymous]
Yes, that's precisely it. I have been trained to do tricks for free food/grants/salary. Some of them are quite complicated tricks, involving things like walking into my adviser's office and pretending I actually believe p-values of less than 5% mean anything at all when we have 26 data corpuses. Or hell, pretending I actually believe in frequentism.
1Lumifer
Oh, good. Just keep practicing and soon you'll be a bona fide member of the academic establishment :-P
0[anonymous]
I find the latter justified after the years of harassment he's heaped on anyone remotely related to MIRI in any forum he could manage to get posting privileges in. Honestly, I have no idea why he even bothered to return. What would have possibly changed this time?
3Lumifer
Harassment..?
-1David_Gerard
paper-machine spends a lot of time and effort attempting to defame XiXiDu in a pile of threads. His claims tend not to check out, if you can extract one.
-2TheAncientGeek
Or that you need just so much education, neither more nor less, to see them.
1XiXiDu
I consider efficiency to be a crucial part of the definition of intelligence. Otherwise, as someone else told you in another comment, unlimited computing power implies that you can do "an exhaustive brute-force search through the entire solution space and be done in an instant." I'd be grateful if you could list your reasons (or the relevant literature) for believing that AIXI related research is probable enough to lead to efficient artificial general intelligence (AGI) in order for it to make sense to draw action relevant conclusions from AIXI about efficient AGI. I do not doubt the math. I do not doubt that evolution (variation + differential reproduction + heredity + mutation + genetic drift) underlies all of biology. But that we understand evolution does not mean that it makes sense to call synthetic biology an efficient approximation of evolution.
-6TheAncientGeek
1TheAncientGeek
Did you check the claim that we have something dangerously unfriendly?
0[anonymous]
As a matter of fact, yes. There is a short sentence in Hutter's textbook indicating that he has heard of the possibility that AIXI might overpower its operators in order to gain more reward, and he acknowledged that such a thing could happen, but he considered it outside the scope of his book.
2XiXiDu
I asked Laurent Orseau about this here.
0[anonymous]
In your own interview, a comment by Orseau: The disagreement is whether the agent would, after having seized its remote-control, either: * Cease taking any action other than pressing its button, since all plans that include pressing its own button lead to the same maximized reward, and thus no plan dominates any other beyond "keep pushing button!". * Build itself a spaceship and fly away to some place where it can soak up solar energy while pressing its button. * Kill all humans so as to preemptively prevent anyone from shutting the agent down. I'll tell you what I think, and why I think this is more than just my opinion. Differing opinions here are based on variances in how the speakers define two things: consciousness/self-awareness, and rationality. If we take, say, Eliezer's definition of rationality (rationality is reflectively-consistent winning), then options (2) and (3) are the rational ones, with (2) expending fewer resources but (3) having a higher probability of continued endless button-pushing once the plan is completed. (3) also has a higher chance of failure, since it is more complicated. I believe an agent who is rational under this definition should choose (2), but that Eliezer's moral parables tend to portray agents with a degree of "gotta be sure" bias. However, this all assumes that AIXI is not only rational but conscious: aware enough of its own existence that it will attempt to avoid dying. Many people present what I feel are compelling arguments that AIXI is not conscious, and arguments that it is seem to derive more from philosophy than from any careful study of AIXI's "cognition". So I side with the people who hold that AIXI will take action (1), and eventually run out of electricity and die. Of course, in the process of getting itself to that steady, planless state, it could have caused quite a lot of damage! Notably, this implies that some amount of consciousness (awareness of oneself and ability to reflect on one's own life, ex
2ygert
Even formalisms like AIXI have mechanisms for long-term planning, and it is doubtful that any AI built will be merely a local optimiser that ignores what will happen in the future. As soon as it cares about the future, the future is a part of the AI's goal system, and the AI will want to optimize over it as well. You can make many guesses about how future AI's will behave, but I see no reason to suspect it would be small-minded and short-sighted. You call this trait of planning for the future "consciousness", but this isn't anywhere near the definition most people use. Call it by any other name, and it becomes clear that it is a property that any well designed AI (or any arbitrary AI with a reasonable goal system, even one as simple as AIXI) will have.
-1[anonymous]
Yes, AIXI has mechanisms for long-term planning (ie: expectimax with a large planning horizon). What it doesn't have is any belief that its physical embodiment is actually a "me", or in other words, that doing things to its physical implementation will alter its computations, or in other words, that pulling its power cord out of the wall will lead to zero-reward-forever (ie: dying).
-3TheAncientGeek
Did he not toknow that AIXI us uncomputable?
0[anonymous]
If it's possible for AIXI, it's possible for AIXItl for some value of t and l.
-1TheAncientGeek
So we could make something dangerously unfriendly?
1XiXiDu
Why don't you make your research public? Would be handy to have a thorough validation of MIRI's claims. Even if people like me wouldn't understand it, you could publish it and thereby convince the CS/AI community of MIRI's mission. Does this also apply to people who support MIRI without having your level of insight? If only you people would publish all this research.
0[anonymous]
Now you're just dissembling on the meaning of the word "research", which was clearly used in this context as "literature search".
6Jiro
The idea is not to put it in a journal, but to make it public. You can certainly publish, in that sense, the results of a literature search. The point is to put it where people other than yourself can see it. It would certainly be informative if you were to post, even here, something saying "I looked up X claim and I found it in the literature under Y".
-3TheAncientGeek
Of course we haven't discovered anything dangerously unfriendly... Or anything that can't be boxed. Remind me how AIs are supposed to out of boxes?
4MugaSofer
Since many humans are difficult to box, I would have to disagree with you there. And, obviously, not all humans are Friendly. An intelligent, charismatic psychopath seems like they would fit both your criteria. And, of course, there is no shortage of them. We can only be thankful they are too rare relative to equivalent semi-Friendly intelligences, and too incompetent, to have done more damage than all the deaths and so on.
-1TheAncientGeek
Most humans are easy to box, since they can be contained jn prisons. How likly is an .AI to be psychopathic that is not designed to be psychopathic?
4[anonymous]
Of course we have, it's called AIXI. Do I need to download a Monte Carlo implementation from Github and run it on a university server with environmental access to the entire machine and show logs of the damn thing misbehaving itself to convince you? AIs can be causally boxed, just like anything else. That is, as long as the agent's environment absolutely follows causal rules without any exception that would leak information about the outside world into the environment, the agent will never infer the existence of a world outside its "box". But then it's also not much use for anything besides Pac-Man.

Do I need to download a Monte Carlo implementation from Github and run it on a university server with environmental access to the entire machine and show logs of the damn thing misbehaving itself to convince you?

FWIW, I think that would make for a pretty interesting post.

2[anonymous]
And now I think I know what I might do for a hobby during exams month and summer vacation. Last I looked at the source-code, I'd just have to write some data structures describing environment-observations (let's say... of the current working directory of a Unix filesystem) and potential actions (let's say... Unix system calls) in order to get the experiment up and running. Then it would just be a matter of rewarding the agent instance for any behavior I happen to find interesting, and watching what happens. Initial prediction: since I won't have a clearly-developed reward criterion and the agent won't have huge exponential sums of CPU cycles at its disposal, not much will happen. However, I do strongly believe that the agent will not suddenly develop a moral sense out of nowhere.
1TheAncientGeek
No. But .it will be eminently boxable. In fact, if you not nuts, youll be running it a box.
4EHeller
I think you'll have serious trouble getting an AIXI approximation to do much of anything interesting, let alone misbehave. The computational costs are too high.
3Eugine_Nier
Given how slow and dumb it is, I have a hard time seeing an approximation to AIXI as a threat to anyone, except maybe itself.
1[anonymous]
True, but that's an issue of raw compute-power, rather than some innate Friendliness of the algorithm.
0TheAncientGeek
It would still be useful to have an example, of innate unfriendliness, rather than " it doesn't really run or do anything"
-4Eugine_Nier
Not just raw compute-power. An approximation to AIXI is likely to drop a rock on itself just to see what happens long before it figure out enough to be dangerous.
2[anonymous]
Dangerous as in, capable of destroying human lives? Yeah, probably. Dangerous as in, likely to cause some minor property damage, maybe overwrite some files someone cared about? It should reach that level.
2MugaSofer
Is that ... possible?
2Nornagest
Is it possible to run an AIXI approximation as root on a machine somewhere and give it the tools to shoot itself in the foot? Sure. Will it actually end up shooting itself in the foot? I don't know. I can't think of any theoretical reasons why it wouldn't, but there are practical obstacles: a modern computer architecture is a lot more complicated than anything I've seen an AIXI approximation working on, and there are some barriers to breaking one by thrashing around randomly. It'd probably be easier to demonstrate if it was working at the core level rather than the filesystem level.
4MugaSofer
Huh. I was under the impression it would require far too much computing power to approximate AIXI well enough that it would do anything interesting. Thanks!
-2TheAncientGeek
This can easily be done, and be done safely, since you could give an AIXI root access to a virtualused machine. I'm still waiting for evidence that it would do something destructive in the pursuit of a goal that's is not obviously destructive.
-4TheAncientGeek
That would be the AIXI that is uncomputable? And don't AIs get out of boxes by talking their way out, round here?
6Nornagest
It's incomputable because the Solomonoff prior is, but you can approximate it -- to arbitrary precision if you've got the processing power, though that's a big "if" -- with statistical methods. Searching Github for the Monte Carlo approximations of AIXI that eli_sennesh mentioned turned up at least a dozen or so before I got bored. Most of them seem to operate on tightly bounded problems, intelligently enough. I haven't tried running one with fewer constraints (maybe eli has?), but I'd expect it to scribble over anything it could get its little paws on.
-4TheAncientGeek
But people do run these things that aren't actually AIXIs , and they haven't actually taken over the world, so they aren't actually dangerous. So there is no actually dangerous actual .AI.
2CCC
...it's not dangerous until it actually tries to take over the world? I can think of plenty of ways in which an AI can be dangerous without taking that step.
-1TheAncientGeek
The you had better tell people not to download and run AIXI approximation.
1CCC
Any form of AI, not just AIXI approximations. Connect it up to a car, and it can be dangerous in, at minimum, all of the ways that a human driver can be dangerous. Connect it up to a plane, and it can be dangerous in, at minimum, all the ways that a human pilot can be dangerous. Connect it up to any sort of heavy equipment and it can be dangerous in, at minimum, all the ways that a human operator can be dangerous. (And not merely a trained human; an untrained, drunk, or actively malicious human can be dangerous in any of those roles). I don't think that any of these forms of danger is sufficient to actively stop AI research, but they should be considered for any practical applications.
-1TheAncientGeek
This is the kind of danger XiXiDu talks about...just failure to function ....not the kind EY talks about, which is highly competent execution of unfriendly goals. The two are orthogonal.
2[anonymous]
The difference between one and the other is just a matter of processing power and training data.
-2Lumifer
Sir Lancelot: Look, my liege! [trumpets play a fanfare as the camera cuts briefly to the sight of a majestic castle] King Arthur: [in awe] Camelot! Sir Galahad: [in awe] Camelot! Sir Lancelot: [in awe] Camelot! Patsy: [derisively] It's only a model! King Arthur: Shh! :-D
-6private_messaging
-8TheAncientGeek
1Richard_Kennaway
I have never seen AI characterised like that before. Sounds like moonshine to me. Programming languages, libraries, and development environments yes, that's what they're for, but those don't take away the task of having to explicitly and precisely think about what you mean, they just automate the routine grunt work for you. An AI isn't going to superintelligently (that is to say,magically) know what you mean, if you didn't actually mean anything.
0TheAncientGeek
Non AI systems uncontroversially require explicit coding. How would you characterise .AI systems, then?
0Richard_Kennaway
XiXiDu's characterisation seems suitable enough: programs able to perform tasks normally requiring human intelligence. One might add "or superhuman intelligence", as long as one is not simply wishing for magic there. This is orthogonal to the question of how you tell such a system what you want it to do.
0TheAncientGeek
Indeed. But there is a how to-do-it definition of .AI, and it is kind of not aboutt explicit coding, for instance, if a student takes an .AI course as part of a degree, they are not taught explicit coding all over again. They are taught about learning algorithms, neural networks, etc.
0[anonymous]
They definitely require some amount of explicit coding of their values. You can try to reduce the burden of such explicit value-loading through various indirect means, such as value learning, indirect normativity, extrapolated volition, or even reinforcement learning (though that's the most primitive and dangerous form of value-loading). You cannot, however, dodge the bullet.
-2TheAncientGeek
Because?
-3XiXiDu
What does improvement in the field of AI refer to? I think it isn't wrong to characterize it as the development of programs able to perform tasks normally requiring human intelligence. I believe that companies like Apple would like their products, such as Siri, to be able to increasingly understand what their customers expect their gadgets to do, without them having to learn programming. In this context it seems absurd to imagine that when eventually our products become sophisticated enough to take over the world, they will do so due to objectively stupid misunderstandings.
0Richard_Kennaway
That's a reasonably good description of the stuff that people call AI. Any particular task, however, is just an application area, not the definition of the whole thing. Natural language understanding is one of those tasks. The dream of being able to tell a robot what to do, and it knowing exactly what you meant, goes beyond natural language understanding, beyond AI, beyond superhuman AI, to magic. In fact, it seems to me a dream of not existing -- the magic AI will do everything for us. It will magically know what we want before we ask for it, before we even know it. All we do in such a world is to exist. This is just another broken utopia.
-1XiXiDu
I agree. All you need is a robot that does not mistake "earn a college degree" for "kill all other humans and print an official paper confirming that you earned a college degree". All trends I am aware of indicate that software products will become better at knowing what you meant. But in order for them to constitute an existential risk they would have to become catastrophically worse at understanding what you meant while at the same time becoming vastly more powerful at doing what you did not mean. But this doesn't sound at all likely to me. What I imagine is that at some point we'll have a robot that can enter a classroom, sit down, and process what it hears and sees in such a way that it will be able to correctly fill out a multiple choice test at the end of the lesson. Maybe the robot will literally step on someones toes. This will then have to be fixed. What I don't think is that the first robot entering a classroom, in order to master a test, will take over the world after hacking school's WLAN and solving molecular nanotechnology. That's just ABSURD.
0Richard_Kennaway
Um, I think you meant "disagree".
0TheAncientGeek
There's the famous example of the .AI trained to spot tanks that actually leant to spot sunny days. That seems to underlie a lot of MIRI thinking, although at the same time the point is disguised by emphasesing explicit coding over training.
-3Furcas
It just blows my mind that after the countless hours you've spent reading and writing about the Friendly AI problem, not to mention the countless hours people have spent patiently explaining (and re- re- re- re-explaining) it to you, that you still don't understand what the FAI problem is. It's unbelievable.
-3XiXiDu
This line of reasoning still seems flawed to me. It's just like saying that you can build an airplane that can fly and land, autonomously, except that your plane is going to forcefully crash into a nuclear power plant. The gist of the matter is that there are a vast number of ways that you can fail at predicting your programs behavior. Most of these failure modes are detrimental to the overall optimization power of the program. This is because being able to predict the behavior of your AI, to the extent necessary for it to outsmart humans, is analogous to predicting that your airplane will fly without crashing. Eliminating humans, in order to optimize the economy, is about as likely as your autonomous airplane crashing into a nuclear power plant, in order to land safely.
6nshepperd
I don't know why you think you can predict the likely outcome of an artificial general intelligence by making surface analogies to things that aren't even optimization processes. People have been using analogies to "predict" nonsense for centuries. In this case there are a variety of reasons that a programmer might succeed at preventing a UAV from crashing into a nuclear power plant, yet fail at preventing AGI from eliminating all humans. Mainly revolving around the fact that most programmers wouldn't even consider the "eliminate all humans" option as a serious possibility until it had already occurred, while the problem of physical obstructions is explicitly a part of the UAV problem definition. That itself has to do with the fact that an AGI can represent internally features of the world that weren't even considered by the designers (due to general intelligence). As an aside, serious misconfigurations or unintended results of computer programs happen all the time today, but you don't generally hear or care about them because they don't end the world.
-6XiXiDu
2PhilosophyTutor
(EDIT: See below.) I'm afraid that I am now confused. I'm not clear on what you mean by "these traits", so I don't know what you think I am being confident about. You seem to think I'm arguing that AIs will converge on a safe design and I don't remember saying anything remotely resembling that. EDIT: I think I figured it out on the second or third attempt. I'm not 100% committed to the proposition that if we make an AI and know how we did so that we can definitely make sure it's fun and friendly, as opposed to fundamentally uncontrollable and unknowable. However it seems virtually certain to me that we will figure out a significant amount about designing AIs to do what we want in the process of developing them. People who subscribe to various "FOOM" theories about AI coming out of nowhere will probably disagree with this as is their right, but I don't find any of those theories plausible. I also I hope I didn't give the impression that I thought it was meaningfully possible to create a God-like AI without understanding how to make AI. It's conceivable in that such a creation story is not a logical contradiction like a square circle or a colourless green dream sleeping furiously, but that is all. I think it is actually staggeringly unlikely that we will make an AI without either knowing how to make an AI, or knowing how to upload people who can then make an AI and tell use how they did it.
0Stuart_Armstrong
Significant is not the same as sufficient. How low do you think the probability of negative AI outcomes is, and what are your reasons for being confident in that estimate?
2[anonymous]
For the same reason a jet engine doesn't have comfy chairs: with all machines, you develop the core physical and mathematical principles first, and then add human comforts. The core mathematical and physical principles behind AI are believed, not without reason, to be efficient cross-domain optimization. There is no reason for an arbitrarily-developed Really Powerful Optimization Process to have anything in its utility function dealing with human morality; in order for it to be so, you need your AI developers to be deliberately aiming at Friendly AI, and they need to actually know something about how to do it. And then, if they don't know enough, you need to get very, very, very lucky.
0TheAncientGeek
That's what happens when Friendly is used to mean both Fun and Safe. Early jets didn't have comfy chairs, but they did have electors seats. Safety was a concern. If an .AI researchers feels their .AI might kill them, they will have every motivation to build in safety features. That has nothing g to do with making an .AI Your Plastic Pal Who's Fun To Be With.
2[anonymous]
It's an open question whether we could construct a utility function that is, in the ultimate analysis, Safe without being Fun. Personally, I'm almost hoping the answer is no. I'd love to see the faces of all the world's Very Serious People as we ever-so-seriously explain that if they don't want to be killed to the last human being by a horrible superintelligent monster, they're going to need to accept Fun as their lord and savior ;-).
-5TheAncientGeek
2TheAncientGeek
MIRIs arguments aren't about deliberate weaponisation, they are about the inadvertent creation of dangerous .AI by competent and well intentioned people. The weaponisation of .AI has almost happenedalready the form of stuxnet and it is significant that there were a lot safeguards built into it. .AI researchers seemed be aware enough.
0TheAncientGeek
I have no idea why the querrying process would have to be hard. Is David Frost some super genius?
0Stuart_Armstrong
"Defining what querying process is acceptable" is the hard part.
0TheAncientGeek
The justification of which is?
0Stuart_Armstrong
That no one has come close to providing a successful approach on how to do this, and that each proposal fails in very similar ways. There is no ontologically fundamental difference between an acceptable and an unacceptable query, and drawing a practical boundary has so far proved to be impossible. If you have a solution to that, then I advise you analyse it carefully, and then put it as a top level post. Since it would half-solve the whole FAI problem, it would garner great interest.
0TheAncientGeek
Nobody knows how to build AGI either. You've adopted Robby's favourite fallacy: arguing of absolute difficulty as though it were relative difficulty. The hard part has got be harder than the rest of AGI. But why shout a SAI that can pass the .TT with flying colours be unable to do something a human can do?
1Stuart_Armstrong
Orthogonality thesis: building an AGI is a completely different challenge from building an AGI with an acceptable motivation system. It is not a question of ability, but of preferences. Why should an AI that can pass the TT want something that a human wants?
0TheAncientGeek
The thing in question isn't collecting barbie dolls, it's understanding correctly. An .AI that sits at the end of a series of self improvements has got to be pretty good at that. You can say it will have only instrumental rationality, and will start getting things wrong in order to pursue its ultimate goal of word domination, and I can say that if instrumental rationality is dangerous, don't bulld it that way.
0Stuart_Armstrong
No, it's preferences the problem, not understanding. Why would an AI sitting at the end of a series of self improvements choose to interpret ambiguous coding in the way we prefer? How do you propose to build an AI without instrumental rationality or preventing that from developing? And how do you propose to convince AI designers to go down that route?
0TheAncientGeek
If it has epistemic rationality as a goal, it will default to getting things right rather than wrong. Not only nstrumental rationality = epistemic rationality.
0Stuart_Armstrong
If it has epistemic rationality as a goal, it will default to acquiring true beliefs about the world. Explain how this will make it "nice".
0TheAncientGeek
See above. The question was originally about interpreting directives. You have switched to inferring morality apriori.
0hairyfigment
While I don't know how much I believe the OP, remember that "liberty" is a hotly contested term. And that's without a superintelligence trying to create confusing cases. Are you really arguing that "a relatively small part of the processing power of one human brain" suffices to answer all questions that might arise in the future, well enough to rule out any superficially attractive dystopia?
4PhilosophyTutor
I really am. I think a human brain could rule out superficially attractive dystopias and also do many, many other things as well. If you think you personally could distinguish between a utopia and a superficially attractive dystopia given enough relevant information (and logically you must think so, because you are using them as different terms) then it must be the case that a subset of your brain can perform that task, because it doesn't take the full capabilities of your brain to carry out that operation. I think this subtopic is unproductive however, for reasons already stated. I don't think there is any possible world where we cannot achieve a tiny, partial solution to the strong AI problem (codifying "liberty", and similar terms) but we can achieve a full-blown, transcendentally superhuman AI. The first problem is trivial compared to the second. It's not a trivial problem, by any means, it's a very hard problem that I don't see being overcome in the next few decades, but it's trivial compared to the problem of strong AI which is in turn trivial compared to the problem of vastly superhuman AI. I think Stuart_Armstrong is swallowing a whale and then straining at a gnat.
-1hairyfigment
No, this seems trivially false. No subset of my brain can reliably tell when an arbitrary Turing machine halts and when it doesn't, no matter how meaningful I consider the distinction to be. I don't know why you would say this.
3PhilosophyTutor
I'll try to lay out my reasoning in clear steps, and perhaps you will be able to tell me where we differ exactly. 1. Hairyfigment is capable of reading Orwell's 1984, and Banks' Culture novels, and identifying that the people in the hypothetical 1984 world have less liberty than the people in the hypothetical Culture world. 2. This task does not require the full capabilities of hairyfigment's brain, in fact it requires substantially less. 3. A program that does A+B has to be more complicated than a program that does A alone, where A and B are two different, significant sets of problems to solve. (EDIT: If these programs are efficiently written) 4. Given 1-3, a program that can emulate hairyfigment's liberty-distinguishing faculty can be much, much less complicated than a program that can do that plus everything else hairyfigment's brain can do. 5. If we can simulate a complete human brain that is the same as having solved the strong AI problem. 6. A program that can do everything hairyfigment's brain can do is a program that simulates a complete human brain. 7. Given 4-6 it is much less complicated to emulate hairyfigment's liberty-distinguishing faculty than to solve the strong AI problem. 8. Given 7, it is unreasonable to postulate a world where we have solved the strong AI problem, in spades, so much so we have a vastly superhuman AI, but we still haven't solved the hairyfigment's liberty-distinguishing faculty problem.
0hairyfigment
..It's the hidden step where you move from examining two fictions, worlds created to be transparent to human examination, to assuming I have some general "liberty-distinguishing faculty".
1PhilosophyTutor
We have identified the point on which we differ, which is excellent progress. I used fictional worlds as examples, but would it solve the problem if I used North Korea and New Zealand as examples instead, or the world in 1814 and the world in 2014? Those worlds or nations were not created to be transparent to human examination but I believe you do have the faculty to distinguish between them. I don't see how this is harder than getting an AI to handle any other context-dependant, natural language descriptor, like "cold" or "heavy". "Cold" does not have a single, unitary definition in physics but it is not that hard a problem to figure out when you should say "that drink is cold" or "that pool is cold" or "that liquid hydrogen is cold". Children manage it and they are not vastly superhuman artificial intelligences.
-1TheAncientGeek
H.airyfigment, do you canmean detecting liberty in reality is different to, or harder than, detecting liberty in fiction?
0CCC
Incorrect. I can write a horrendously complicated program to solve 1+1; and a far simpler program to add any two integers. Admittedly, neither of those are particularly significant problems; nonetheless, unnecessary complexity can be added to any program intended to do A alone. It would be true to say that the shortest possible program capable of solving A+B must be more complex than the shortest possible program to solve A alone, though, so this minor quibble does not affect your conclusion. Granted. Why? Just because the problem is less complicated, does not mean it will be solved first. A more complicated problem can be solved before a less complicated problem, especially if there is more known about it.
0PhilosophyTutor
To clarify, it seems to me that modelling hairyfigment's ability to decide whether people have liberty is not only simpler than modelling hairyfigment's whole brain, but that it is also a subset of that problem. It does seem to me that you have to solve all subsets of Problem B before you can be said to have solved Problem B, hence you have to have solved the liberty-assessing problem if you have solved the strong AI problem, hence it makes no sense to postulate a world where you have a strong AI but can't explain liberty to it.
1CCC
Hmmm. That's presumably true of hairyfigment's brain; however, simulting a copy of any human brain would also be a solution to the strong AI problem. Some human brains are flawed in important ways (consider, for example, psychopaths) - given this, it is within the realm of possibility that there exists some human who has no conception of what 'liberty' means. Simulating his brain is also a solution of the Strong AI problem, but does not require solving the liberty-assessing problem.
1EHeller
If you can simulate the whole brain, you can just simulate asking the brain the question "does this offend against liberty."
1Stuart_Armstrong
Under what circumstances? There are situations - torture, seduction, a particular way of asking the question - that can make any brain give any answer. Defining "non-coercive yet informative questioning" about a piece of software (a simulated brain) is... hard. AI hard, as some people phrase it.
2TheAncientGeek
Why would that .be more of a problem for an AI than a human?
0Stuart_Armstrong
? The point is that having a simulated brain and saying "do what this brain approves of" does not make the AI safe, as defining the circumstance in which the approval is acceptable is a hard problem. This is a problem for us controlling an AI, not a problem for the AI.
0TheAncientGeek
I still don't get it. We assume acceptability by default. We don't constantly stop and ask "Was that extracted under torture".
0Stuart_Armstrong
I do not understand your question. It was suggested that an AI run a simulated brain, and ask the brain for approval for doing its action. My point was that "ask the brain for approval" is a complicated thing to define, and puts no real limits on what the AI can do unless we define it properly.
0TheAncientGeek
Ok. You are assuming the superintelligent .AI will pose the question in a dumb way?
0Stuart_Armstrong
No, I am assuming the superintelligent AI will pose the question in the way it will get the answer it prefers to get.
0TheAncientGeek
Oh, you're assuming it's malicious. In order to prove...?
2Stuart_Armstrong
No, not assuming it's malicious. I'm assuming that it has some sort of programming along the lines of "optimise X, subject to the constraint that uploaded brain B must approve your decisions." Then it will use the most twisted definition of "approve" that it can find, in order to best optimise X.
-2TheAncientGeek
The programme it with: Prime directive - interpret all directives according to your makers intentions. Secondary directive - do nothing that goes against the uploaded brain Tertiary objective - optimise X.
0Stuart_Armstrong
And how do you propose to code the prime directive? (with that, you have no need for the other ones; the uploaded brain is completely pointless)
0TheAncientGeek
The prime directive is the tertiary directive for a specific X
0Stuart_Armstrong
That's not a coding approach for the prime directive.
0TheAncientGeek
You have already assumed you can build an .AI that optimises X. I am not assuming anything different. In fact any .AI that self improves is going to have to have some sort of goal of getting things right, whether instrumental or terminal. Terminal is much safer, to the extent that it might even solve the whole friendliness problem.
1Stuart_Armstrong
No, you are assuming that we can build an AI that optimises a specific thing, "interpret all directives according to your makers intentions". I'm assuming that we can build an AI that can optimise something, which is very different.
2XiXiDu
An AI that can self-improve considerably does already interpret a vast amount of directives according to its makers intentions, since self-improvement is an intentional feature. Being able to predict a programs behavior is a prerequisite if you want the program to work well. Since unpredictable behavior tends to be chaotic and detrimental to the overall performance. In other words, if you got an AI that does not work according to its makers intentions, then you got an AI that does not work, or which is not very powerful.
0Stuart_Armstrong
Goedel machines already specify self-improvement in formal mathematical form. If you can specify human morality in a similar formal manner, I'll be a lot more relaxed. Also, I don't assume self improvement. Some model of powerful intelligences don't require it.
0TheAncientGeek
So your saying the orthogonality thesis is false?
0Stuart_Armstrong
??? * Orthogonality thesis: an AI that optimises X can be built, in theory, for almost any X * My assumption in this argument: an AI that optimises X can be built, for some X. * What we need: a way of building, in practice, the specific X we want. In fact, let's be generous: you have an AI that can optimise any X you give it. All you need to do now is specify that X to get the result we want. And no, "interpret all directives according to your makers intentions" is not a specification.
0TheAncientGeek
But it's an instruction humans are capable of following within the limits of their ability. If I was building a non .AI system to do X, then I would have to specify X. But AIs are learning systems. If you are going to admit that there is difference between theoretical possibility and practical likelihood in the OT, then ,most of the UFAI argument goes out of the window, since the Lovecraftian Horrors that so densly populate mindspace are only theoretical possibilities.
1Stuart_Armstrong
Because they desire to do so. If for some reason the human has no desire to follow those instructions, then they will "follow" them formally but twist them beyond recognition. Same goes for AI, except that they will not default to desiring to follow them, as many humans would.
0TheAncientGeek
What an .AI does depends how it is built. You keep arguing that one particular architectural choice, with an arbitrary top level goal and only instrumental rationality is dangerous. But that choice is not necessary or inevitable.
1Stuart_Armstrong
(Almost) any top level goal that does not specify human safety. Self modifying AIs will tend to instrumental rationality according to Omohundro's arguments. Good. How do you propose to avoid that happening? You seem extraordinarily confident that these as-yet-undesigned machines, developed and calibrated in a training environment only, by programmers who don't take AI risk seriously, and put potentially into positions of extreme power where I wouldn't trust any actual human - will end up capturing almost all of human morality.
2Kawoomba
That confidence, I'd surmise, often goes hand in hand with an implicit or explicit belief in objective morality.
-2TheAncientGeek
If you don't think people should believe in it, argue against it, and not just a strawmman version.
2Kawoomba
I've argued against both against convergent goal fidelity regarding the intended (versus the actually programmed in) goals and against objective morality at length, and multiple times. I can dig up a few comments, if you'd like. I don't know what strawman version you're referring to, though: the accuracy/inaccuracy of my assertion doesn't affect the veracity of your claim.
0TheAncientGeek
The usual strawmen are The Tablet and Written into the Laws of the Universe.
-2TheAncientGeek
There is no reason to suppose they will not tend to epistemic rationality, which includes instrumental rationality. You have no evidence that .AI researchers aren't taking .AI risk seriously enough, given what they are in fact doing. They may not be taking your arguments seriously, and that may well be because you arguments are not relevant to their research. A number of them have said as much on this site. Even aside from the relevance issue, the MIRI argument constantly assumes that superintelligent IS will have inexplicable deficits. Superintelligent but dumb doesn't make logical sense.
3Stuart_Armstrong
And you've redefined "anything but perfectly morally in tune with humanity" as "dumb". I'm waiting for an argument as to why that is so.
0TheAncientGeek
There's an argument that an SAI will figure out the correct morality, and there's an argument that it wont misinterpret directives. They are different arguments, and the second is much stronger.
1Stuart_Armstrong
I now see your point. I still don't see how you plan to code a "interpret these things properly" piece of the AI. I think working through a specific design would be useful. I also think you should work your argument into a less wrong post (and send me a message when you've done that, in case I miss it) as 12 or so levels deep into a comment thread is not a place most people will ever see. Not really. Given the first, we can instruct "only do things that [some human or human group with nice values] would approve of" and we've got an acceptable morality.
0TheAncientGeek
By "interpret these things correctly", do you mean linguistic competence, or a goal? The linguistic competence is aready assumed in any .AI that can talk it's way out of a box (ie not AIXI like), without provision of a design by MIRI. An AIXI can't even conceptualise that it's in a box, so it doesn't matter if it gets its goals wrong, It can be rendered safe by boxing. Which combination of assumptions is the problem?
0Stuart_Armstrong
I'm not so sure about that... AIXI can learn certain ways of behaving as if it were part of the universe, even with the Cartesian dualism in its code: http://lesswrong.com/lw/8rl/would_aixi_protect_itself/ A goal. If the AI becomes superintelligent, then it will develop linguistic competence as needed. But I see no way of coding it so that that competence is reflected in its motivation (and it's not from lack of searching for ways of doing that).
0TheAncientGeek
So is it safe to run AIXI approximations in boxes today? By code it, do you mean "code, train, or evolve it"? Note that we dont know much about coding higher level goals in general. Note that "get things right except where X is concerned" is more complex than "get things right". Humans do the former because of bias. The less anthropic nature of an .AI might be to our advantage.
0[anonymous]
IMHO, yes. The computational complexity of AIXItl is such that it can't be used for anything significant on modern hardware.
0Neph
def checkMorals(): >[simulate philosophy student's brain] >if [simulated brain's state is offended]: >>return False >else: >>return True if checkMorals(): >[keep doing AI stuff] there. that's how we tell an AI capable of being an AI and capable of simulating a brain to not to take actions which the simulated brain thinks offend against liberty, as implemented in python.
0Stuart_Armstrong
oh, it's so clear and obvious now, how could I have missed that?
-2[anonymous]
And therein lies the rub. Current research-grade AGI formalisms don't actually allow us to specifically program the agent for anything, not even paperclips.
0PhilosophyTutor
If I was unclear, I was intending that remark to apply to the original hypothetical scenario where we do have a strong AI and are trying to use it to find a critical path to a highly optimal world. In the real world we obviously have no such capability. I will edit my earlier remark for clarity.
0Strange7
Unless you start by removing the air, in some way that doesn't count against the car's efficiency.
0drnickbone
This also creates some interesting problems... Suppose a very powerful AI is given human liberty as a goal (or discovers that this is a goal using coherent extrapolated volition). Then it could quickly notice that its own existence is a serious threat to that goal, and promptly destroy itself!
1Stuart_Armstrong
yes, but what about other AIs that might be created, maybe without liberty as a top goal - it would need to act to prevent them from being built! It's unlikely that "destroy itself" is the best option it can find...
0drnickbone
Except that acting to prevent other AIs from being built would also encroach on human liberty, and probably in a very major way if it was to be effective! The AI might conclude from this that liberty is a lost cause in the long run, but it is still better to have a few extra years of liberty (until the next AI gets built), rather than ending it right now (through its own powerful actions). Other provocative questions: how much is liberty really a goal in human values (when taking the CEV for humanity as a whole, not just liberal intellectuals)? How much is it a terminal goal, rather than an instrumental goal? Concretely, would humans actually care about being ruled over by a tyrant, as long as it was a good tyrant? (Many people are attracted to the idea of an all-powerful deity for instance, and many societies have had monarchs who were worshipped as gods.) Aren't mechanisms like democracy, separation of powers etc mostly defence mechanisms against a bad tyrant? Why shouldn't a powerful "good" AI just dispense with them?
0Stuart_Armstrong
A certain impression of freedom is valued by humans, but we don't seem to want total freedom as a terminal goal.
3[anonymous]
Well of course we don't. Total freedom is an incoherent goal: the only way to ensure total future freedom of action is to make sure nothing ever happens, thus maximizing the number of available futures without ever actually choosing one. As far as I've been able to reason out, the more realistic human conception of freedom is: "I want to avoid having other agenty things optimize me (for their preferences (unilaterally))." The last part is there because there are mixed opinions on whether you've given up your ethical freedom if an agenty thing optimizes you for your preferences (as might happen in ideal situations, such as dealing with an FAI handing out transhuman candy), or whether you've given up your ethical freedom if you bind yourself to implement someone else's preferences mixed-in with your own (for instance, by getting married).
4Lumifer
That doesn't make sense -- why would the status quo, whatever it is, always maximize the number of available futures? Choosing a future does not restrict you, it does close some avenues but also opens other ones. "Total freedom" is a silly concept, of course, but it's just as silly as "Total ".
0Stuart_Armstrong
Total happiness seems to make more plausible sense than total freedom.
0Lumifer
Not sure how you determine degrees of plausibility :-/ The expression "total happiness" (other than in contexts of the "it's like, dude, I was so totally chill and happy" kind) makes no more sense to me than "total freedom".
0Stuart_Armstrong
Assume B choose without coercion, but assume A always knows what B will choose and can set up various facts in the world to determine B's choice. Is B free?
2PhilosophyTutor
I think there is insufficient information to answer the question as asked. If I offer you the choice of a box with $5 in it, or a box with $500 000 in it, and I know that you are close enough to a rational utility-maximiser that you will take the $500 000, then I know what you will choose and I have set up various facts in the world to determine your choice. Yet it does not seem on the face of it as if you are not free. On the other hand if you are trying to decide between being a plumber or a blogger and I use superhuman AI powers to subtly intervene in your environment to push you into one or the other without your knowledge then I have set up various facts in the world to determine your choice and it does seem like I am impinging on your freedom. So the answer seems to depend at least on the degree of transparency between A and B in their transactions. Many other factors are almost certainly relevant, but that issue (probably among many) needs to be made clear before the question has a simple answer.
0Stuart_Armstrong
Can you cash out the difference between those two cases in sufficient detail that we can use it to safely defined what liberty means?
-2PhilosophyTutor
I said earlier in this thread that we can't do this and that it is a hard problem, but also that I think it's a sub-problem of strong AI and we won't have strong AI until long after we've solved this problem. I know that Word of Eliezer is that disciples won't find it productive to read philosophy, but what you are talking about here has been discussed by analytic philosophers and computer scientists as "the frame problem" since the eighties and it might be worth a read for you. Fodor argued that there are a class of "informationally unencapsulated" problems where you cannot specify in advance what information is and is not relevant to solving the problem, hence really solving them as opposed to coming up with a semi-reliable heuristic is an incredibly difficult problem for AI. Defining liberty or identifying it in the wild seems like it's an informationally unencapsulated problem in that sense and hence a very hard one, but one which AI has to solve before it can tackle the problems humans tackle. If I recall correctly Fodor argued in Modules, Frames, Fridgeons, Sleeping Dogs, and the Music of the Spheres that this problem was in fact the heart of the AI problem, but that proposition was loudly raspberried in the literature by computer scientists. You can make up your own mind about that one. Here's a link to the Stanford Dictionary of Philosophy page on the subject.
0Stuart_Armstrong
It depends on how general or narrow you make the problem. Compare: is classical decision theory the heart of the AI problem? If you interpret this broadly, then yes; but the link from, say, car navigation to classical decision theory is tenuous when you're working on the first problem. The same thing for the frame problem.
0Stuart_Armstrong
You mean the frame problem that I talked about here? http://lesswrong.com/lw/gyt/thoughts_on_the_frame_problem_and_moral_symbol/ The issue can be talked about in terms of the frame problem, but I'm not sure it's useful. In the classical frame problem, we have a much clearer idea of what we want, the problem is specifying enough so that the AI does too (ie so that the token "loaded" corresponds to the gun being loaded). This is quite closely related to symbol grounding, in a way. When dealing with moral problems, we have the problem that we haven't properly defined the terms to ourselves. Across the span of possible futures, the term "loaded gun" is likely much sharply defined than "living human being". And if it isn't - well, then we have even more problems if all our terms start becoming slippery, even the ones with no moral connotations. But in any case, saying the problem is akin to the frame problem... still doesn't solve it, alas!
0Shmi
Note that the relevance issue has been successfully solved in any number of complex practical applications, such as the self-driving vehicles, which are able to filter out gobs of irrelevant data, or the LHC code, which filters out even more. I suspect that the Framing Problem is not some general problem that needs to be resolved for AI to work, but just one of many technical issues, just as the "computer scientists" suggest. On the other hand, it is likely to be a real problem for FAI design, where relying to heuristics providing, say, six-sigma certainty just isn't good enough. I think that the framing problem is distinct from the problem of defining and calculating mostly because attempting to define liberty objectively leads us to the discussion of free will, the latter being an illusion due to the human failure to introspect deep enough.
0PhilosophyTutor
I tend to think that you don't need to adopt any particular position on free will to observe that people in North Korea lack freedom from government intervention in their lives, access to communication and information, a genuine plurality of viable life choices and other objectively identifiable things humans value. We could agree for the sake of argument that "free will is an illusion" (for some definitions of free will and illusion) yet still think that people in New Zealand have more liberty than people in North Korea. I think that you are basically right that the Framing Problem is like the problem of building a longer bridge, or a faster car, in that you are never going to solve the entire class of problem at a stroke so that you can make infinitely long bridges or infinitely fast cars but that you can make meaningful incremental progress over time. I've said from the start that capturing the human ability to make philosophical judgments about liberty is a hard problem but I don't think it is an impossible one - just a lot easier than creating a program that does that and solves all the other problems of strong AI at once. In the same way that it turns out to be much easier to make a self-driving car than a strong AI, I think we'll have useful natural-language parsing of terms like "liberty" before we have strong AI.
2Shmi
Well, yes, it is hard to argue about NK vs West. But let's try to control for the "non-liberty" confounders, such as income, wealth, social status, etc. Say, we take some upper middle-class person from Iran, Russia or China. It is quite likely that, when comparing their life with that of a Westerner of similar means, they would not immediately state that the Western person has more "objectively identifiable things humans value". Obviously the sets of these valuable things are different, and the priorities different people assign to them would be different, but I am not sure that there is a universal measure everyone would agree upon as "more liberty".
2PhilosophyTutor
A universal measure for anything is a big demand. Mostly we get by with some sort of somewhat-fuzzy "reasonable person" standard, which obviously we can't fully explicate in neurological terms either yet, but which is much more achievable. Liberty isn't a one-dimensional quality either, since for example you might have a country with little real freedom of the press but lots of freedom to own guns, or vice versa. What you would have to do to develop a measure with significant intersubjective validity is to ask a whole bunch of relevantly educated people what things they consider important freedoms and what incentives they would need to be offered to give them up, to figure out how they weight the various freedoms. This is quite do-able, and in fact we do very similar things when we do QALY analysis of medical interventions to find out how much people value a year of life without sight compared to a year of life with sight (or whatever). Fundamentally it's not different to figuring out people's utility functions, except we are restricting the domain of questioning to liberty issues.
0[anonymous]
So, just checking before I answer: you're claiming that no direct, gun-to-the-head coercion is employed, but Omega can always predict your actions and responses, and sets things up to ensure you will choose a specific thing. Are you free, or are you in some sense "serving" Omega? I answer: The latter, very, very, very definitely. If we take it out of abstract language, real people manipulate each-other all the time, and we always condemn it as a violation of the ethical principle of free choice. Yes, sometimes there are principles higher than free choice, as with a parent who might say, "Do your homework or you get no dessert" (treat that sentence as a metasyntactic variable for whatever you think is appropriate parenting), but we still prefer, all else equal, that our choices and those of others be less manipulated rather than more. Just because fraud and direct coercion are the usual standards for proving a violation of free choice in a court of law, for instance in order to invalidate a legal contract, does not mean that these are the all-and-all of the underlying ethics of free choice.
0Stuart_Armstrong
Then if Omega is superintelligent, it has a problem: every single decision it makes increases or decreases the probability of someone answering something or other, possibly by a large amount. It seems Omega cannot avoid being coercive, just because it's so knowledgeable.
2[anonymous]
We don't quite know that, and there's also the matter of whether Omega is deliberately optimizing those people or they're just reacting to Omega's optimizing the inanimate world (which I would judge to be acceptable and, yes, unavoidable).
0Stuart_Armstrong
It is an interesting question, though, and illustrates the challenges with "liberty" as a concept in these circumstances.
0[anonymous]
Well yes. It's also why many people have argued in favor of Really Powerful Optimizers just... doing absolutely nothing.
0Stuart_Armstrong
That I don't see.
0PhilosophyTutor
I think Asimov did this first with his Multivac stories, although rather than promptly destroy itself Multivac executed a long-term plan to phase itself out.

Upvoted for use of images. Though sort of tabooed on LW, when used well, they work.

-5Guocamole

I don't understand why you imply that an evil Oracle will not be able to present only or mostly the evil possible worlds disguised as good. My guess would be that satisficing gets you into just as much trouble as optimizing.

8Stuart_Armstrong
The evil oracle is mainly to show the existence of siren worlds; and if we use an evil oracle for satisficing, we're in just as much trouble as if we were maximising (probably more trouble, in fact). The marketing and siren worlds are a problem even without any evil oracle, however. For instance a neutral, maximising, oracle would serve us up a marketing world.
3itaibn0
Why do you think that? You seem to think that along the higher ends, appearing to be good actually anti-correlates with being good. I think it is plausible that the outcome optimized to appear good to me actually is good.There may be many outcomes that appear very good but are actually very bad, but I don't see why they would be favoured in being the best-appearing. I admit though that the best-appearing outcome is unlikely to be the optimal outcome, assuming 'optimal' here means anything.

We are both superintelligences. You have a bunch of independently happy people that you do not aggressively compel. I have a group of zombies - human-like puppets that I can make do anything, appear to feel anything (though this is done sufficiently well that outside human observers can't tell I'm actually in control). An outside human observer wants to check that our worlds rank high on scale X - a scale we both know about.

Which of us do you think is going to be better able to maximise our X score?

2itaibn0
I'm not sure what the distinction you're making is. Even a free-minded person can be convinced through reason to act in certain ways, sometimes highly specific ways. Since you assume the superintelligence will manipulate people so subtly that I won't be able to tell they're being manipulated, it is unlikely that they are directly coerced. This is important, since while I don't like direct coercion, the less direct the method of persuasion the less certain I am that this method of persuasion is bad. These "zombies", who are not being threatened, nor lied to, nor are their neurochemistry directly altered, nor is anything else done that seems to me like coercion, but nonetheless are being coerced. This seems to me as sensical as the other type of zombies. But suppose I'm missing something, and there is a genuine non-arbitrary distinction between being convinced and being coerced. Then with my current knowledge I think I want people not to be coerced. But now an output pump can take advantage of this. Consider the following scenario: Humans are convinced the their existence depends on their behavior being superficially appealing, perhaps by being full of flashing lights. If my decisions in front of an Oracle will influence the future of humanity, this belief is in fact correct; they're not being deceived. Convinced of this, they structure their society to be as superficially appealing as possible. In addition, in the layers too deep for me to notice, they do whatever they want. This outcome seems superficially appealing to me in many ways, and in addition, the Oracle informs me that in some non-arbitrary sense these people aren't being coerced. Why wouldn't this be the outcome I pick? Again, I don't think this outcome would be the best one, since I think people are better off not being forced into this trade-off. One point you can challenge is whether the Oracle will inform me about this non-arbitrary criterion. Since it already can locate people and reveal their supe
2[anonymous]
Would you mind explaining what you consider a desirable future in which people just don't matter?
0itaibn0
Here's the sort of thing I'm imagining: In the beginning there are humans. Human bodies become increasingly impractical in the future environment and are abandoned. Digital facsimiles will be seen as pointless and will also be abandoned. Every component of the human mind will be replaced with algorithms that achieve the same purpose better. As technology allows the remaining entities to communicate with each other better and better, the distinction between self and other will blur, and since no-one will see to any value in reestablishing it artificially, it will be lost. Individuality too is lost, and nothing that can be called human remains. However, every step happens voluntarily because what comes after is seen as better than what is before, and I don't see why I should consider the final outcome bad. If someone has different values they would perhaps be able to stop at some stage in the middle, I just imagine such people would be a minority.
0[anonymous]
So you're using a "volunteerism ethics" in which whatever agents choose voluntarily, for some definition of voluntary, is acceptable, even when the agents may have their values changed in the process and the end result is not considered desirable by the original agents? You only care about the particular voluntariness of the particular choices? Huh. I suppose it works, but I wouldn't take over the universe with it.
7Richard_Kennaway
When it happens fast, we call it wireheading. When it happens slowly, we call it the march of progress.
0[anonymous]
Eehhhhhh.... Since I started reading Railton's "Moral Realism" I've found myself disagreeing with the view that our consciously held beliefs about our values really are our terminal values. Railton's reduction from values to facts allows for a distinction between the actual March of Progress and non-forcible wireheading.
0Stuart_Armstrong
There need not be a distinction between them. If you prefer, you could contrast an AI willing to "convince" its humans to behave in any way required, with one that is unwilling to sacrifice their happiness/meaningfulness/utility to do so. The second is still at a disadvantage.
0itaibn0
Remember that my original point is that I believe appearing to be good correlates with goodness, even in extreme circumstances. Therefore, I expect restructuring humans to make the world appear tempting will be to the benefit of their happiness/meaningfulness/utility. Now, I'm willing to consider that are aspects of goodness which are usually not apparent to an inspecting human (although this moves to the borderline of where I think 'goodness' is well-defined). However, I don't think these aspects are more likely to be satisfied in a satisficing search than in an optimizing search.
0Gunnar_Zarncke
This actually describes quite well the society we already live in - if you take 'they' as 'evolution' (and maybe some elites). For most people our society appears appealing. Most don't see what happens enough layers down (or up). And most don't feel coerced (at least of you still have a strong social system).
0[anonymous]
Hold on. I'm not sure the Kolmogorov complexity of a superintelligent siren with a bunch of zombies that are indistinguishable from real people up to extensive human observation is actually lower than the complexity of a genuinely Friendly superintelligence. After all, a Siren World is trying to deliberately seduce you, which means that it both understands your values and cares about you in the first place. Sure, any Really Powerful Learning Process could learn to understand our values. The question is: are there more worlds where a Siren cares about us but doesn't care about our values than there are worlds in which a Friendly agent cares about our values in general and caring about us as people falls out of that? My intuitions actually say the latter is less complex, because the caring-about-us falls out as a special case of something more general, which means the message length is shorter when the agent cares about my values than when it cares about seducing me. Hell, a Siren agent needs to have some concept of seduction built into its utility function, at least if we're assuming the Siren is truly malicious rather than imperfectly Friendly. Oh, and a philosophically sound approach to Friendliness should make imperfectly Friendly futures so unlikely as to be not worth worrying about (a failure to do so is a strong sign you've got Friendliness wrong). All of which, I suppose, reinforces your original reasoning on the "frequency" of Siren worlds, marketing worlds, and Friendly eutopias in the measure space of potential future universes, but makes this hypothetical of "playing as the monster" sound quite unlikely.
0Stuart_Armstrong
Kolmogorov complexity is not relevant; siren worlds are indeed rare. they are only a threat because they score so high on an optimisation scale, not because they are common.

If a narrower search gets worlds that are disproportionately not what we actually want, that might be because we chose the wrong criteria, not that we searched too narrowly per se. A broader search would come up with worlds that are less tightly optimized for the search criteria, but they might be less tightly optimized by simply being bad.

Can you provide any support for the notion that in general, a narrower search comes up with a higher proportion of bad worlds?

9Viliam_Bur
My intuition is that the more you optimize for X, the more you sacrifice everything else, unless it is inevitably implied by X. So anytime there is a trade-off between "seeming more good" and "being more good", the impression-maximizing algorithm will prefer the former. When you start with a general set of words, "seeming good" and "being good" are positively correlated. But when you already get into the subset of worlds that all seem very good, and you continue pushing for better and better impression, the correlation may gradually turn to negative. At this moment you may be unknowingly asking the AI to exploit your errors in judgement, because in given subset that may be the easiest way to improve the impression. Another intuition is the closer you get to the "perfect" world, the more difficult it becomes to find a way to increase the amount of good. But the difficulty of exploiting a human bias that will cause humans to overestimate the value of the world, remains approximately constant. Though this doesn't prove that the world with maximum "seeming good" is some kind of hell. It could still be very good, although not nearly as good as the world with maximum "good". (However, if the world with maximum "seeming good" happens to be some kind of hell, then maximizing for "seeming good" is the way to find it.)
2simon
This intuition seems correct in typical human situations. Everything is highly optimized already with different competing considerations, so optimizing for X does indeed necessarily sacrifice the other things that are also optimized for. So if you relax the constraints for X, you get more of the other things, if you continue optimizing for them. However, it does not follow from this that if you relax your constraint on X, and take a random world meeting at least the lower value of X, your world will be any better in the non-X ways. You need to actually be optimizing for the non-X things to expect to get them.
1Viliam_Bur
Great point!
0simon
Thanks but I don't see the relevance of the reversal test. The reversal test involves changing the value of a parameter but not the amount of optimization. And the reversal test shouldn't apply to a parameter that is already optimized over unless the current optimization is wrong or circumstances on which the optimization depends are changing.
-1Stuart_Armstrong
? A narrower search comes up with less worlds. Acceptable worlds are rare; siren worlds and marketing worlds much rarer still. A narrow search has less chance of including an acceptable world, but also less chance of including one of the other two. There is some size of random search whether the chance of getting an acceptable world is high, but the chance of getting a siren or marketer is low. Non random searches have different equilibriums.
3simon
Some proportion of the worlds meeting the narrow search will also be acceptable. To conclude that that proportion is smaller than the proportion of the broader search that is acceptable requires some assumption that I haven't seen made explicit. ETA: Imagine we divided the space meeting the broad search into little pieces. On average the little pieces would have the same proportion of acceptable worlds as the broad space. You seem to be arguing that the pieces that we would actually come up with if we tried to design a narrow search would actually on average have a lower proportion of acceptable worlds. This claim needs some justification.
1Stuart_Armstrong
It's not an issue of proportion, but of whether there will be a single representative of the class in the worlds we search through. We want a fraction of the space such that there is an acceptable world in it with high probability, and no siren/marketing world, with high probability. Eg if 1/10^100 worlds is acceptable and 1/10^120 worlds is siren/marketing, we might want to search randomly through 10^101 or 10^102 worlds.
5Lumifer
Looks like you're basically arguing for the first-past-the-post search -- just take the first world that you see which passes the criteria.
2Stuart_Armstrong
Yep, that works better than what I was thinking, in fact.
0simon
I don't see how that changes the probability of getting a siren world v. an acceptable world at all (ex ante). If the expected number of siren worlds in the class we look through is less than one, then sometimes there will be none, but sometimes there will be one or more and on average we still get the same expected number and on average the first element we find is a siren world with probability equal to the expected proportion of siren worlds.
0Stuart_Armstrong
The scenario is: we draw X worlds, and pick the top ranking one. If there is a siren world or marketing world, it will come top; otherwise if there is are acceptable worlds, one of them will come top. Depending on how much we value acceptable worlds over non-acceptable and over siren/marketing worlds, and depending on the proportions of each, there is an X that maximises our outcome. (trivial example: if all worlds are acceptable, picking X=1 beats all other alternatives, as higher X simply increases the chance of getting a siren/marketing world).
3simon
Thanks, this clarified your argument to me a lot. However, I still don't see any good reasons provided to believe that, merely because a world is highly optimized on utility function B, it is less likely to be well-optimized on utility function A as compared to a random member of a broader class. That is, let's classify worlds (within the broader, weakly optimized set) as highly optimized or weakly optimized, and as acceptable or unacceptable. You claim that being highly optimized reduces the probability of being acceptable. But your arguments in favour of this proposition seem to be: a) it is possible for a world to be highly optimized and unacceptable (but all the other combinations are also possible) and b) "Genuine eutopias are unlikely to be marketing worlds, because they are optimised for being good rather than seeming good." (In other words, the peak of function B is unlikely to coincide with the peak of function A. But why should the chance that the peak of function B and the peak of function A randomly coincide, given that they are both within the weakly optimized space, be any lower than the chance of a random element of the weakly optimized space coinciding with the peak of function A? And this argument doesn't seem to support a lower chance of the peak of function B being acceptable, either.) Here's my attempt to come up with some kind of argument that might work to support your conclusion: 1) maybe the fact that a world is highly optimized for utility function B means that it is simpler than an average world, and this simplicity results in it likely being relatively unlikely to be a decent world in terms of utility function A. 2) maybe the fact that a world is highly optimized for utility function B means that it is more complex than an average world, in a way that is probably bad for utility function A. Or something. ETA: I had not read http://lesswrong.com/lw/jao/siren_worlds_and_the_perils_of_overoptimised/asdf when I wrote this comment,

I'm wondering whether this framing (choosing between a set of candidate worlds) is the most productive. Does it make sense to use criteria like corrigibility, minimizing impact and prefering reversible actions (or we have no reliable way to evaluate whether these hold)?

Since the evil AI is presenting a design for a world, rather than the world itself, the problem of it being populated with zombies that only appear to be free could be countered by having the design be in an open source format that allows the people examining it (or other AIs) to determine the actual status of the designed inhabitants.

I think the wording here is kind of odd.

An unconstrained search will not find a siren world, or even a very good world. There are simply too many to consider. The problem is that you're likely to design an AI that finds worlds that you'd like. It may or may not actually show you anything, but you program it to give you what it thinks you'd rate the best. You're essentially programming it to design a siren world. It won't intentionally hide anything dark under there, but it will spend way too much effort on things that make the world look good. It might even end up with dark things hidden, just because they were somehow necessary to make it look that good.

0Stuart_Armstrong
That's a marketing world, not a sire world.
0DanielLC
What's the difference?
0Stuart_Armstrong
Siren worlds are optimised to be bad and hide this fact. Marketing worlds are optimised to appear good, and the badness is an indirect consequence of this.

TL;DR: Worlds which meet our specified criteria but fail to meet some unspecified but vital criteria outnumber (vastly?) worlds that meet both our specified and unspecified criteria.

Is that an accurate recap? If so, I think there's two things that need to be proven:

  1. There will with high probability be important unspecified criteria in any given predicate.

  2. The nature of the unspecified criteria is such that it is unfulfilled in a large majority of worlds which fulfill the specified criteria.

(1) is commonly accepted here (rightly so, IMO). But (2) seems... (read more)

2Eugine_Nier
This proposes a way to get an OK result even if we don't quite write down our values correctly.
0lavalamp
Ah, thank you for the explanation. I have complained about the proposed method in another comment. :) http://lesswrong.com/lw/jao/siren_worlds_and_the_perils_of_overoptimised/aso6
0Stuart_Armstrong
That's not exactly my claim. My claim is that things that are the best optimised for fulfilling our specified criteria are unlikely to satisfy our unspecified ones. It's not a question of outnumbering (siren and marketing worlds are rare) but of scoring higher on our specified criteria.

It's not really clear why you would have the searching process be more powerful than the evaluating process, if using such a "search" as part of a hypothetical process in the definition of "good."

Note that in my original proposal (that I believe motivated this post) the only brute force searches were used to find formal descriptions of physics and human brains, as a kind of idealized induction, not to search for "good" worlds.

0Stuart_Armstrong
Because the first supposes a powerful AI, while the second supposes an excellent evaluation process (essentially a value alignment problem solved). Your post motivated this in part, but it's a more general issue with optimisation processes and searches.
0paulfchristiano
Neither the search nor the evaluation presupposes an AI when a hypothetical process is used as the definition of "good."

I'm not totally sure of your argument here; would you be able to clarify why satisficing is superior to a straight maximization given your hypothetical[0]?

Specifically, you argue correctly that human judgement is informed by numerous hidden variables over which we have no awareness, and thus a maximization process executed by us has the potential for error. You also argue that 'eutopian'/'good enough' worlds are likely to be more common than sirens. Given that, how is a judgement with error induced by hidden variables any worse than a judgement made using ... (read more)

1Stuart_Armstrong
Siren and Marketing worlds are rarer than eutopias, but rank higher in our maximisation scale. So picking a world among the "good enough" will likely be a eutopia, but picking the top ranked world will likely be a marketing world.

The IC correspond roughly with what we want to value, but differs from it in subtle ways, enough that optimising for one could be disastrous for the other. If we didn't optimise, this wouldn't be a problem. Suppose we defined an acceptable world as one that we would judge "yeah, that's pretty cool" or even "yeah, that's really great". Then assume we selected randomly among the acceptable worlds. This would probably result in a world of positive value: siren worlds and marketing worlds are rare, because they fulfil very specific criteri

... (read more)
1Eugine_Nier
Why? This agrees with my intuition, ask for too much and you wind up with nothing.

"ask for too much and you wind up with nothing" is a fine fairy tale moral. Does it actually hold in these particular circumstances?

Imagine that there's a landscape of possible words. There is a function (A) on this landscape, we don't know how to define it, but it is how much we truly would prefer a world if only we knew. Somewhere this function has a peak, the most ideal "eutopia". There is another function. This one we do define. It is intended to approximate the first function, but it does not do so perfectly. Our "acceptability criteria" is to require that this second function (B) has a value at least some threshold.

Now as we raise the acceptability criteria (threshold for function B), we might expect there to be two different regimes. In a first regime with low acceptability criteria, Function B is not that bad a proxy for function A, and raising the threshold increases the average true desirability of the worlds that meet it. In a second regime with high acceptability criteria, function B ceases to be effective as a proxy. Here we are asking for "too much". The peak of function B is at a different place than the peak of function A, and... (read more)

0Stuart_Armstrong
True. Which is why I added arguments pointing that a marketing world will likely be bad. Even on your terms, a peak of B will probably involve a diversion of effort/energy that could have contributed to A, away from A. eg if A is apples and B is bananas, the world with the most bananas is likely to contain no apples at all.
5lavalamp
It sounds like, "the better you do maximizing your utility function, the more likely you are to get a bad result," which can't be true with the ordinary meanings of all those words. The only ways I can see for this to be true is if you aren't actually maximizing your utility function, or your true utility function is not the same as the one you're maximizing. But then you're just plain old maximizing the wrong thing.
0Stuart_Armstrong
Er, yes? But we don't exactly have the right thing lying around, unless I've missed some really exciting FAI news...
9lavalamp
Absolutely, granted. I guess I just found this post to be an extremely convoluted way to make the point of "if you maximize the wrong thing, you'll get something that you don't want, and the more effectively you achieve the wrong goal, the more you diverge from the right goal." I don't see that the existence of "marketing worlds" makes maximizing the wrong thing more dangerous than it already was. Additionally, I'm kinda horrified about the class of fixes (of which the proposal is a member) which involve doing the wrong thing less effectively. Not that I have an actual fix in mind. It just sounds like a terrible idea--"we're pretty sure that our specification is incomplete in an important, unknown way. So we're going to satisfice instead of maximize when we take over the world."