This is a road that does not lead to Friendly AI, only to AGI. I doubt this has anything to do with Lenat’s motives—but I’m glad the source code isn’t published and I don’t think you’d be doing a service to the human species by trying to reimplement it.
I doubt this has anything to do with Lenat’s motives—but I’m glad the source code isn’t published
A stunning proof-of-concept for AI, with the source code lost to the mists of time; then immersion in an apparently massive dead-end project. Is anyone else worried that Lenat might secretly have a phase 3 to his plan?
You may stop worrying for the moment. I just tried to wade through the papers RichardKennaway has put up, and it seems reimplementing EURISKO given Lenat’s most detailed descriptions of it will likely be a big creative endeavor. Download them, read them and then tell me just one thing: do you now know (even approximately) what the main loop looks like, or not? Cause I couldn’t make it out in half an hour’s reading.
Depends how cool. I don’t know the space of self-modifying programs very well. Anything cooler than anything that’s been tried before, even marginally cooler, has a noticeable subjective probability of going to shit. I mean, if you kept on making it marginally cooler and cooler, it’d go to “oh, shit” one day after a sequence of “ooh, cools” and I don’t know how long that sequence is.
I mean, if you kept on making it marginally cooler and cooler, it’d go to “oh, shit” one day after a sequence of “ooh, cools” and I don’t know how long that sequence is.
This means we should feel pretty safe, since AI does not appear to be making even incremental progress.
Really, it’s hard for anyone who is well-versed in the “state of the art” of AI to feel any kind of alarm about the possibility of an imminent FOOM. Take a look at this paper. Skim through the intro, note the long and complicated reinforcement learning algorithm, and check out the empirical results section. The test domain involves a monkey in a 5x5 playroom. There are some fun little complications, like a light switch and a bell. Note that these guys are top-class (Andrew Barto basically invented RL), and the paper was published at one of the top-tier machine learning conferences (NIPS), in 2005.
Call me a denier, but I just don’t think the monkey is going to bust out of his playroom and take over the world. At least, not anytime soon.
Taking progress in AI to mean more real world effectiveness:
Intelligence seems to have jumps in real world effectiveness, e.g. the brains of great apes and humans are very similar, the difference in effectiveness is obvious.
So coming to the conclusion that we are fine based on the state of the art not being any more effective (not making progress) would be very dangerous. Perhaps tomorrow, some team of AI researchers will combine the current state of the art solutions in just the right way, resulting in a massive jump in real world effectiveness? maybe enough to have an “oh, shit” moment?
Regardless of the time frame, if the AI community is working towards AGI rather than FAI, we will likely have (eventually) an AI go FOOM or at the very least, and “oh, shit” moment (I’m not sure if they are equivalent).
Does the specific distance even matter? UFAI vs FAI is zero sum, and we have no idea how long FAI will take us. Any progress toward AGI that isn’t “matched” by progress toward FAI is regressive, even if AGI is still 100 years off.
I just site Googled lesswrong.org and overcomingbias.com, and my best guess so far is that AGI = Artificial General Intelligence. (Augmented?) Is that what AGI stands for? Non site-specific google results are definitely not the applicable ones. Apparently it has been used as an abbreviation in this community for quite awhile, but finding out what it stands for takes a few jumps. (Should there be an easily found FAQ? One is not obvious here or on the wiki. Or is a certain degree of obscurity desired?)
Seeing the recent thread necromancy, it looks like this is a much more important question than I realized at the time, since it bears on AI-related existential risk.
The question, to summarize, was, “How exactly do you keep a good AGI prospect from being developed into unFriendly AGI, other than Security by Obscurity?”
My answer is that SbO (Security by Obscurity, not Antimony(II) Oxide, which doesn’t even exist) is not a solution here for the same reason it’s criticized everywhere else (which I assume is that it increases the probability of a rogue outsmarting mainstream researchers). Better to let the good guys be as well informed as the bad guys so they can deploy countermeasures (their own AGI) when the bad guys develop theirs.
But then, I haven’t researched this 24⁄7 for the last several years, so this may be too trite a dismissal.
This is a road that does not lead to Friendly AI, only to AGI.
Lazy question: Did you explain this in an OB post? I have a crappy mental index of the topics you covered, and Google didn’t yield any immediate results, just an explanation of why Eurisko didn’t foom. If not, I’d love to see a post explaining how you’re extrapolating the Eurisko approach into something incompatible with FAI. Or are you just applying the general rule, “AGI that’s not explicitly friendly is explicitly unfriendly”?
Not exactly, Thom. Roughly, for FAI you need precise self-modification. For precise self-modification, you need a precise theory of the intelligence doing the self-modification. To get to FAI you have to walk the road that leads to precise theories of intelligence—something like our present-day probability theory and decision theory, but more powerful and general and addressing issues these present theories don’t.
Eurisko is the road of self-modification done in an imprecise way, ad-hoc, throwing together whatever works until it gets smart enough to FOOM. This is a path that leads to shattered planets, if it were followed far enough. No, I’m not saying that Eurisko in particular is far enough, I’m saying that it’s a first step along that path, not the FAI path.
Perhaps a writeup of what you have discovered, or at least surmise, about walking that road would encourage bright young minds to work on those puzzles instead of reimplementing Eurisko.
It’s not immediately clear that studying and playing with specific toy self-referential systems won’t lead to ideas that might apply to precise members of that class.
I’ve written up some of the concepts of precise self-modification, but need to collect the posts on a Wiki page on “lawfulness of intelligence” or something.
Well, in this sense computing is also a first step on that path, Moore’s law of mad science and all. Eurisko in particular doesn’t seem to deserve more mention than that.
...indeed. It seems that I failed to figure out just what I was arguing against. Let me re-make that point.
As far as first steps along that path go, they have already been taken: we have gone from a world without computers to a world with one, and we can’t reverse that. The logical place to focus our efforts would seem to be the next step which has not been taken, which could very well be reimplementing EURISKO. (Though it could also very well be running a neural net on a supercomputer or some guy making the video game “Operant Conditioning Hero”.)
We have gone from a world without dictators to a world with one, and we can’t reverse that. The logical place to focus our efforts would seem to be the next step which has not been taken, which could very well be resurrecting Hitler.
While grandparent was probably a miscalculation of some sort, I feel that mentioning Hitler is more acceptable if the context is Nazi super science than outrage maximization.
Grandparent was probably a miscalculation of some sort, but I think mention of Hitler is acceptable if the context is Nazi super science rather than outrage maximization.
Resurrecting Hitler would probably teach us a lot about medicine, actually. If we can generalize the process by which we resurrect Hitler, we could save a lot of lives.
True, if resurrecting Hitler is a good idea and we can cause it to happen; if resurrecting Hitler is inevitable and we can ensure that he ends up being a good guy; or if resurrecting Hitler would be bad and we can prevent it from happening.
Do you suppose that developing a FAI will require at least some experience trying whatever works? I don’t know of any major computer programs that were written entirely before they were first compiled...
Edit: I see SoullessAutomaton has written a very similar comment.
But can we learn anything useful for a complete theory of intelligence based on something like EURISKO? Sure, it’s an ad hoc, throw things at the wall and see what sticks approach—but so are our brains, and if something like EURISKO can show limited, non-foomy levels of optimization power it would at least provide another crappy data point other than vertebrate brains on how intelligence works.
I used to think it’s useful to study ad-hoc attempts at AGI, but it now seems to me that knowledge of these chaotic things is both very likely a dead end, even for destroying the world, and of a wrong character to progress towards FAI.
I think one of the factors that contributes to interest in ad-hoc techniques is the prospect of a “thrilling discovery”. One is allowed to fantasize that all of their time and effort may pay off suddenly and unpredictably, which makes the research seem that much more fun and exciting. This is in contrast to a more formal approach in which understanding and progress are incremental by their very nature.
I bring this up because I see it as a likely underlying motive for arguments of the form “ad-hoc technique X is worth pursuing even though it’s not a formal approach”.
There are two kinds of scientific progress: the methodical experimentation and categorization which gradually extend the boundaries of knowledge, and the revolutionary leap of genius which redefines and transcends those boundaries. Acknowledging our debt to the former, we yearn nonetheless for the latter. - Academician Prokhor Zakharov, “Address to the Faculty”
No, it actually looks (just barely) feasible to get a FOOM out of something ad-hoc, and there are even good reasons for expecting that. But it doesn’t seem to be on the way towards deeper understanding. The best to hope for is catching it right when the FOOMing is imminent and starting to do serious theory, but the path of blind experimentation doesn’t seem to be the most optimal one even towards blind FOOM.
I didn’t mean to imply it was an invalid motive, merely a potential underlying motive. If it is valid in the sense that you mean (and I think it is), that’s just reason to scrutinize such claims even more closely.
Starting to seriously think about FAI and studying more rigorous system modeling techniques/theories changed my mind. There seems to be very little overlap between wild intuitions of ad-hoc AGI and technical challenges of careful inference/simulation or philosophical issues with formalizing decision theories for intelligence on overdrive.
Some of the intuitions from thinking about ad-hoc seem to carry over, but it’s just that: intuitions, and understanding of approaches to more careful modeling, even if they are applicable only on “toy” applications, gives deeper insight than knowledge of a dozen “real projects”. Intuitions gained from ad-hoc do apply, but only as naive clumsy caricatures.
Ad hoc AI is like ad hoc aircraft design. It flaps, it’s got wings, it has to fly, right? If we keep trying stuff, we’ll stumble across a wing that works. Maybe it’s the feathers?
Since such aircraft design actually worked, and produced aeroplanes before pure theory-based design, perhaps it’s not the best analogy. [Edit: Unless that was your point]
There are multiple concepts in the potential of ad-hoc. There is Strong AI, Good AI (Strong AI that has a positive effect), and Useful AI (Strong AI that can be used as a prototype or inspiration for Good AI, but can go Paperclip maximizer if allowed to grow). These concepts can be believed to be in quite different relations to each other.
Your irony states that there is no potential for any Strong AI in ad-hoc. Given that stupid evolution managed to get there, I think that with enough brute force of technology it’s quite feasible to get to Strong AI via this road.
Many reckless people working on AGI think that Strong AI is likely to also be a Good AI.
My previous position was that ad-hoc gives a good chance (in the near future) for Strong AI that is likely a Useful AI, but unlikely a Good AI. My current position is that ad-hoc has a small but decent chance (in the near future) for Strong AI, that is unlikely to be either Useful AI or Good AI.
Good AI is a category containing Friendly AI, that doesn’t require the outcome to be precisely right. This separates more elaborated concept of Friendly AI from an informal concept (requirement) of good outcome.
I believe the concepts are much more close than it seems, that is it’s hard to construct an AI that is not precisely Friendly, but still Good.
FAI is about being reliably harmless. Whether the outcome seems good in the short term is tangential. Even a “good” AI ought to be considered unfriendly if it’s opaque to proof—what can you possibly rely upon? No amount of demonstrated good behavior can be trusted. It could be insincere, it could be sincere but fatally misguided, it could have a flaw that will distort its goals after a few recursions. We would be stupid to just run it and see.
That’s what I had in mind, though I didn’t state it explicitly. It’s what I meant by ‘worked out’. It’s clear that you want these things worked out formally, as strong as being provably friendly.
I’m still skeptical on the world-destroying. My money’s on chaos to FOOM. Dynamism FTW. But then, I think AGI will come from robots.
What if its domain is restricted to math and self-modification? Then, if it fooms, it will be a safe math Oracle, possibly even provably safe. Then it would be a huge help in the road to FAI, both directly and as a case study.
It may very well be possible to build such an AI. However there are several issues with it:
The AI can be adapted for other, less restricted, domains if knowledge on how it works spreads. There would be a large incentive to since such an oracle would only be of limited utility.
The AI adds code that will evolve into another AI into it’s output. It’s remotely possible, depending on what kind of problems you have it working on. If you were using it to design more efficient algorithms, in some cases an AI of some form might be the optimal solution.
Even if you 100% trust the AI to provide the optimal output, you can’t trust that the optimal output to the problem you’ve specified is what you actually want.
The AI could self-modify incorrectly and result in unfriendly AI. In order to be provably friendly/restricted, it would have to be 100% certain of any modification. That’s a very tall order, especially in AI where everything has to be approximations or probabilistic.
It might not be as safe as you think it is. The AI runs some code and gets an unexpected result. Possibly because of a bug in the environment itself. Look up how difficult it is to sandbox untrusted code and you will get some appreciation for how a superintelligence could figure a way out of it’s box.
But it can’t do anything with any exploits it finds because it is restricted to hard-coded axioms? Well, maybe. If it’s using probabilities and some form of machine learning, it might be able to learn that “executing this code give me this result” and then learn to take advantage of that. I don’t believe that a system can work only in formal proofs. However I might be completely wrong about this one, it’s just a thought.
The AI can be adapted for other, less restricted, domains
That the ideas from a safe AI can be used to build an unsafe AI is a general argument against working on (or even talking about) any kind of AI whatsoever.
The AI adds code that will evolve into another AI into it’s output
The output is to contain only proofs of theorems. Specifically, a proof (or refutation) of the theorem in the input. The state of the system is to be reset after each run so as to not accumulate information.
The AI could self-modify incorrectly and result in unfriendly AI
Any correct or incorrect self-modification is still restricted to the math domain, and so cannot result in an unsafe AI.
bug in the environment itself
Guarding against software bugs is easy in this case. You design an abstract virtual machine environment for the AI, then design the software that implements this environment, then formally prove that the software is correct.
Guarding against errors caused by cosmic rays is also easy. You estimate the probability of such an error, and then add redundancy until the probability is so low that it won’t happen until the end of the universe.
Look up how difficult it is to sandbox untrusted code
Sandboxing untrusted code is easy. The difficult thing is sandboxing it while making it think that it runs normally. This is irrelevant here.
I don’t believe that a system can work only in formal proofs
It is quite probable that a “pure math Oracle” system cannot work. The point was, it can be made safe to try.
As we see here, the important difference between flying saucers and Unfriendly AI is that the latter belief is not harmless.
Eliezer, I know you’re not expecting your efforts in spreading parasite memes to result in snuffing out the future of intelligent life in the universe, but I will ask you to consider the possibility that you may be mistaken.
I notice you phrased this in terms of belief. I’m curious, what would you consider to be the minimum estimate of UFAI’s probability necessary to “reasonably” motivate concern or action?
If I’m right, the effect of widespread propagation of such memes will be to snuff out what chance of survival and success humanity might have had. Unlike UFAI which is pure science fiction, the strangling of progress is something that occurs—has occurred before—in real life.
What would you consider to be the minimum estimate of the probability that I’m right, necessary to “reasonably” motivate concern or action?
I’m not quite sure what “you being right” means here. Your thesis is that propagating the UFAI meme will suppress scientific and technological progress such as to contribute non-negligibly to our destruction?
I’m afraid I don’t have much background on how that’s supposed to work. If you can explain what you mean or point me to an existing explanation, I’ll try and give you an answer, rather than reactively throwing your question back at you.
Basically yes. Civilizations, species and worlds are mortal; there are rare long-lived species whose environment has remained unchanged for long periods of time, but the environment in which we evolved is long gone and our current one is not merely not stable, it is not even in equilibrium. And as long as we remain confined to one little planet running off a dwindling resource base and with everyone in weapon range of everyone else, there is nothing good about our long-term prospects. (For a fictional but eloquent discussion of some of the issues involved, see Permanence by Karl Schroeder.)
To change that, we need more advanced technology, for which we need software tools smart enough to help us deal with complexity. If our best minds start buying into the UFAI meme and turning away from building anything more ambitious than a social networking mashup, we may simply waste whatever chance we had. That is why UFAI belief is not as its proponents would have it the road of safety, but the road of oblivion.
rhollerith raised some reasonable objections to this response that I’d like to see answered, but I’ll try and answer your question without that information:
What would you consider to be the minimum estimate of the probability that I’m right, necessary to “reasonably” motivate concern or action?
As as far as concern goes, I think my threshold for concern over your proposition is identical to my threshold for concern over UFAI, as they postulate similar results (UFAI still seems marginally worse due to the chance of destroying intelligent alien life, but I’ll write this off as entirely negligible for the current discussion). I’d say 1:10,000 is a reasonable threshold for concern of the vocalized form, “hey, is anyone looking into this?” I’d love to see some more concrete discussion on this.
“Action” in your scenario is complicated by its direct opposition to acceptance of UFAI, so I can only give you some rough constraints. To simplify, I’ll assume all risks allow equally effective action to compensate for them, even though this is clearly not the case.
Let R = the scenario you’ve described, E = the scenario in which UFAI is a credible threat. “R and E” could be described as “damned if we do, damned if we don’t”, in which case action is basically futile, so I’ll consider the case where R and E are disjoint. In that case, action would only be justifiable if p(R) > p(E). My intuition says that such justification is proportional to p(R) - p(E), but I’d prefer more clarity in this step.
So that’s a rough answer… if T is my threshold probability for action in the face of existential risk, T (p(R) - p(E)) is my threshold for action on your scenario. If R and E aren’t disjoint, it looks something like T (p(R and ~E) - p(E and ~R)).
Though I’m not convinced “R and E” necessarily means “damned either way”. If I believed E in addition to R, I think what I would do is:
Forget about memetics in either direction as likely to do more harm than good, and concentrate all available resources on developing Friendly AI as reliably and quickly as possible.
However, provably Friendly AI is still not possible with 2009 vintage tools.
So I’d do it in stages, a series of self improving AIs, the early ones with low intelligence and crude Friendliness architecture, using them to develop better Friendliness architecture in tandem with increasing intelligence for the later ones. No guarantees, but if recursive self-improvement actually worked, I think that approach would have a reasonable chance of success.
rwallace has been arguing the position that AI researchers are too concerned (or will become too concerned) about the existential risk from UFAI. He writes that
we need software tools smart enough to help us deal with complexity.
rwallace: can we deal with complexity sufficiently well without new software that engages in strongly-recursive self-improvement?
Without new AGI software?
One part of the risk that rwallace says outweighs the risk of UFAI is that
we remain confined to one little planet . . . with everyone in weapon range of everyone else
The only response rwallace suggests to that risk is
we need more advanced technology, for which we need software tools smart enough to help us deal with complexity
rwallace: please give your reasoning for how more advanced technology decreases the existential risk posed by weapons more than it increases it.
Another part of the risk that rwallace says outweighs the risk of UFAI is that
we remain confined to one little planet running off a dwindling resource base
Please explain how dwindling resources presents a significant existential risk. I can come up with several argument, but I’d like to see the one or two you consider the strongest arguments.
If we have uploads we can get off the planet and stay in space for a fraction of the resources it currently costs to do manned space flight. We can spread ourselves between the stars.
But an upload might go foom, so we should stop all upload research.
It is this kind of conundrum I see humanity in at the moment.
First, an upload isn’t going to “go foom”: a digital substrate doesn’t magically confer superpowers, and early uploads will likely be less powerful than their biological counterparts in several ways.
Second, stopping upload research is not the path of safety, because ultimately we must advance or die.
First, an upload isn’t going to “go foom”: a digital substrate doesn’t magically confer superpowers, and early uploads will likely be less powerful than their biological counterparts in several ways.
Foom is about rate of power increase, not initial power level. Copy/paste isn’t everything, but still a pretty good superpower.
Second, stopping upload research is not the path of safety, because ultimately we must advance or die.
It’s not at all obvious to me that the increased risk of stagnation death outweighs the reduced risk of foom death.
Your third point is valid, but your first is basicallywrong; our environments occupy a small and extremely regular subset of the possibility space, so that success on a certain few tasks seems to correlate extremely well with predicted success across plausible future domains. Measuring success on these tasks is something AIs can easily do; EURISKO accomplished it in fits and starts. More generally, intelligence isn’t magical: if there’s any way we can tell whether a change in an AGI represents a bug or an improvement, then there’s an algorithm that an AI can run to do the same.
As for the second problem, one idea that may not have occurred to you is that an AI could write a future version of itself, bug-check and test out various subsystems and perhaps even the entire thing on a virtual machine first, and then shut itself down and start up the successor. If there’s a way for Lenat to see that EURISKO isn’t working properly and then fix it, then there’s a way for AI (version N) to see that AI (version N+1) isn’t working properly and fix it before making the change-over.
In those posts you are arguing something different from what I was talking about. Sure chimps will never make better technology than humans, but sometimes making more advanced clever technology is not what you want to do and be positively detrimental to your chances of shaping the world to a desirable state. The arms race for nuclear weapons for example or bio-weapons research.
If humans manage to invent a virus that wipes us out, would you still call that intelligent? If so it is not that sort of intelligence we need to create… we need to create things that win in the end, not have short term wins and then destroy itself.
“More generally, intelligence isn’t magical: if there’s any way we can tell whether a change in an AGI represents a bug or an improvement, then there’s an algorithm that an AI can run to do the same.”
Except that we don’t—can’t—do it by pure armchair thought, which is what the recursive self-improvement proposal amounts to.
The approach of testing a new version in a sandbox had occurred to me, and I agree it is a very promising one for many things—but recursive self-improvement isn’t among them! Consider, what’s the primary capability for which version N+1 is being tested? Why, the ability to create version N+2… which involves testing N+2… which involves creating N+3… etc.
Again, there’s enough correlation between ability to perform certain tasks that you don’t need an infinite recursion. To test AIv(N+1)‘s ability to program to exact specification, instead of having it program AIv(N+2) have it instead program some other things that AIvN finds difficult (but whose solutions are within AIvN’s power to verify). That we will be applying AIv(N+1)’s precision programming to itself doesn’t mean we can’t test it on non-recursive data first.
ETA: Of course, since we want the end result to be a superintelligence, AIvN might also ask AIv(N+1) for verifiable insight into an array of puzzling questions, some of which AIvN can’t figure out but suspects are tractable with increased intelligence.
If you observed something to work 15 times, how do you know that it’ll work 16th time? You obtain a model of increasing precision with each test, that lets you predict what happens next, on a test you haven’t performed yet. The same way, you can try to predict what happens on the first try, before any observations took place.
Another point is that testing can be a part of the final product: instead of building a working gizmo, you build a generic self-testing adaptive gizmo that finds the right parameters itself, and that is pre-designed to do that in the most optimal way.
Unfortunately there are three fundamental problems, any one of which would suffice to make this approach unworkable even in principle.
Keep the bolded context in mind.
What constitutes an improvement is not merely a property of the program, but also a property of the world. Even if an AI program could find improvements, it could not distinguish them from bugs.
There are large classes of problem for which this is not the case. For example, “make accurate predictions about future sensory data based on past sensory data” relates program to world, but optimizing for this task is a property of the time, memory, and accuracy trade-offs involved. At the very least, your objection fails to apply “even in principle”.
Any significant AI program is the best effort of one or more highly skilled human programmers. To qualitatively improve on that work, the program would have to be at least as smart as the humans, so even if recursive self-improvement were workable, which it isn’t, it would be irrelevant until after human level intelligence is reached by other means.
This depends on the definition of “qualitatively improve”. It seems Eurisko improved itself in important ways that Lenat couldn’t have done by hand, so I think this too fails the “even in principle” test.
Any self-modifying system relies, for the ability to constructively modify upper layers of content, on unchanging lower layers of architecture. This is a general principle that applies to all complex systems from engineering to biology to politics.
This seems the most reasonable objection of the three. Interestingly enough, Eliezer claims it’s this very difference that makes self-improving AI unique.
I think this also fails to apply universally, for somewhat more subtle reasons involving the nature of software. A self-improving AI possessing precise specifications of its components has no need for a constant lower layer of architecture (except if you mean the hardware, which is a very different question). The Curry-Howard isomorphism offers a proof-of-concept here: An AI composed of precise logical propositions can arbitrarily rewrite its proofs/programs with any other program of an equivalent type. If you can offer a good argument for why this is impossible in principle, I’d be interested in hearing it.
“It seems Eurisko improved itself in important ways that Lenat couldn’t have done by hand”
As far as I can see, the only self improvements it came up with were fairly trivial ones that Lenat could easily have done by hand. Where it came up with important improvements that Lenat wouldn’t have thought of by himself was in the Traveler game—a simple formal puzzle, fully captured by a set of rules that were coded into the program, making the result a success in machine learning but not self-improvement.
“The Curry-Howard isomorphism offers a proof-of-concept here”
Indeed this approach shows promise, and is the area I’m currently working on. For example if you can formally specify an indexing algorithm, you are free to look for optimized versions provided you can formally prove they meet the specification. If we can make practical tools based on this idea, we could make software engineering significantly more productive.
But this can only be part of the story. Consider the requirement that a program have an intuitive user interface. We have nothing remotely approaching the ability to formally specify this, nor could an AI ever come up with such by pure introspection because it depends on entities that are not part of the AI. Nor, obviously, is a formal specification of human psychology the kind of thing that would ever be approached accidentally by experimentation with Eurisko-style programs as was the original topic. And if the science of some future century—that would need to be to today’s science as the latter is to witchcraft—ever does manage to accomplish such, why then, that would be the key to enabling the development of provably Friendly AI.
Where it came up with important improvements that Lenat wouldn’t have thought of by himself was in the Traveler game—a simple formal puzzle, fully captured by a set of rules that were coded into the program, making the result a success in machine learning but not self-improvement.
...And applied these improvements to the subsequent modified set of rules. “That was machine learning, not self-improvement” sounds like a fully general counter-argument, especially considering your skepticism toward the very idea of self-improvement. Perhaps you can clarify the distinction?
Consider the requirement that a program have an intuitive user interface. We have nothing remotely approaching the ability to formally specify this, nor could an AI ever come up with such by pure introspection because it depends on entities that are not part of the AI.
An AI is allowed to learn from its environment, no one’s claiming it will simply meditate on the nature of being and then take over the universe. That said, this example has nothing to do with UFAI. A paperclip maximizer has no need for an intuitive user interface.
And if [science] ever does manage to accomplish [a formal specification of human psychology], why then, that would be the key to enabling the development of provably Friendly AI.
Indeed! Sadly, such a specification is not required to interact with and modify one’s environment. Humans were killing chimpanzees with stone tools long before they even possessed the concept of “psychology”.
I’ll call it self improvement when a substantial, nontrivial body of code is automatically developed, that is applicable to domains other than gameplaying, as opposed to playing a slightly different version of the same game. (Note that it was substantially the same strategy that won the second time as the first time, despite the rules changes.)
“A paperclip maximizer has no need for an intuitive user interface.”
True if you’re talking about something like Galactus that begins the story already possessing the ability to eat planets. However, UFAI believers often talk about paperclip maximizers being able to get past firewalls etc. by verbal persuasion of human operators. That certainly isn’t going to happen without a comprehensive theory of human psychology.
“Strongly-recursive self-improvement” is a figment of the imagination; among the logical errors involved is confusion between properties of a program and properties of the world.
As for the rest: do you believe humanity can survive permanently as we are now, confined to this planet? If you do, then I will point you to the geological evidence to the contrary. If not, then it follows that without more advanced technology, we are dead. Neither I nor anybody else can know what will be the proximate cause of death for the last individual, or in what century, but certain extinction is certain extinction nonetheless.
Let us briefly review the discussion up to now since many readers use the the comments page which does not provide much context. rwallace has been arguing that AI researchers are too concerned (or will become too concerned) about the existential risk from reimplementing EURISKO and things like that.
You have mentioned two or three times, rwallace, that without more advanced technology, humans will eventually go extinct. (I quote one of those 2 or 3 mentions below.) You mention that to create and to manage that future advanced technology, civilization will need better tools to manage complexity. Well, I see one possible objection to your argument right there, in that better science and better technology might well decrease the complexity of the cultural information humans are required to keep on top of. Consider that once Newton gave our civilization a correct theory of dynamics, almost all of the books written before Newton on dynamics could safely be thrown away (the exceptions being books by Descartes and Galileo that help people understand Newton and put Newton in historical context) which of course constitutes a net reduction in the complexity of the cultural information that our civilization has to keep on top of. (If it does not seem like a reduction, that is because the possession of Newtonian dynamical theory made our civilization more ambitious about what goals to try for.)
do you believe humanity can survive permanently as we are now, confined to this planet? If you do, then I will point you to the geological evidence to the contrary. If not, then it follows that without more advanced technology, we are dead.
But please explain to me what your argument has to do with EURISKO and things like that: is it your position that the complexity of future human culture can be managed only with better AGI software?
And do you maintain that that software cannot be developed fast enough by AGI researchers such as Eliezer who are being very careful about existential risks?
In general, the things you argue are dangerous are slow dangers. You yourself refer to “geological evidence” which suggests that they are dangerous on geological timescales.
In contrast, research into certain areas of AI seems to me genuinely fast dangers: things with a high probability of wiping out our civilization in the next 30, 50 or 100 years. It seems unwise to increase fast dangers to decrease slow dangers. But I suppose you disagree that AGI research if not done very carefully is a fast danger. (I’m still studying your arguments on that.)
Reduction in complexity is at least conceivable, I’ll grant. For example if someone invented a zero point energy generator with the cost and form factor of an AA battery, much of the complexity associated with the energy industry could disappear.
But this seems highly unlikely. All the current evidence suggests the contrary: the breakthroughs that will be necessary and possible in the coming century will be precisely those of complex systems (in both senses of the term).
Human level AGI in the near future is indeed neither necessary nor possible. But there is a vast gap between that and what we have today, and we will, yes, need to fill some of that gap. Perhaps a key breakthrough would have come from a young researcher who would have re-implemented Eurisko and from the experiment acquired a critical jump in understanding—and who has now quietly left, thinking Eurisko might blow up the world, to reconsider that job offer from Electronic Arts.
I do disagree that AGI research is a fast danger. I will grant you that there is a sense in which the dangers I am worried about are slow ones - barring unlikely events like a large asteroid impact (which is likely only over longer time scales), I’m confident humanity will still exist 100 years from now.
But our window of opportunity may not. Consider that civilizations are mortal, for reasons unrelated to this conversation. Consider that environments conducive to scientific progress are even considerably rarer and more transient than civilization itself. Consider also that the environment in which our civilization arose is gone, and is not coming back. (For the simplest example, while fossil fuels still exist, the easily accessible deposits thereof, so important for bootstrapping, are largely gone.) I think it quite possible that the 21st-century may be the last hard step in the Great Filter, that by the year 2100 the ultimate fate of humanity may in fact have been decided, even if nobody on that date yet knows it. I cannot of course be certain of this, but I think it likely enough that we cannot afford to risk wasting this window of opportunity.
One problem with this argument is how conjunctive it is: “(A) Progress crucially depends on breakthroughs in complexity management and (B) strong recursive self-improvement is impossible and (C) near-future human level AGI is neither dangerous nor possible but (D) someone working on it is crucial for said complexity management breakthroughs and (E) they’re dissuaded by friendliness concerns and (F) our scientific window of opportunity is small.”
My back-of-the-envelope, generous probabilities:
A. 0.5, this is a pretty strong requirement.
B. 0.9, for simplicity, giving your speculation the benefit of the doubt.
C. 0.9, same.
D. 0.1, a genuine problem of this magnitude is going to attract a lot of diverse talent.
E. 0.01, this is the most demanding element of the scenario, that the UFAI meme itself will crucially disrupt progress.
F. 0.05, this would represent a large break from our current form of steady scientific progress, and I haven’t yet seen much evidence that it’s terribly likely.
That product comes out to roughly 1:50,000. I’m guessing you think the actual figure is higher, and expect you’ll contest those specific numbers, but would you agree that I’ve fairly characterized the structure of your objection to FAI?
Neither is a belief in flying saucers, taken seriously, for most values of “harmless”.
To be clear, you are saying that you consider spontaneous FAI overwhelmingly likely, and thus consider any failure to promote FOOMing both a humanitarian catastrophe and a gateway to other existential risks? Because that seems like a claim worth proving.
...aaaand that’s why I don’t go around discussing the danger paths until someone (who I can realistically influence) actually starts to advocate going down them. Plenty of idiots to take it as an instruction manual. So I discuss the safe path but make no particular advance effort to label the dangerous ones.
I am rather surprised that you accept all of the claimed achievements of Eurisko and even regard it as “dangerous”, despite the fact that no one save the author has ever seen even a fragment of its source code. I firmly believe that we are dealing with a “mechanical Turk.”
I am also curious why you believe that meaningful research on Friendly AI is at all possible without prior exposure to a working AGI. To me it seems a bit like trying to invent the ground fault interrupter before having discovered electricity.
Aside from that: If I had been following your writings more carefully, I might already have learned the answer to this, but: just why do you prioritize formalizing Friendly AI over achieving AI in the fist place? You seem to side with humanity over a hypothetical Paperclip Optimizer. Why is that? It seems to me that unaugmented human intelligence is itself an “unfriendly (non-A)I”, quite efficient at laying waste to whatever it touches.
There is every reason to believe that if an AGI does not appear before the demise of cheap petroleum, our species is doomed to “go out with a whimper.” I for one prefer the “bang” as a matter of principle.
I would gladly accept taking a chance at conversion to paperclips (or some similarly perverse fate at the hands of an unfriendly AGI) when the alternative appears to be the artificial squelching of the human urge to discover and invent, with the inevitable harvest of stagnation and eventually oblivion.
I accept Paperclip Optimization (and other AGI failure modes) as an honorable death, far superior to being eaten away by old age or being killed by fellow humans in a war over dwindling resources. I want to live in interesting times. Bring on the AGI. It seems to me that if any intelligence, regardless of its origin, is capable of wrenching the universe out of our control, it deserves it.
Why is the continued hegemony of Neolithic flesh-bags so precious to you?
Aside from that: If I had been following your writings more carefully, I might already have learned the answer to this, but: just why do you prioritize formalizing Friendly AI over achieving AI in the fist place?
It seems to me that if any intelligence, regardless of its origin, is capable of wrenching the universe out of our control, it deserves it.
I don’t think you understand the paperclip maximizer scenario. An UnFriendly AI is not necessarily conscious; it’s just this device that tiles the light cone with paperclips. Arguably it helps to say “really powerful optimization process” rather than “intelligence.” Consider that we would not say that a thermostat deserves to control the temperature of a room, even if we happen to be locked in the room and are to be roasted to death.
I’d like to take my chances of being cooked vs. a world without furnaces, thermostats or even rooms—something I believe we’re headed for by default, in the very near future.
I will accept that “AGI-now” proponents should carry the blame for a hypothetical Paperclip apocalypse when Friendliness proponents accept similar blame for an Earth-bound humanity flattened by a rogue asteroid (or leveled by any of the various threats a superintelligence—or, say, the output of a purely human AI research community unburdened by Friendliness worries—might be able to counter. I previously gave Orlov’s petrocollapse as yet another example.)
ASCII—the onus is on you to give compelling arguments that the risks you are taking are worth it
Status quo bias, anyone?
I presently believe, not without justification, that we are headed for extinction-level disaster as things are; and that not populating the planet with the highest achievable intelligence is in itself an immediate existential risk. In fact, our current existence may well be looked back on as an unthinkably horrifying disaster by a superintelligent race (I’m thinking of Yudkowsky’s Super-Happies.)
Since your justification is omitted here, I’ll go ahead and suspect it’s at least as improbable as this one. The question isn’t simply “do we need better technology to mitigate existential risk”, it’s “are the odds that technological suppression due to friendliness concerns wipes us out greater than the corresponding AGI risk”.
If you assume friendliness is not a problem, AI is obviously a beneficial development. Is that really the major concern here? All this talk of the benefits of scientific and technological progress seems wasted. Take friendliness out of the picture, I doubt many here would disagree with the general point that progress mitigates long-term risk.
So please, be more specific. The argument “lack of progress contributes to existential risk” contains no new information. Either tell us why this risk is far greater than we suspect, or why AGI is less risky than we suspect.
This is a road that does not lead to Friendly AI, only to AGI. I doubt this has anything to do with Lenat’s motives—but I’m glad the source code isn’t published and I don’t think you’d be doing a service to the human species by trying to reimplement it.
A stunning proof-of-concept for AI, with the source code lost to the mists of time; then immersion in an apparently massive dead-end project. Is anyone else worried that Lenat might secretly have a phase 3 to his plan?
Seems too obvious.
Exactly.
You may stop worrying for the moment. I just tried to wade through the papers RichardKennaway has put up, and it seems reimplementing EURISKO given Lenat’s most detailed descriptions of it will likely be a big creative endeavor. Download them, read them and then tell me just one thing: do you now know (even approximately) what the main loop looks like, or not? Cause I couldn’t make it out in half an hour’s reading.
Are you really afraid that AI is so easy that it’s a very short distance between “ooh, cool” and “oh, shit”?
Depends how cool. I don’t know the space of self-modifying programs very well. Anything cooler than anything that’s been tried before, even marginally cooler, has a noticeable subjective probability of going to shit. I mean, if you kept on making it marginally cooler and cooler, it’d go to “oh, shit” one day after a sequence of “ooh, cools” and I don’t know how long that sequence is.
This means we should feel pretty safe, since AI does not appear to be making even incremental progress.
Really, it’s hard for anyone who is well-versed in the “state of the art” of AI to feel any kind of alarm about the possibility of an imminent FOOM. Take a look at this paper. Skim through the intro, note the long and complicated reinforcement learning algorithm, and check out the empirical results section. The test domain involves a monkey in a 5x5 playroom. There are some fun little complications, like a light switch and a bell. Note that these guys are top-class (Andrew Barto basically invented RL), and the paper was published at one of the top-tier machine learning conferences (NIPS), in 2005.
Call me a denier, but I just don’t think the monkey is going to bust out of his playroom and take over the world. At least, not anytime soon.
Taking progress in AI to mean more real world effectiveness:
Intelligence seems to have jumps in real world effectiveness, e.g. the brains of great apes and humans are very similar, the difference in effectiveness is obvious.
So coming to the conclusion that we are fine based on the state of the art not being any more effective (not making progress) would be very dangerous. Perhaps tomorrow, some team of AI researchers will combine the current state of the art solutions in just the right way, resulting in a massive jump in real world effectiveness? maybe enough to have an “oh, shit” moment?
Regardless of the time frame, if the AI community is working towards AGI rather than FAI, we will likely have (eventually) an AI go FOOM or at the very least, and “oh, shit” moment (I’m not sure if they are equivalent).
Zing!
Also, good point, but this post is designed produce such progress, is it not?
Does the specific distance even matter? UFAI vs FAI is zero sum, and we have no idea how long FAI will take us. Any progress toward AGI that isn’t “matched” by progress toward FAI is regressive, even if AGI is still 100 years off.
I am.
I just site Googled lesswrong.org and overcomingbias.com, and my best guess so far is that AGI = Artificial General Intelligence. (Augmented?) Is that what AGI stands for? Non site-specific google results are definitely not the applicable ones. Apparently it has been used as an abbreviation in this community for quite awhile, but finding out what it stands for takes a few jumps. (Should there be an easily found FAQ? One is not obvious here or on the wiki. Or is a certain degree of obscurity desired?)
It’s Artificial general intelligence, I added a redirect and a stub page on the wiki.
Yeah, AGI would be Artificial General Intelligence, where the term “General” is to contrast it with much of current AI work which, well, isn’t.
ie, a human, for instance, can figure out how to play an instrument, do deep number theory, build an engine, etc etc etc etc...
Security By Obscurity: When you can’t be bothered to implement a real solution!(tm)
And what, pray tell, does a “real solution” look like?
Seeing the recent thread necromancy, it looks like this is a much more important question than I realized at the time, since it bears on AI-related existential risk.
The question, to summarize, was, “How exactly do you keep a good AGI prospect from being developed into unFriendly AGI, other than Security by Obscurity?”
My answer is that SbO (Security by Obscurity, not Antimony(II) Oxide, which doesn’t even exist) is not a solution here for the same reason it’s criticized everywhere else (which I assume is that it increases the probability of a rogue outsmarting mainstream researchers). Better to let the good guys be as well informed as the bad guys so they can deploy countermeasures (their own AGI) when the bad guys develop theirs.
But then, I haven’t researched this 24⁄7 for the last several years, so this may be too trite a dismissal.
Lazy question: Did you explain this in an OB post? I have a crappy mental index of the topics you covered, and Google didn’t yield any immediate results, just an explanation of why Eurisko didn’t foom. If not, I’d love to see a post explaining how you’re extrapolating the Eurisko approach into something incompatible with FAI. Or are you just applying the general rule, “AGI that’s not explicitly friendly is explicitly unfriendly”?
Yeah, he means that. “Please don’t work on AGI until you’ve worked out FAI.”
ETA: read Eliezer’s reply.
Not exactly, Thom. Roughly, for FAI you need precise self-modification. For precise self-modification, you need a precise theory of the intelligence doing the self-modification. To get to FAI you have to walk the road that leads to precise theories of intelligence—something like our present-day probability theory and decision theory, but more powerful and general and addressing issues these present theories don’t.
Eurisko is the road of self-modification done in an imprecise way, ad-hoc, throwing together whatever works until it gets smart enough to FOOM. This is a path that leads to shattered planets, if it were followed far enough. No, I’m not saying that Eurisko in particular is far enough, I’m saying that it’s a first step along that path, not the FAI path.
Perhaps a writeup of what you have discovered, or at least surmise, about walking that road would encourage bright young minds to work on those puzzles instead of reimplementing Eurisko.
It’s not immediately clear that studying and playing with specific toy self-referential systems won’t lead to ideas that might apply to precise members of that class.
I’ve written up some of the concepts of precise self-modification, but need to collect the posts on a Wiki page on “lawfulness of intelligence” or something.
Any of these posts ever go up?
Cf. “Lawful intelligence.”
Well, in this sense computing is also a first step on that path, Moore’s law of mad science and all. Eurisko in particular doesn’t seem to deserve more mention than that.
Doesn’t seem to deserve more mention than the creation of computing? Sure. But computing has already been created.
Um, so has Eurisko.
...indeed. It seems that I failed to figure out just what I was arguing against. Let me re-make that point.
As far as first steps along that path go, they have already been taken: we have gone from a world without computers to a world with one, and we can’t reverse that. The logical place to focus our efforts would seem to be the next step which has not been taken, which could very well be reimplementing EURISKO. (Though it could also very well be running a neural net on a supercomputer or some guy making the video game “Operant Conditioning Hero”.)
We have gone from a world without dictators to a world with one, and we can’t reverse that. The logical place to focus our efforts would seem to be the next step which has not been taken, which could very well be resurrecting Hitler.
Seriously? I did not think a discussion of Eurisko could be Godwinned. Bravo.
While grandparent was probably a miscalculation of some sort, I feel that mentioning Hitler is more acceptable if the context is Nazi super science than outrage maximization.
Grandparent was probably a miscalculation of some sort, but I think mention of Hitler is acceptable if the context is Nazi super science rather than outrage maximization.
Resurrecting Hitler would probably teach us a lot about medicine, actually. If we can generalize the process by which we resurrect Hitler, we could save a lot of lives.
True, if resurrecting Hitler is a good idea and we can cause it to happen; if resurrecting Hitler is inevitable and we can ensure that he ends up being a good guy; or if resurrecting Hitler would be bad and we can prevent it from happening.
Do you suppose that developing a FAI will require at least some experience trying whatever works? I don’t know of any major computer programs that were written entirely before they were first compiled...
Edit: I see SoullessAutomaton has written a very similar comment.
But can we learn anything useful for a complete theory of intelligence based on something like EURISKO? Sure, it’s an ad hoc, throw things at the wall and see what sticks approach—but so are our brains, and if something like EURISKO can show limited, non-foomy levels of optimization power it would at least provide another crappy data point other than vertebrate brains on how intelligence works.
I used to think it’s useful to study ad-hoc attempts at AGI, but it now seems to me that knowledge of these chaotic things is both very likely a dead end, even for destroying the world, and of a wrong character to progress towards FAI.
I think one of the factors that contributes to interest in ad-hoc techniques is the prospect of a “thrilling discovery”. One is allowed to fantasize that all of their time and effort may pay off suddenly and unpredictably, which makes the research seem that much more fun and exciting. This is in contrast to a more formal approach in which understanding and progress are incremental by their very nature.
I bring this up because I see it as a likely underlying motive for arguments of the form “ad-hoc technique X is worth pursuing even though it’s not a formal approach”.
Upvoted for alpha centuri reference. God I love that game!
No, it actually looks (just barely) feasible to get a FOOM out of something ad-hoc, and there are even good reasons for expecting that. But it doesn’t seem to be on the way towards deeper understanding. The best to hope for is catching it right when the FOOMing is imminent and starting to do serious theory, but the path of blind experimentation doesn’t seem to be the most optimal one even towards blind FOOM.
That doesn’t contradict what logi said. It could still be a motive.
It could, but it won’t be invalid motive, as I (maybe incorrectly) heard implied.
I didn’t mean to imply it was an invalid motive, merely a potential underlying motive. If it is valid in the sense that you mean (and I think it is), that’s just reason to scrutinize such claims even more closely.
What changed your mind?
Starting to seriously think about FAI and studying more rigorous system modeling techniques/theories changed my mind. There seems to be very little overlap between wild intuitions of ad-hoc AGI and technical challenges of careful inference/simulation or philosophical issues with formalizing decision theories for intelligence on overdrive.
Some of the intuitions from thinking about ad-hoc seem to carry over, but it’s just that: intuitions, and understanding of approaches to more careful modeling, even if they are applicable only on “toy” applications, gives deeper insight than knowledge of a dozen “real projects”. Intuitions gained from ad-hoc do apply, but only as naive clumsy caricatures.
Ad hoc AI is like ad hoc aircraft design. It flaps, it’s got wings, it has to fly, right? If we keep trying stuff, we’ll stumble across a wing that works. Maybe it’s the feathers?
Since such aircraft design actually worked, and produced aeroplanes before pure theory-based design, perhaps it’s not the best analogy. [Edit: Unless that was your point]
There are multiple concepts in the potential of ad-hoc. There is Strong AI, Good AI (Strong AI that has a positive effect), and Useful AI (Strong AI that can be used as a prototype or inspiration for Good AI, but can go Paperclip maximizer if allowed to grow). These concepts can be believed to be in quite different relations to each other.
Your irony states that there is no potential for any Strong AI in ad-hoc. Given that stupid evolution managed to get there, I think that with enough brute force of technology it’s quite feasible to get to Strong AI via this road.
Many reckless people working on AGI think that Strong AI is likely to also be a Good AI.
My previous position was that ad-hoc gives a good chance (in the near future) for Strong AI that is likely a Useful AI, but unlikely a Good AI. My current position is that ad-hoc has a small but decent chance (in the near future) for Strong AI, that is unlikely to be either Useful AI or Good AI.
BTW, none of the above classifications are “friendly”.
Good AI is a category containing Friendly AI, that doesn’t require the outcome to be precisely right. This separates more elaborated concept of Friendly AI from an informal concept (requirement) of good outcome.
I believe the concepts are much more close than it seems, that is it’s hard to construct an AI that is not precisely Friendly, but still Good.
FAI is about being reliably harmless. Whether the outcome seems good in the short term is tangential. Even a “good” AI ought to be considered unfriendly if it’s opaque to proof—what can you possibly rely upon? No amount of demonstrated good behavior can be trusted. It could be insincere, it could be sincere but fatally misguided, it could have a flaw that will distort its goals after a few recursions. We would be stupid to just run it and see.
At which point you are starting to think of what it takes to make not just informally “Good” AI, but an actually Friendly AI.
Right.
That’s what I had in mind, though I didn’t state it explicitly. It’s what I meant by ‘worked out’. It’s clear that you want these things worked out formally, as strong as being provably friendly.
I’m still skeptical on the world-destroying. My money’s on chaos to FOOM. Dynamism FTW. But then, I think AGI will come from robots.
Why are you still thinking it has any potential? (Assuming it doesn’t, the above comment sounds ridiculous.)
Does this imply that you think there’s a significant risk of EURISKO recursively self-improving, or do you discourage it for other reasons?
What if its domain is restricted to math and self-modification? Then, if it fooms, it will be a safe math Oracle, possibly even provably safe. Then it would be a huge help in the road to FAI, both directly and as a case study.
It may very well be possible to build such an AI. However there are several issues with it:
The AI can be adapted for other, less restricted, domains if knowledge on how it works spreads. There would be a large incentive to since such an oracle would only be of limited utility.
The AI adds code that will evolve into another AI into it’s output. It’s remotely possible, depending on what kind of problems you have it working on. If you were using it to design more efficient algorithms, in some cases an AI of some form might be the optimal solution.
Even if you 100% trust the AI to provide the optimal output, you can’t trust that the optimal output to the problem you’ve specified is what you actually want.
The AI could self-modify incorrectly and result in unfriendly AI. In order to be provably friendly/restricted, it would have to be 100% certain of any modification. That’s a very tall order, especially in AI where everything has to be approximations or probabilistic.
It might not be as safe as you think it is. The AI runs some code and gets an unexpected result. Possibly because of a bug in the environment itself. Look up how difficult it is to sandbox untrusted code and you will get some appreciation for how a superintelligence could figure a way out of it’s box.
But it can’t do anything with any exploits it finds because it is restricted to hard-coded axioms? Well, maybe. If it’s using probabilities and some form of machine learning, it might be able to learn that “executing this code give me this result” and then learn to take advantage of that. I don’t believe that a system can work only in formal proofs. However I might be completely wrong about this one, it’s just a thought.
That the ideas from a safe AI can be used to build an unsafe AI is a general argument against working on (or even talking about) any kind of AI whatsoever.
The output is to contain only proofs of theorems. Specifically, a proof (or refutation) of the theorem in the input. The state of the system is to be reset after each run so as to not accumulate information.
Any correct or incorrect self-modification is still restricted to the math domain, and so cannot result in an unsafe AI.
Guarding against software bugs is easy in this case. You design an abstract virtual machine environment for the AI, then design the software that implements this environment, then formally prove that the software is correct. Guarding against errors caused by cosmic rays is also easy. You estimate the probability of such an error, and then add redundancy until the probability is so low that it won’t happen until the end of the universe.
Sandboxing untrusted code is easy. The difficult thing is sandboxing it while making it think that it runs normally. This is irrelevant here.
It is quite probable that a “pure math Oracle” system cannot work. The point was, it can be made safe to try.
As we see here, the important difference between flying saucers and Unfriendly AI is that the latter belief is not harmless.
Eliezer, I know you’re not expecting your efforts in spreading parasite memes to result in snuffing out the future of intelligent life in the universe, but I will ask you to consider the possibility that you may be mistaken.
I notice you phrased this in terms of belief. I’m curious, what would you consider to be the minimum estimate of UFAI’s probability necessary to “reasonably” motivate concern or action?
If I’m right, the effect of widespread propagation of such memes will be to snuff out what chance of survival and success humanity might have had. Unlike UFAI which is pure science fiction, the strangling of progress is something that occurs—has occurred before—in real life.
What would you consider to be the minimum estimate of the probability that I’m right, necessary to “reasonably” motivate concern or action?
I’m not quite sure what “you being right” means here. Your thesis is that propagating the UFAI meme will suppress scientific and technological progress such as to contribute non-negligibly to our destruction?
I’m afraid I don’t have much background on how that’s supposed to work. If you can explain what you mean or point me to an existing explanation, I’ll try and give you an answer, rather than reactively throwing your question back at you.
Basically yes. Civilizations, species and worlds are mortal; there are rare long-lived species whose environment has remained unchanged for long periods of time, but the environment in which we evolved is long gone and our current one is not merely not stable, it is not even in equilibrium. And as long as we remain confined to one little planet running off a dwindling resource base and with everyone in weapon range of everyone else, there is nothing good about our long-term prospects. (For a fictional but eloquent discussion of some of the issues involved, see Permanence by Karl Schroeder.)
To change that, we need more advanced technology, for which we need software tools smart enough to help us deal with complexity. If our best minds start buying into the UFAI meme and turning away from building anything more ambitious than a social networking mashup, we may simply waste whatever chance we had. That is why UFAI belief is not as its proponents would have it the road of safety, but the road of oblivion.
rhollerith raised some reasonable objections to this response that I’d like to see answered, but I’ll try and answer your question without that information:
As as far as concern goes, I think my threshold for concern over your proposition is identical to my threshold for concern over UFAI, as they postulate similar results (UFAI still seems marginally worse due to the chance of destroying intelligent alien life, but I’ll write this off as entirely negligible for the current discussion). I’d say 1:10,000 is a reasonable threshold for concern of the vocalized form, “hey, is anyone looking into this?” I’d love to see some more concrete discussion on this.
“Action” in your scenario is complicated by its direct opposition to acceptance of UFAI, so I can only give you some rough constraints. To simplify, I’ll assume all risks allow equally effective action to compensate for them, even though this is clearly not the case.
Let R = the scenario you’ve described, E = the scenario in which UFAI is a credible threat. “R and E” could be described as “damned if we do, damned if we don’t”, in which case action is basically futile, so I’ll consider the case where R and E are disjoint. In that case, action would only be justifiable if p(R) > p(E). My intuition says that such justification is proportional to p(R) - p(E), but I’d prefer more clarity in this step.
So that’s a rough answer… if T is my threshold probability for action in the face of existential risk, T (p(R) - p(E)) is my threshold for action on your scenario. If R and E aren’t disjoint, it looks something like T (p(R and ~E) - p(E and ~R)).
A fair answer, thanks.
Though I’m not convinced “R and E” necessarily means “damned either way”. If I believed E in addition to R, I think what I would do is:
Forget about memetics in either direction as likely to do more harm than good, and concentrate all available resources on developing Friendly AI as reliably and quickly as possible.
However, provably Friendly AI is still not possible with 2009 vintage tools.
So I’d do it in stages, a series of self improving AIs, the early ones with low intelligence and crude Friendliness architecture, using them to develop better Friendliness architecture in tandem with increasing intelligence for the later ones. No guarantees, but if recursive self-improvement actually worked, I think that approach would have a reasonable chance of success.
rwallace has been arguing the position that AI researchers are too concerned (or will become too concerned) about the existential risk from UFAI. He writes that
rwallace: can we deal with complexity sufficiently well without new software that engages in strongly-recursive self-improvement?
Without new AGI software?
One part of the risk that rwallace says outweighs the risk of UFAI is that
The only response rwallace suggests to that risk is
rwallace: please give your reasoning for how more advanced technology decreases the existential risk posed by weapons more than it increases it.
Another part of the risk that rwallace says outweighs the risk of UFAI is that
Please explain how dwindling resources presents a significant existential risk. I can come up with several argument, but I’d like to see the one or two you consider the strongest arguments.
If we have uploads we can get off the planet and stay in space for a fraction of the resources it currently costs to do manned space flight. We can spread ourselves between the stars.
But an upload might go foom, so we should stop all upload research.
It is this kind of conundrum I see humanity in at the moment.
I agree, and will add:
First, an upload isn’t going to “go foom”: a digital substrate doesn’t magically confer superpowers, and early uploads will likely be less powerful than their biological counterparts in several ways.
Second, stopping upload research is not the path of safety, because ultimately we must advance or die.
Foom is about rate of power increase, not initial power level. Copy/paste isn’t everything, but still a pretty good superpower.
It’s not at all obvious to me that the increased risk of stagnation death outweighs the reduced risk of foom death.
You can’t copy paste hardware; and no, an upload won’t be able to run on a botnet.
Not to mention the bizarre assumption that an uploading patient will turn into a comic book villain whose sole purpose is to conquer the world.
Source?
Upvoted for this.
You know, I don’t think I’ve ever seen someone argue that. Does anyone have any links?
I’ve written a more detailed explanation of why recursive self-improvement is a figment of our imaginations: http://code.google.com/p/ayane/wiki/RecursiveSelfImprovement
Your third point is valid, but your first is basically wrong; our environments occupy a small and extremely regular subset of the possibility space, so that success on a certain few tasks seems to correlate extremely well with predicted success across plausible future domains. Measuring success on these tasks is something AIs can easily do; EURISKO accomplished it in fits and starts. More generally, intelligence isn’t magical: if there’s any way we can tell whether a change in an AGI represents a bug or an improvement, then there’s an algorithm that an AI can run to do the same.
As for the second problem, one idea that may not have occurred to you is that an AI could write a future version of itself, bug-check and test out various subsystems and perhaps even the entire thing on a virtual machine first, and then shut itself down and start up the successor. If there’s a way for Lenat to see that EURISKO isn’t working properly and then fix it, then there’s a way for AI (version N) to see that AI (version N+1) isn’t working properly and fix it before making the change-over.
In those posts you are arguing something different from what I was talking about. Sure chimps will never make better technology than humans, but sometimes making more advanced clever technology is not what you want to do and be positively detrimental to your chances of shaping the world to a desirable state. The arms race for nuclear weapons for example or bio-weapons research.
If humans manage to invent a virus that wipes us out, would you still call that intelligent? If so it is not that sort of intelligence we need to create… we need to create things that win in the end, not have short term wins and then destroy itself.
Super-plagues and other doomsday tools are possible with current technology. Effective countermeasures are not. Ergo, we need more intelligence, ASAP.
“More generally, intelligence isn’t magical: if there’s any way we can tell whether a change in an AGI represents a bug or an improvement, then there’s an algorithm that an AI can run to do the same.”
Except that we don’t—can’t—do it by pure armchair thought, which is what the recursive self-improvement proposal amounts to.
The approach of testing a new version in a sandbox had occurred to me, and I agree it is a very promising one for many things—but recursive self-improvement isn’t among them! Consider, what’s the primary capability for which version N+1 is being tested? Why, the ability to create version N+2… which involves testing N+2… which involves creating N+3… etc.
Again, there’s enough correlation between ability to perform certain tasks that you don’t need an infinite recursion. To test AIv(N+1)‘s ability to program to exact specification, instead of having it program AIv(N+2) have it instead program some other things that AIvN finds difficult (but whose solutions are within AIvN’s power to verify). That we will be applying AIv(N+1)’s precision programming to itself doesn’t mean we can’t test it on non-recursive data first.
ETA: Of course, since we want the end result to be a superintelligence, AIvN might also ask AIv(N+1) for verifiable insight into an array of puzzling questions, some of which AIvN can’t figure out but suspects are tractable with increased intelligence.
If you observed something to work 15 times, how do you know that it’ll work 16th time? You obtain a model of increasing precision with each test, that lets you predict what happens next, on a test you haven’t performed yet. The same way, you can try to predict what happens on the first try, before any observations took place.
Another point is that testing can be a part of the final product: instead of building a working gizmo, you build a generic self-testing adaptive gizmo that finds the right parameters itself, and that is pre-designed to do that in the most optimal way.
Where is the evidence that EURISKO ever accomplished anything? No one but the author has seen the source code.
On the subject of self-improving AI, you say:
Keep the bolded context in mind.
There are large classes of problem for which this is not the case. For example, “make accurate predictions about future sensory data based on past sensory data” relates program to world, but optimizing for this task is a property of the time, memory, and accuracy trade-offs involved. At the very least, your objection fails to apply “even in principle”.
This depends on the definition of “qualitatively improve”. It seems Eurisko improved itself in important ways that Lenat couldn’t have done by hand, so I think this too fails the “even in principle” test.
This seems the most reasonable objection of the three. Interestingly enough, Eliezer claims it’s this very difference that makes self-improving AI unique.
I think this also fails to apply universally, for somewhat more subtle reasons involving the nature of software. A self-improving AI possessing precise specifications of its components has no need for a constant lower layer of architecture (except if you mean the hardware, which is a very different question). The Curry-Howard isomorphism offers a proof-of-concept here: An AI composed of precise logical propositions can arbitrarily rewrite its proofs/programs with any other program of an equivalent type. If you can offer a good argument for why this is impossible in principle, I’d be interested in hearing it.
“It seems Eurisko improved itself in important ways that Lenat couldn’t have done by hand”
As far as I can see, the only self improvements it came up with were fairly trivial ones that Lenat could easily have done by hand. Where it came up with important improvements that Lenat wouldn’t have thought of by himself was in the Traveler game—a simple formal puzzle, fully captured by a set of rules that were coded into the program, making the result a success in machine learning but not self-improvement.
“The Curry-Howard isomorphism offers a proof-of-concept here”
Indeed this approach shows promise, and is the area I’m currently working on. For example if you can formally specify an indexing algorithm, you are free to look for optimized versions provided you can formally prove they meet the specification. If we can make practical tools based on this idea, we could make software engineering significantly more productive.
But this can only be part of the story. Consider the requirement that a program have an intuitive user interface. We have nothing remotely approaching the ability to formally specify this, nor could an AI ever come up with such by pure introspection because it depends on entities that are not part of the AI. Nor, obviously, is a formal specification of human psychology the kind of thing that would ever be approached accidentally by experimentation with Eurisko-style programs as was the original topic. And if the science of some future century—that would need to be to today’s science as the latter is to witchcraft—ever does manage to accomplish such, why then, that would be the key to enabling the development of provably Friendly AI.
...And applied these improvements to the subsequent modified set of rules. “That was machine learning, not self-improvement” sounds like a fully general counter-argument, especially considering your skepticism toward the very idea of self-improvement. Perhaps you can clarify the distinction?
An AI is allowed to learn from its environment, no one’s claiming it will simply meditate on the nature of being and then take over the universe. That said, this example has nothing to do with UFAI. A paperclip maximizer has no need for an intuitive user interface.
Indeed! Sadly, such a specification is not required to interact with and modify one’s environment. Humans were killing chimpanzees with stone tools long before they even possessed the concept of “psychology”.
“Perhaps you can clarify the distinction?”
I’ll call it self improvement when a substantial, nontrivial body of code is automatically developed, that is applicable to domains other than gameplaying, as opposed to playing a slightly different version of the same game. (Note that it was substantially the same strategy that won the second time as the first time, despite the rules changes.)
“A paperclip maximizer has no need for an intuitive user interface.”
True if you’re talking about something like Galactus that begins the story already possessing the ability to eat planets. However, UFAI believers often talk about paperclip maximizers being able to get past firewalls etc. by verbal persuasion of human operators. That certainly isn’t going to happen without a comprehensive theory of human psychology.
rhollerith:
“Strongly-recursive self-improvement” is a figment of the imagination; among the logical errors involved is confusion between properties of a program and properties of the world.
As for the rest: do you believe humanity can survive permanently as we are now, confined to this planet? If you do, then I will point you to the geological evidence to the contrary. If not, then it follows that without more advanced technology, we are dead. Neither I nor anybody else can know what will be the proximate cause of death for the last individual, or in what century, but certain extinction is certain extinction nonetheless.
Let us briefly review the discussion up to now since many readers use the the comments page which does not provide much context. rwallace has been arguing that AI researchers are too concerned (or will become too concerned) about the existential risk from reimplementing EURISKO and things like that.
You have mentioned two or three times, rwallace, that without more advanced technology, humans will eventually go extinct. (I quote one of those 2 or 3 mentions below.) You mention that to create and to manage that future advanced technology, civilization will need better tools to manage complexity. Well, I see one possible objection to your argument right there, in that better science and better technology might well decrease the complexity of the cultural information humans are required to keep on top of. Consider that once Newton gave our civilization a correct theory of dynamics, almost all of the books written before Newton on dynamics could safely be thrown away (the exceptions being books by Descartes and Galileo that help people understand Newton and put Newton in historical context) which of course constitutes a net reduction in the complexity of the cultural information that our civilization has to keep on top of. (If it does not seem like a reduction, that is because the possession of Newtonian dynamical theory made our civilization more ambitious about what goals to try for.)
But please explain to me what your argument has to do with EURISKO and things like that: is it your position that the complexity of future human culture can be managed only with better AGI software?
And do you maintain that that software cannot be developed fast enough by AGI researchers such as Eliezer who are being very careful about existential risks?
In general, the things you argue are dangerous are slow dangers. You yourself refer to “geological evidence” which suggests that they are dangerous on geological timescales.
In contrast, research into certain areas of AI seems to me genuinely fast dangers: things with a high probability of wiping out our civilization in the next 30, 50 or 100 years. It seems unwise to increase fast dangers to decrease slow dangers. But I suppose you disagree that AGI research if not done very carefully is a fast danger. (I’m still studying your arguments on that.)
Reduction in complexity is at least conceivable, I’ll grant. For example if someone invented a zero point energy generator with the cost and form factor of an AA battery, much of the complexity associated with the energy industry could disappear.
But this seems highly unlikely. All the current evidence suggests the contrary: the breakthroughs that will be necessary and possible in the coming century will be precisely those of complex systems (in both senses of the term).
Human level AGI in the near future is indeed neither necessary nor possible. But there is a vast gap between that and what we have today, and we will, yes, need to fill some of that gap. Perhaps a key breakthrough would have come from a young researcher who would have re-implemented Eurisko and from the experiment acquired a critical jump in understanding—and who has now quietly left, thinking Eurisko might blow up the world, to reconsider that job offer from Electronic Arts.
I do disagree that AGI research is a fast danger. I will grant you that there is a sense in which the dangers I am worried about are slow ones - barring unlikely events like a large asteroid impact (which is likely only over longer time scales), I’m confident humanity will still exist 100 years from now.
But our window of opportunity may not. Consider that civilizations are mortal, for reasons unrelated to this conversation. Consider that environments conducive to scientific progress are even considerably rarer and more transient than civilization itself. Consider also that the environment in which our civilization arose is gone, and is not coming back. (For the simplest example, while fossil fuels still exist, the easily accessible deposits thereof, so important for bootstrapping, are largely gone.) I think it quite possible that the 21st-century may be the last hard step in the Great Filter, that by the year 2100 the ultimate fate of humanity may in fact have been decided, even if nobody on that date yet knows it. I cannot of course be certain of this, but I think it likely enough that we cannot afford to risk wasting this window of opportunity.
One problem with this argument is how conjunctive it is: “(A) Progress crucially depends on breakthroughs in complexity management and (B) strong recursive self-improvement is impossible and (C) near-future human level AGI is neither dangerous nor possible but (D) someone working on it is crucial for said complexity management breakthroughs and (E) they’re dissuaded by friendliness concerns and (F) our scientific window of opportunity is small.”
My back-of-the-envelope, generous probabilities:
A. 0.5, this is a pretty strong requirement.
B. 0.9, for simplicity, giving your speculation the benefit of the doubt.
C. 0.9, same.
D. 0.1, a genuine problem of this magnitude is going to attract a lot of diverse talent.
E. 0.01, this is the most demanding element of the scenario, that the UFAI meme itself will crucially disrupt progress.
F. 0.05, this would represent a large break from our current form of steady scientific progress, and I haven’t yet seen much evidence that it’s terribly likely.
That product comes out to roughly 1:50,000. I’m guessing you think the actual figure is higher, and expect you’ll contest those specific numbers, but would you agree that I’ve fairly characterized the structure of your objection to FAI?
Neither is a belief in flying saucers, taken seriously, for most values of “harmless”.
To be clear, you are saying that you consider spontaneous FAI overwhelmingly likely, and thus consider any failure to promote FOOMing both a humanitarian catastrophe and a gateway to other existential risks? Because that seems like a claim worth proving.
You just made me want to participate even more!
...aaaand that’s why I don’t go around discussing the danger paths until someone (who I can realistically influence) actually starts to advocate going down them. Plenty of idiots to take it as an instruction manual. So I discuss the safe path but make no particular advance effort to label the dangerous ones.
Eliezer,
I am rather surprised that you accept all of the claimed achievements of Eurisko and even regard it as “dangerous”, despite the fact that no one save the author has ever seen even a fragment of its source code. I firmly believe that we are dealing with a “mechanical Turk.”
I am also curious why you believe that meaningful research on Friendly AI is at all possible without prior exposure to a working AGI. To me it seems a bit like trying to invent the ground fault interrupter before having discovered electricity.
Aside from that: If I had been following your writings more carefully, I might already have learned the answer to this, but: just why do you prioritize formalizing Friendly AI over achieving AI in the fist place? You seem to side with humanity over a hypothetical Paperclip Optimizer. Why is that? It seems to me that unaugmented human intelligence is itself an “unfriendly (non-A)I”, quite efficient at laying waste to whatever it touches.
There is every reason to believe that if an AGI does not appear before the demise of cheap petroleum, our species is doomed to “go out with a whimper.” I for one prefer the “bang” as a matter of principle.
I would gladly accept taking a chance at conversion to paperclips (or some similarly perverse fate at the hands of an unfriendly AGI) when the alternative appears to be the artificial squelching of the human urge to discover and invent, with the inevitable harvest of stagnation and eventually oblivion.
I accept Paperclip Optimization (and other AGI failure modes) as an honorable death, far superior to being eaten away by old age or being killed by fellow humans in a war over dwindling resources. I want to live in interesting times. Bring on the AGI. It seems to me that if any intelligence, regardless of its origin, is capable of wrenching the universe out of our control, it deserves it.
Why is the continued hegemony of Neolithic flesh-bags so precious to you?
This was addressed in “Value is Fragile.”
I don’t think you understand the paperclip maximizer scenario. An UnFriendly AI is not necessarily conscious; it’s just this device that tiles the light cone with paperclips. Arguably it helps to say “really powerful optimization process” rather than “intelligence.” Consider that we would not say that a thermostat deserves to control the temperature of a room, even if we happen to be locked in the room and are to be roasted to death.
I was going to reply, but it appears that someone has eloquently written the reply for me.
I’d like to take my chances of being cooked vs. a world without furnaces, thermostats or even rooms—something I believe we’re headed for by default, in the very near future.
This reminds me of the response I got when I criticized an acquaintance for excessive, reckless speeding: “Life is all about taking risks.”
The only difference was that he was mostly risking his own life, whereas asciilifeform is risking mine, yours and everyone else’s too.
ASCII—the onus is on you to give compelling arguments that the risks you are taking are worth it.
Actually, I fully intended the implication that he was risking more than his own life. Self-inflicted risks don’t concern me.
Now you’ve got me wondering what the casualty distribution for speeding-induced accidents looks like.
Well if ASCII has his way, there may be one data point at casualty level 6.6 billion …
I will accept that “AGI-now” proponents should carry the blame for a hypothetical Paperclip apocalypse when Friendliness proponents accept similar blame for an Earth-bound humanity flattened by a rogue asteroid (or leveled by any of the various threats a superintelligence—or, say, the output of a purely human AI research community unburdened by Friendliness worries—might be able to counter. I previously gave Orlov’s petrocollapse as yet another example.)
Status quo bias, anyone?
I presently believe, not without justification, that we are headed for extinction-level disaster as things are; and that not populating the planet with the highest achievable intelligence is in itself an immediate existential risk. In fact, our current existence may well be looked back on as an unthinkably horrifying disaster by a superintelligent race (I’m thinking of Yudkowsky’s Super-Happies.)
Since your justification is omitted here, I’ll go ahead and suspect it’s at least as improbable as this one. The question isn’t simply “do we need better technology to mitigate existential risk”, it’s “are the odds that technological suppression due to friendliness concerns wipes us out greater than the corresponding AGI risk”.
If you assume friendliness is not a problem, AI is obviously a beneficial development. Is that really the major concern here? All this talk of the benefits of scientific and technological progress seems wasted. Take friendliness out of the picture, I doubt many here would disagree with the general point that progress mitigates long-term risk.
So please, be more specific. The argument “lack of progress contributes to existential risk” contains no new information. Either tell us why this risk is far greater than we suspect, or why AGI is less risky than we suspect.