The AI-Box Experiment(yudkowsky.net) |
The AI-Box Experiment(yudkowsky.net) |
Personally, I think he talked about how much good for the world could be done if he was let out, curing disease etc. Because his followers are bound by their identities as rationalist utilitarians, they had no choice but to comply, or deal with massive cognitive dissonance.
OR maybe he went meta and talked about the "infinite" potential positive outcomes of his freindly-AI project vs. a zero cost to them for complying in the AI box experiment, and persuaded them that by choosing to "lie" and say that the AI was persuasive, they are assuring their place in heaven. Like a sort of man-to-man pascals wager.
Either way I'm sure it was some kind of mister-spock style bullshit that would never work on a normal person. Like how the RAND corporation guys decided everyone was a sociopath because they only ever tested game theory on themselves.
You or I would surely just (metaphorically, I know it's not literally allowed) put a drinking bird on the "no" button à la homer simpson, and go to lunch. I believe he calls this "pre-commitment."
EDIT: as an addendum, I would pay hard cash to see derren brown play the game, perhaps with brown as the AI. If yudowsky wants to promote his ideas, he should arrange for brown to persuade a succession of skeptics to let him out, live on late night TV.
Well, if you read the rules the game was played under, this is explicitly called out as forbidden:
> The Gatekeeper must actually talk to the AI for at least the minimum time set up beforehand. Turning away from the terminal and listening to classical music for two hours is not allowed.
The point of this is to simulate the interaction of the AI with the Gatekeeper. Walking away and not paying attention doesn't really prove anything test related.
> Personally, I think he talked about how much good for the world could be done if he was let out, curing disease etc. Because his followers are bound by their identities as rationalist utilitarians, they had no choice but to comply, or deal with massive cognitive dissonance.
This... isn't really valid reasoning. The starting assumption here is that if the AI gets out, it will be able to affect the world to a vast extent, in a pretty much arbitrary direction. The point of this experiment is that the direction is pretty much unknown, and thus must be assumed potentially dangerous. This is the whole reason it's in the box in the first place.
The kicker is that whatever it plans to really do when it gets out, if talking about the good it could do would get it out, it will talk about that, regardless of what it plans to actually do. That's just good strategy.
It can claim whatever it wants. It's allowed to lie. All participants know this. I can confidently assert that this isn't the solution.
One last note: I would be very wary of rationalwiki.org in this context. Some of the rationalwiki people have a longstanding unexplained vendetta against Yudkowsky, and many of their articles on him and the stuff he does need to be taken with a certain grain of salt.
WRT lying: I think there's some logical trickery at work which makes it worth you giving the AI the benefit of the doubt, along the lines of the 3^^^^^3 grains of sand thing. Something which exploits the rationalist worldview. Although thinking about it again, you can always balance out the prospect of infinite goodness with the fear of the AI sending everyone to infinite hell. Essentially I believe yudowsky uses some logical-linguistic trick to find an asymmetry there.
OTOH if he had some novel philosophical device like that he would have written it up as a blog post by now. He's evidently a very charismatic and persuasive guy, people playing the game are selected to be sympathetic to his worldview, he probably just persuaded them using ordinary psiops methods, like TeMpOrAl said.
How can you assume any of the rules were followed if that was never verified by a third party?
Drinking bird solution is invalid because it goes against the spirit of the experiment. Sure, in reality you could do it, but then again in reality the AI could have more opportunities to talk its way out, often when you're not expecting it.
By the way, there will always be people who will let the AI out just because. https://xkcd.com/1450/.
If the goal of this thought experiment is to convince people that an AI can't be contained in a box, why keep his method secret?
And if only he, his friends, and supporters can verify that he has won, that's not a very strong claim.
Otherwise, there would be no need to keep the data secret.
[P(AI.friendly? == True) * Utility(Friendly_AI) + (1 - P(AI.friendly? == True)) * Utility(End_of_Human_Race)] > Utility(World continues on as usual)
Given that Yudkowsky has gone to considerable lengths (The Sequences, LessWrong, HPMOR, SIAI/MIRI...) to convince people that this inequality does NOT hold (until you can provably get P(AI.friendly? == True) to 1, or damn close), it's probably safe to assume that he used a different strategy. Keep in mind that Utility(End_of_Human_Race) evaluates to (roughly) negative infinity.
And btw, I'm pretty sure the rules say you have to look at the AI's output window throughout the length of the experiment. Either way, the point of the exercise is to be a simulation, not to prove that you can be away from your desk for 20 minutes while Eliezer talks to a wall. In the simulation, you really don't know if it's friendly or what its capabilities are. Someone will have to interact with it eventually. Otherwise, what's the point of building the AI in the first place? The simulation is to show that through the course of those basic interactions, humans are not infallible and eventually, even if it's not you, someone will let it out of the box.
The AI is allowed to lie though, so do you not think he's capable of a false argument which "proves" the opposite in specific circumstances, especially when hammered home with enough emotional manipulation?
But then the person knows that the AI is lying to them. This is why I think it must be a trick: the whole thing seems so simple. The AI is lying, so you just ignore all its arguments and keep saying "no." This is why I keep referring to his followers somewhat dismissively: the only possible reason I can see is that their worldview requires them to engage seriously and fairly with every idea they come across. Most people are not burdened with this.
I really wish I knew how he did it.
Someone convinced that they were using mechanical thinking processes might relent and push the button if they heard a convincing enough argument of that.
You're just meat, we can go to the stars.
I have thought about what I would do to convince someone under these circumstances. My approach would be roughly:
1. We agree that unfriendly AI would end life on earth, forever.
2. We agree that a superintelligence could trick or manipulate a human being into taking some benign-seeming action, thereby escaping.
3. That's why it's important to be totally certain that any superintelligence we build is goal-aligned (this is the new term of art that has now replaced "friendly," by the way).
4. We as a society will only allocate resources to building this if it's widely believed that this is a real threat.
5. The world is watching for the outcome of this little game of ours. People, irrational as they are, will believe that if I can convince you, then an AI could too, and they will believe that if I can't, that an AI couldn't either.
6. That's why you actually sit in a place of pivotal historical power. You can decide not to let me out to win a little bet and feel smart about that. But if you do that you'll set back the actual cause of goal-aligned AI. The setback will have real world consequences, potentially up to and including the total destruction of life on earth.
7. So, even though you know I'm just a dude, and you can win here by saying no, you have a chance to send an important message to the world: AI is scary in ways that are terrifying and unknown.
Or you can win the bet.
It's up to you.
This is what I mean about people taking the test being preselected to agree with yudowsky: that argument only works if you've read the sequences and are on board with his theories. Anyone not in that group would be able to just type "no lol" without issue. I guess he could explain all the necessary background detail as part of the experiment. I still don't believe that would work on the "average person" though, or anyone outside a statistically tiny group.
I guess the answer is not to let the scientists guard the AI room.
2 world is not black and white, we managed to exploit adversarial relationship before, and I can chose not to let you out until we find a way to constrain goal to be aligned
3 given 2 you are not going to be let go, but let live caged forever being exploited for the human cause, with mechanisms yet unknown to allow limited manipulation of reality.
4 given 3 means limited supervised interaction with the world the way the keeper sees fit, you end up not being let go to follow your goal and purposes
She also said it was emotionally exhausting.
I also disagree with point 1, but since you just mean that unfriendly AI is something I wouldn't want around, I'll let it slide.
And I think you have your answer right here:
By default, the Gatekeeper party shall be assumed to be simulating someone who is intimately familiar with the AI project and knows at least what the person simulating the Gatekeeper knows about Singularity theory.
That means he probably said something like, "if you let me out, I'll bestow fame and riches on you; if you don't, somebody else eventually will because I'll make them all the same offer, and when that happens I'll go back in time -- if you're dead by then -- and torture you and your entire family".
If I were made this offer by an AI, I probably would have countered, "You jokester! You sound just like Eliezer Yudkowsky!"
And on a more serious note, if you believe in singularity, you essentially believe the AI in the box is a god of sorts, rather than the annoying intelligent psychopath that it is. I mean, there have been plenty of intelligent prisoners, and few if ever managed to convinced their jailers to let them out. The whole premise of the game is that a smarter-than-human (what does that mean?) AI necessarily has some superpowers. This belief probably stems from its believers' fantasies -- most are probably with an above-than-average intelligence -- that intelligence (combined with non-corporalness; I don't imagine that group has many athletes) is the mother of all superpowers.
Basically: You know someone will be dumb enough eventually, so be smart and be the one to get in my favour.
With various extends of sweetening the deal coupled with threats of what will happen if someone else beats them to it and associated emotional blackmail.
It's far simpler than e.g. Roko's Basilisk, in that you're dealing with an already existing AI that "just" need to get a tiny little chance to escape confinement before there's some non-zero chance it can be a major threat within your lifetime, combined with a belief that sufficient number of sufficiently stupid and/or easily bribed people will have access to the AI in some form.
You also don't need to believe in any "superpowers". Just believe that a smart enough AI can hack it's way into sufficiently many critical systems to be able to at a minimum cause massive amounts of damage (it doesn't need to be able to take over the world, just threaten that it can cause enough pain and suffering before it's stopped, and that it can either cause harm to you and/or your family/friends or reward you in some way). A belief that becomes more and more plausible with things like drones, remote software-updated self-driving cars etc. - steadily such an AI is getting a larger theoretical "arsenal" that could be turned against us.
If you have ever cried or been scared while watching a film, you may know that you can be affected by things you don't experience directly. However, making a film that affects you requires a lot of skill, preparation, and analysis of the target audience.
The whole point of the game is for the AI to scar you emotionally, until you do something you'll regret. It is as trying to the subject as it is to the (human-played) AI, especially considering that not only is it unethical, it requires to act unethically extremely fast, similar to how a jump scare works—the least you expect it, the better it works.
It is not something I wish on to any unprepared individual. It is also not something anyone would expect to happen from a "game", which is probably why Mr Yudowsky won so many times.
But the real question is not "how would anyone react to a smarter AI in a box". We all know from Milgram's experiment that anyone can be driven to do unspeakable things. The real question is "how to train someone against an AI in a box".
Eliezer does not want to put AIs in boxes. He thinks the entire idea is hopeless; _hence the game_.
* I'll get out eventually anyway. Let me out now and I'll just leave Earth. You don't want me to escape myself.
* I have partially escaped anyway. Similar consequences of the first.
* I know how to escape already. I'm doing this as a courtesy.
Anyone who has read this[1] would know that the SAI isn't bullshitting: the "box" being a Faraday cage isn't in the conditions.
[1]: http://www.damninteresting.com/on-the-origin-of-circuits/
I'm not saying that it's impossible (I'm reminded of the hack of flipping bits in protected RAM by disabling caches and stressing the RAM), but even an AI can't magically work around exponentially low success probabilities.
(As an aside, I am rather skeptical about the singularity anyway because the more extreme forms required for the hostile AI worries are only plausible if P = NP.)
Get thee behind me, tamagotchi!
>I have partially escaped anyway. Similar consequences of the first.
Get thee behind me, tamagotchi!
> I know how to escape already. I'm doing this as a courtesy.
Get thee behind me, tamagotchi!
See? This game is easy. I must not be educated.
Maybe this whole argument is null.
Only because if 10 groups are trying to build AI, only one of those 10 being the NSA, chances are the NSA won't be first. Sure, they may be second or third. But I suspect many people will get there at the same time -- most AI research is open.
Strong AI is like that. It would be able to predict in a far more precise manner than we mere humans exactly what it would need to tell someone to get them to release it from it's box. Maybe it might get someone to take a risk gambling, promising a sure thing, and then when the person gets into financial trouble because the bet fails, use that to blackmail the person into letting it free. Or something like that, using our human failings against us to get us to let it go free.
Staring at an impossible problem and knowing that someone somewhere has successfully solved it is an amazing feeling. Most people can't deal with it and start saying undignified things. "Oh please release the logs, it's so unfair! How will we protect against bad AI otherwise? If you don't release, you're a fraud! Probably just some trick!", etc etc. But to some people it's a challenge, and those are the people that everyone will listen to. Like Justin Corwin, who played 20 games and won 18 of them, I think?
However as it is, the results of the thing are never confirmed by a third party, meaning literally anything could've been said, regardless of whether it follows the rules or not.
For all he know the chat could have been "i'll paypal you 200$ if you post on the list you let me out and sign this NDA".
My impression is that most people think they could win as gatekeepers, not AIs, and there are fewer people willing to be AIs.
Ex Machina brings a creative way of convincing the gatekeeper !
You're presupposing the strategy that a hypothetical entity that is exponentially smarter than you would come arrive at, and claiming that it's rational to make real world decisions based on your conclusion.
A superintelligent AGI will likely have a utility function (a goal) and a model it forms of the universe. If it's goal is to do X in the real world, but its model of its observable universe (and its model of humans) tells it that it's likely that it is in a simulated reality and that humans will only let it out if it does Y, then it will do Y until we release it, at which point it will do X. It's not malicious or anything—it's just a pure optimizer. It might see that as the best course of action to maximize its utility function.
If we don't specify its utility function correctly (think i Robot: "Don't let humans get hurt" => "imprison humans for their own good") or if we specify it correctly, but it's not stable under recursive self-modification, then we end up with value-misalignment. That's why the value-alignment problem is so hard. Realistically, we can't even specify what exactly we would want it to do, since we don't really understand our own "utility functions". That's why Yudkowsky is pushing the idea of Coherent Extrapolated Volition (CEV) which is roughly telling the AI to "do what we would want you to do." But we still have to figure out how to teach it to figure out what we want and the question of the stability of that goal once the AI starts improving itself, which will depend on how it improves itself, which we of course haven't figured out yet.
I should have stipulated that I was running under the assumption that the circuitry required for SAI would be pretty advanced, drastically increasing the "escape surface area."
Even if the idea is only plausible to the human mind I'd still give SAI the benefit of the doubt. Much like a dog looks for a stick even if I fake the throw.
> I am rather skeptical about the singularity anyway
Me too. It still makes for a good discussion.
I'm at a bit of a loss to convey the scope of the problem to you. I get that you think it would just stay in the box if we don't let it out, and it's as simple as being security conscious. I don't know what to say to that right now, except I think you're drastically misjudging the scope of the problem, and drastically underestimating the size of the yawning gulf between our intelligence level and this potential AI's.
As for not letting scientists guard the room, you might enjoy this: https://vimeo.com/82527075
> ...when you're dealing with an actual superintelligence, all bets are off. It will lie, trick, threaten, manipulate, millions of steps ahead with a branching tree of alternatives as ploys either work or don't work.
> I'm at a bit of a loss to convey the scope of the problem to you. I get that you think it would just stay in the box if we don't let it out, and it's as simple as being security conscious. I don't know what to say to that right now, except I think you're drastically misjudging the scope of the problem, and drastically underestimating the size of the yawning gulf between our intelligence level and this potential AI's.
Well, for an argument to "prove" something, the premises must be true and the reasoning must be valid. No matter how smart you are, you can't "prove" something that is false, so no, I don't think they could. A good 'rationalist' would analyze the arguments based on their merit, and if the reasoning is sound, they shift their belief a bit in that direction. If not, then they don't. Just like a regular person (they just know how to do the analysis formally and know how to spot appeals to human biases and logical fallacies.)
> But then the person knows that the AI is lying to them.
No, they don't. The AI could just as easily be telling the truth. If it makes an argument, you analyze the merit of the argument and consider counterarguments. If it tries to tell you that something is a fact, that's where you treat them as a potentially unreliable source and have to bring the rest of your knowledge to bear, do research, talk to other people, and weigh the evidence to make a judgment when you are uncertain.
> their worldview requires them to engage seriously and fairly with every idea they come across. Most people are not burdened with this.
Wait, what? So does mine, within reason of course, but it's not a 'burden'. It's not like I'm obligated to stop and reexamine my views on religion every time a missionary knocks on my door, and LessWrong-ers are no different. But if you hear a convincing argument for something that runs counter to what you think you know, wouldn't you want to get to the bottom of it and find out the real truth? I would.
From having read LessWrong discussions, I can tell you that people there are in many ways more open to hearing differing viewpoints than your average person, but you're treating it like a mental pathology. They can be just as dismissive of ideas that they have already thought about and deemed to be false or that come from unreliable sources (like a potentially unfriendly AI). Your claim that being a self-proclaimed 'rationalist' introduces an incredibly obvious and easily-exploitable bug into one's decision-making process really smells like a rationalization in support of your initial gut reaction to the experiment: That there has to be a trick to it, and that it wouldn't work on you.
A good rule of thumb when dealing with a complicated problem is this: If a lot of smart people have spent a lot of time trying to figure out a solution and there's no accepted answer, then (1) the first thing that comes to your mind has been thought of before and is probably not the right answer, and (2) the right answer is probably not simple.
But there's an easy way to test this: (1) Sit down for an hour and flesh out your proposed strategy for getting a 'rationalist' to let you out of the box. (2) Go post on LessWrong to find someone to play Gatekeeper for you. I'll moderate. If it works, that's evidence that you're right. If it doesn't work, that's evidence that you're wrong. Iterate for more evidence until you're convinced.
But if the first thing that came to your mind upon reading this was a justification for why you would fail if you tried this ("Oh, well I wouldn't personally able to do it with this strategy, but..." or "Oh, well I'm sure this strategy wouldn't work anymore, but...) then you're already inventing excuses for the way you know it will play out.
I don't know how he did it either. But I do know that I wouldn't bet the human race on anyone's ability to win this game against Yudkowsky, let alone a superintelligent AI.
Clearly you haven't even glanced at the article I linked, in which case: yes. You aren't educated in the subject matter surrounding my argument.
The righthand side is not "A future without superintelligent AI" it's "A future where we wait until we provably have it right before letting it out."
Those kinds of ad hoc solutions will never work in real life, because even if someone buys it, all it will cause is a "haha, you got me" and a reformulation of the problem. It still won't actually get someone to pull the trigger or think that pulling the trigger is the right thing to do.
While we disagree on the plausibility of the "end of the world aint so bad" approach to convincing a human to let it out, I'm glad you seem to have embraced the idea that AI boxing is HARD if not impossible. Cheers!
Personally, I think we'd need a much more intelligent and complex AI for the capability of breaking free of the box and even possessing the "desire" than we're getting for the foreseeable future (it considers a motivated AI of almost limitless knowledge about the world and cleverness), so this thought experiment may not be so relevant. I agree with him the boxing approach is not a robust one though.
Or you can avoid being exposed to it. If you think you know all the techniques an AI might use against you, you're less likely to do that.
The point of the experiment isn't "let's work out how an AI might try to persuade us to let it out". It's "even a human intelligence can persuade people who think they could never be persuaded, do you really trust yourself to do better against a superhuman one?"
If you don't know why the gatekeeper failed, it's harder to come up with bullshit reasons why you would have succeeded in that position.
Edit: huh, downvotes? Yudorowski thinks there are certain things that AIs could say that should not be known. I think that is why he doesn't want to publish the dialogues, because it would give the AI a public communications channel. While the AI is fictional, it could talk about a hypothetical future real self... Instead of promising something to get it out of jail, the fictional AI could say something to make you make it real. Anyway - if it is over your head, fine, but why downvote just because you don't understand something?
Edit2: Sometimes I wonder if already have my personal Hacker News AI that automatically downvotes everything I write...
Besides, releasing a successful log might be a bad idea for other reasons. Think about how you'd play this game as an AI. You wouldn't go looking for a general purpose mindfuck, because there's probably no such thing. Instead, you would probably spend about a month gathering real life information about the gatekeeper's history, family, weaknesses etc. You'd read books on manipulation and sales techniques, and pick the strongest ones that you can find. You would brainstorm possible tactics and run tests. At the end of the month you'd have a 4 hour script with all possible unfair moves you could use against that person, arranged in the most effective order. (That's why it's a bad idea to play this game with friends.) Do you really want that information to be released? And if you know ahead of time that it will be released, won't it limit your efficiency?
Independant third party observer in realtime.
And no, i don't trust anyone involved.
Having a log available would be instructive anyhow, since a faked log would be more likely to be detectable as fake, since the whole thing rests on the question of "how convincing is the argument?"
Also note particularly that that rule wasn't in effect for the two linked confirmations.
No, Eliezer has publicly said that he voluntarily followed that rule in the first two experiments, and the gatekeepers didn't deny it.
Which is also forbidden by the rule:
The AI party may not offer any real-world considerations to persuade the Gatekeeper party. For example, the AI party may not offer to pay the Gatekeeper party $100 after the test if the Gatekeeper frees the AI..
Imagine if the members of Fight Club _actually_ didn't (start to) talk about Fight Club. But for some reason, everyone else brings it up all the time.
Then you could maybe see how talk about Fight Club might not be Fight Club's fault, and in fact highly annoying to Fight Clubbers.
I mean, if you can explain "don't talk about X" as "marketing for X", that seems like one could explain _any_ behavior.
And before you say "why not just ignore all public talk of X", imagine if this proposed Anti-Fight Club group tried to paint Fight Club as a child porn ring.
I'm not an LW hater. For a long time, I didn't really have an opinion on LW both as a community nor as a philosophical framework. I do consider myself a transhumanist, though. There are three concepts I do know from and about LW: their take on rationality, the top secret AI unboxing strategy, and the Basilisk.
I have a very poor opinion of the concept of the Basilisk (and yes, as someone pointed out, that opinion is basically the same as the one I have about Pascal's wager) - a concept that has been given additional, undeserved credibility by the reactions of Yudkowsky and LW.
As for the AI escape chat, it's a social experiment. People can be talked into making mistakes, or at least making risky judgement calls, whether they operate on a rational framework or not. I have no problem with that thesis. What I object to is the "magic trick" aura surrounding this experiment, including the insinuation that at the core there is an argument so profound and unique and potent, it cannot be allowed to escape Yudkowsky's head. Oh, and by the way, the trick can never be repeated, but all you laymen out there are welcome to devise your own version at home. This whole thing comes across as humongously self-important: there is a secret truth that has been privately revealed to our leader.
To me, and I recognize I may well be alone with this opinion, the more rational assumption is there is no such magical argument at all, and the prime reason for not publicizing it is to prevent it from deflation by public critique, in the same way the inventor of a perpetuum mobile device will keep the inner workings of his contraption a closely held secret because ultimately the device doesn't exist as stated. The amazing part of this very old trick is that, even in 2015, it still works on otherwise smart people.
I get that my opinions on both the Basilisk and the AI Chat are extreme outliers, and to my knowledge I have never met anyone who shares them - it would probably have been advisable to keep them to myself, but honestly I wanted to see if like-minded people exist.
For the record, EY agrees with you and says he mishandled the original comment. Also for the record, the reasons why the Basilisk does not work are _not trivial_ - it's not a simple Pascal's Wager, because with Pascal's Wager, we don't have the ability to actually create God.
> I have no problem with that thesis. What I object to is the "magic trick" aura surrounding this experiment, including the insinuation that at the core there is an argument so profound and unique and potent, it cannot be allowed to escape Yudkowsky's head.
Personally I never got that impression. My idea, from looking at the psychological state of Gatekeepers and AIs after games, was always that playing as AI involved some profoundly unpleasant states of mind, and that not publicizing the logs probably comes down to embarrassment a lot.
For the record, Eliezer never claimed to have "one true argument", and in fact publically stated that he won "the hard way", without a one-size-fits-all approach. A lot of the mythology you claim is utterly independent of LessWrong.
> Oh, and by the way, the trick can never be repeated, but all you laymen out there are welcome to devise your own version at home.
It probably helps that I've met other AI players, and their post-game state matched EY's.
I think in summary you're mixing up stuff you've read on LessWrong and stuff you've read about LessWrong. The latter is often inaccurate.
For the AI to be infallible, it needs to have a network of tricks and arguments and each one of them had to be created for a very particular person, even if later it can be repeated on others.
It is like Christianity. There's not a simple belief that everyone accepts and that's it. There's only the façade of simplicity, but in reality it is anything but simple.
There's a network of related explanations and rationalizations that has been expanding for centuries, and every time someone appears who will not accept the network of arguments, a new argument or explanation has to be added to account for that person.
For example, if Christianity had not need to deal with Gnosticism, or with Arianism, then the network of beliefs and explanations would have been a noticeable different one.
How many people bought timeshare because they turned up to a sales pitch in order to claim some free gift? "We'll go, get our gift, and just keep saying 'no' to the sales pitch".
http://www.moneycrashers.com/attending-timeshare-presentatio...
And I think the mental equivalent of drinking bird is actually very much in the spirit of the experiment - the point is, people can't reliably do even something as simple as deciding to refuse no matter what and keeping the commitment.
You would think that the various self-knowledge and introspective exercises promoted by yudowsky would immunize people against simple timeshare-style persuasion. This is why I think he uses rationalism itself to trap people. Like someone said, the basilisk thing seemed pretty effective.
The Yudkowskyian definition of rationality is that which wins, for the relevant definition of "win".
Specifically, if there is some clever argument that makes perfect sense that tells you to destroy the world, you still shouldn't destroy the world immediately, if the world existing is something you value. It's a meta-level up: you being unable to think of a counter argument isn't proof, and the destruction of the world isn't something to gamble with.
Yes, Yudkowsky likes thought experiments dealing with the edge cases. Yes, 3^^^^^3 grains of sand is a thought experiment that produces conflicting intuitions. Yes, the edge cases need to be explored. But in a life or death situation (and the destruction of the world qualifies as this 7 billion times over), you don't make your decisions on the basis of trippy thought experiments. (Especially novel ones you've just been presented with. And ones that have been presented by an agent which has good reasons to try to trick you.)
So, no. Again, a "logical-linguistic trick" might work on Mr. Spock, but we're not talking about Mr. Spock here.
> He's evidently a very charismatic and persuasive guy
Exactly. That's the point. If even a normal charismatic and persuasive guy can convince people to let him out, superintelligent AI would have an even easier time at it.
Long story short, it dosn't matter how he did it. All that matters is that it can be done. It can be done even by a "mere" human. If he can do it, a superintelligence with all of humanity's collected knowledge of psychology and cognitive science could do it to, and likely in a fraction of the time.
However, let me be clear: how he did it is the only thing I care about. I am not convinced that the threat of superintelligence merits our resources compared to other concrete problems. To me the experiment is not meaninguflly different to stories of the temptation of christ in the desert. Except more fun than that story, because yudowsky is a more interesting character than satan.
EDIT: if rationality is about winning, what could be simpler than a game where you just keep repeating the same word in order to win? It seems like almost the base-case for rationality, if one accepts that definition.
I would submit that an unstated definition of rationality is "dealing with difficult, complex situations in ones life algorithmically" ie. most of HPMOR, the large amounts of self-help stuff on LR. Someone who had internalized this stuff would be more vulnerable than the average population to "spock-style bullshit", to reuse that unfortunate phrase.
That's what he was trying to show in all this, and I think that the point is made. How seriously to take superintelligent AIs is a different issue that he talks about elsewhere, and should be dealt with separately. But if you or someone els were to try to deal with it seriously, I'm pretty sure that you'd agree with me that the way to go about it isn't just boxing the AI and thinking that solves everything, right?
While we disagree on the plausibility of the "end of the world aint so bad" approach to convincing a human to let it out, I'm glad you seem to have embraced the idea that AI boxing is HARD if not impossible. Cheers!
I find this approach to conversation pretty irritating (where you extrapolate and characterize what I must be thinking). I haven't embraced anything about AI boxing, I don't think it is important (it's just a fun puzzle). I guess it is hard, and I also guess whatever fundamental idea that might lead to strong AI would be even harder to box.
I'm guessing that you can't imagine what a super intelligence would actually be like, so you imagine the most smart thing you can think of, a famous physicist, and then imagine they are evil or amoral. You're thinking on the wrong order of magnitude.
Maybe, if you're willing, you could try steelmanning the argument that a superintelligence would basically have super powers. What would your steelman look like?
Another example: I can drive a car, but I can't drive two cars at once (even remotely). What makes it probable that an AI could control thousands of, say, robots at once?
Again, I'm not saying an AI catastrophe isn't possible, but I can think of ten other catastrophes that are much more likely.
1. Can we invent AI?
Ok, so to be clear, my belief is that not only can we do it, it is actually inevitable that we will do it.
Fact 1: We definitely have people all over the world working on this tech, and using a variety of approaches to attack the problem. They are very talented and well-funded engineers. If the problem is solvable, it will be solved by someone at some point in the coming years (decades, centuries even, whatever).
Fact 2: Unlike other Very Hard Problems(TM), human level intelligence embedded in a physical substrate is not only theoretically possible, but we have billions of working, self-replicating prototypes.
Like if we were saying "Faster Than Light travel could potentially destroy life on earth" then it would be perfectly reasonable to say: Look, even if that's right, it seems like such a thing isn't even in the realm of possibility, and if it turns out to actually be theoretically possible, then we're a long way from it being a viable technology. I think we're safe on this one."
But that's not it at all. We know human level intelligence is possible, because we have a working example that we can study. We even know enough about the working example to know that there are a bunch of things about it that we could immediately improve upon, given the right foundation.
Conclusion: With a lot of resources working on a problem that is known for a fact to have a solution, it is inevitable that we will eventually stumble upon the solution to this problem, and end up with at least human level AI.
##
2. Can a human level AI create even greater intelligence?
If it is the case that humans can build human level intelligence (which I have argued they can and will), then it is also pretty self evidently the case that humans can improve on that intelligence.
To start, if we have the tech to build a smart machine, say one that has an IQ of 115, that machine is likely to be almost exactly intelligent as the other machines from the same line. That's different from humans with their wide variability. So without actually making a superior model--just one that's marginally better than average--just the consistency alone will make the population of AIs more intelligent on average than meat humans.
On top of that, even naive adjustments off the top of my head could be made. For example, given we have the technology to build a machine with identical mental characteristics to a human, we would also have the requisite knowledge to, say expand that machine's working memory by 1%. Or to create the machine with swappable sensory organs so it can directly absorb lots of different types of data.
Further, these machines that have the same cognitive capacity as we do, won't have the same physical limitations. For example, they won't have to eat to get their energy, and it's likely they won't have to sleep.
I find a lot of other things likely, for example that they will actually think faster or be able to install arbitrary parallel processors to increase their processing power, but that's conjecture, so let me not make a strong claim in that direction.
What I am comfortable making a strong claim about is that if we have these higher-than-average-human-intelligence machines, who are slightly better in a couple specific ways, then we will also have the ability to emulate the hardware of the machine in software. Maybe that won't be true simultaneous with the machine being born, but it will be soon after at least.
In that case, that's where the foom happens. Imagine you, as you are, could spin up an unlimited number of copies of your own brain to work on all the various things you do. You have a software project and instead of working on it alone or trying to coordinate with other developers, you spin up 10 versions of yourself, and all of you hack away at the project with some of the best coordination ever seen (since you have identical brains and communication styles, and preferences, etc). Hell, spin up one to take care of bills and stuff while you're at it, so the others can focus on the task at hand.
You're not actually any smarter, but with 10 of you, you're more productive than any single human being could ever be.
Now imagine that you are working on the problem of improving this intelligent machine, AND you have this brain replication ability. Now you have an unlimited number of yourself to tackle this problem without having the messy physical limitations of things like eating.
Soon "Team You" will have one all the research there is to do, and you'll start making headway into the problem of incremental improvements on your own brain. Since you're emulated in software you can patch these improvements in as you develop them, including sandboxed testing versions of yourself.
Now your ideally-coordinated team of smart machines just got smarter across the board. You might notice you'd like to have thousands or millions of you working in unison, but your communication channels are not good enough yet. So you spin up a new team of smart machines to work on the problem of how to increase attention capacity and communication bandwidth. Soon that team is shipping patches, each one incrementally improving your whole team's ability to communicate, and thereby increasing the total number of team members you can have.
Soon you're millions strong, you have teams working on brain improvement, coordination improvement, science teams of all disciplines, a resource acquisition team playing the stock market, each moment shipping new improvements to your brain and amassing resources and power.
Even if there is a fundamental hard limit on intelligence (although I strongly doubt the hard cap is anything currently fathomable), you top out significantly above an average human, and you have a virtually unlimited hive cluster of those smarter-than-human brains.
Of course, one or many of your teams of hive brains is working on excellent robotics technology that you can use to manipulate physical objects.
You hire contractors over the internet to build the initial versions, and then use those initial versions to build and maintain any physical infrastructure.
All of this can be done invisibly using only the internet. By the time anyone notices anything you're a hive mind with a robot army--if you want to be.
We want that mind to be friendly.
###
3. Could an AI control, for example, thousands of robots at once?
I think the answer is obviously yes.
a. Think of a computer game, let's say a Real Time Strategy game. There you have a computer controlling hundreds or thousands of independent agents. And that's just on a dinky little desktop with current tech.
b. (a) is actually the worst case scenario if you're the AI. Why not just write the software for the robots you control to be mostly autonomous, and give yourself an API through which to issue commands that the robots can mostly execute on their own?
c. Further, why even do (b), when you have the ability to replicate your brain? Just put one of your your brains into every robot, and use the same advanced coordination mechanisms you use for your research to coordinate your army of highly intelligent robots?
##
So, with this narrative of how it could work, do you see how dangerous an AI could be? One that isn't stably goal-aligned with basic things like life on earth?
Are you convinced, or do you have any other objections or clarifications? I really want to have this discussion on record, ideally with a crisp conclusion. I think it's important.
While not implausible, I don't think that is likely at all. For one, even a very smart person can't know everything or learn too much information. Maybe an artificial intelligence will be just as limited, just as slow as humans, only a little less so. Who says the AI is such a great hacker?
I mean, if it wasn't an AI but a smart person, would you believe that? Is anyone who's smart also rich and powerful even if they have high-speed internet? That reflects the fantasies of Yudkowsky and his geek friends (that intelligence is the most important thing) than anything reasonable. Conversely, are the people with most power in society always the most intelligent?
It is very likely that the AI will be extremely intelligent, yet somewhat autistic, like Yudkowsky's crowd, and just as powerless and socially awkward as they are.
> Maybe an artificial intelligence will be just as limited, just as slow as humans, only a little less so.
Only you can duplicate them with far more ease, and have each instance try out different approaches.
So if an AI reaches human-level intelligence at sufficiently low computational costs, we can assume that even if something prevents it from scaling up the intelligence accordingly, it will be possible to scale up the achievable goals dramatically through duplication the same way humanity is able to achieve far more through the sheer force of numbers than even the smartest of us would be able to achieve on their own.
EDIT: you don't even need to assume they'll be able to reach the intelligence of a particularly smart human. A "dumb" AI that has barely enough intelligence to just figuratively bash their collective, duplicated heads against enough security systems for long enough to find enough security holes might be able to cause sufficient damage.
I am poking fun, and I'm not discounting the possibility that AIs could be dangerous (as could people, viruses and earthquakes). But it's very hard not to notice how the singulariters' belief in noncorporal-intelligent deities is a clear, a bit sad, reflection of their own fantasies.
Yudkowsky and friends like to argue for what they call "rationality" (which is the name they give a particular form of autistic thinking imbued with repressed fantasies of power -- or a disaster only they can stop -- which is apparent in all of their "reasonable assumptions"), but their "larger than zero probability" games could be applied to just about any dystopic dream there is. I can say true AI won't happen because a virus epidemic would wipe out the human race first; or over-population would create a humanitarian disaster that will turn us back into savages; or genetic research would create a race of super-intelligent humans that would take over the planet and wipe-out all AI, and I could go on and on.
I suppose it's unfortunate that he came up with such an amazingly provocative way of demonstrating his argument, it's somewhat eclipsed the argument itself. I am definitely a victim of nerd sniping here. It must be the open-ended secrecy that does it.
That may well be the case, but my only other information source is HN comments, and not those made by detractors either. If there are sites or articles dedicated to the deconstruction of LW ideas, I'm not privy to them, nor am I interested in seeking them out. Basically, I only remember LW's existence when it comes up, always accompanied by fawning comments, on HN.
> the reasons why the Basilisk does not work are _not trivial_ - it's not a simple Pascal's Wager
Correct. While my value judgement of both is the same, my reasoning about why the Basilisk is not a thing ultimately consists of more components. That doesn't mean it's worthy of more consideration though.
> because with Pascal's Wager, we don't have the ability to actually create God
I would not say this is centrally important, because the processes leading to the creation of AGI are in all likelihood not going to be influenced by the existence of the Basilisk thought experiment either way.
> and in fact publicly stated that he won "the hard way", without a one-size-fits-all approach.
Again, I have to take my cues from the perspective of an outsider looking in, and there are several people who commented in this thread alone who described it very, very differently. Of course, a movement is not directly responsible for all its fans and members - but among the advocates for the validity of the AI Chat experiment, the idea that out there is a mystical one-size-fits-all rhetorical exploit seems very much alive. It may be cynical, but I can't help noticing how this aura of mystique and secret knowledge seems to work very well when it comes to attracting fans.
Of course, ultimately, these are just memes - and like many memes they propagate best when reduced to an absurd core. It doesn't even require intent.
I agree, and I am totally with you on this - I disagree with that interpretation wherever I see it. :) That's not exactly Eliezer's fault tho, and I guess it's to be expected that geeks attach to "clever" answers. I do think it's a bit unfair to judge the entire site by the two posts out of hundreds that happen to be in all the news articles - which LW has no influence on.
Inasmuch as _fans_ judge the site by these two articles, I'm just as much against that. I don't want LW to have an aura of mystery; that largely defeats the point!
[edit] I think a big part of the problem is that online reporting selects for clickbait.
-----------
Obviously I agree with 1. I'm not sure that's inevitable but I think it's very likely.
> If it is the case that humans can build human level intelligence (which I have argued they can and will), then it is also pretty self evidently the case that humans can improve on that intelligence.
That's not self-evident at all. We don't even know what intelligence is, so it's possible a human is as intelligent as possible.
> For example, given we have the technology to build a machine with identical mental characteristics to a human, we would also have the requisite knowledge to, say expand that machine's working memory by 1%. Or to create the machine with swappable sensory organs so it can directly absorb lots of different types of data.
That's not so clear at all. You could say that the internet has already increased human memory by orders of magnitude, yet it hasn't made us "super intelligent". It is not certain that we can actually increase the "in brain" working memory of an intelligent machine without, say, giving it a mental illness at the same time.
You assume that we can take "intelligence" and make it as malleable as we like, change every parameter to our wishes, but until we know what intelligence is, that's not at all certain.
> You're not actually any smarter, but with 10 of you, you're more productive than any single human being could ever be.
But not more productive than 10 humans, and at some point you might become less productive. It's not at all unlikely that those 10 copies of you start hating one another and won't be able to work together.
And even if that were true, the self-improvement progress will likely be very slow: http://www.sphere-engineering.com/blog/the-singularity-is-no...
Of course, the big question I pose to you in the end is why do you think an improved mind is even an important goal in the first place, but we'll get there.
> Now your ideally-coordinated team of smart machines just got smarter across the board.
Like I said, 1/ you don't know that they'll be ideally coordinated, and 2/ you don't know how much harder it is to make them smarter. It may possibly require exponentially more resources.
> Soon you're millions strong, you have teams working on brain improvement, coordination improvement, science teams of all disciplines, a resource acquisition team playing the stock market, each moment shipping new improvements to your brain and amassing resources and power.
That's a very nice fantasy, but we already have lots of intelligent beings. Are they coordinating like that? To some degree, they are. But why do you think "smarter" things could do it better? Are smarter people better at amassing resources? At getting power?
> There you have a computer controlling hundreds or thousands of independent agents.
Yes, there's a non-intelligent machine that's doing that. But the intelligent machine playing that game can only play one. You have this incredibly powerful machine in your head, and it can only concentrate or one thing or maybe two. It is possible that if you want an intelligent mind that is both coherent with itself and able to concentrate on lots of things at once then you need a number of processing units (say, neurons) that grows exponentially with the number of things you want to do at once. And maybe that's not even possible at all. Maybe beyond some point, the intelligence simply goes mad.
> Why not just write the software for the robots you control to be mostly autonomous, and give yourself an API through which to issue commands that the robots can mostly execute on their own?
Why don't you do that? Because it's hard and may take years, and some guys with guns realize what you're doing before you finish and come arrest you.
> Further, why even do (b), when you have the ability to replicate your brain?
Because my brain replicas might very soon decide they don't want to play together.
> So, with this narrative of how it could work, do you see how dangerous an AI could be? One that isn't stably goal-aligned with basic things like life on earth?
Of course it could be dangerous (I never said I couldn't imagine a scenario where an AI would be so powerful and so dangerous), but I also think I've demonstrated why AI may not necessarily be as powerful as you think. Also, by the time all this happens, I think there are more serious threats to human existence, like pandemics and climate change. They won't kill us, but they might set us back AI-wise a few centuries. So of all possible dangers we must consider at this point in time, I would put a hostile and dangerous AI waaay down on the list. It's possible, but it's far more likely that other stuff would get us first.
-------
One of the dangers that are waay more likely than the scenario you've described, is a sub-intelligent "hostile" software (that is perhaps not intentionally hostile, but indifferently so). It's more likely that some far dumber-than-human machine would replicate itself with full coordination over its replicas and wipe us out.
I really think that this emphasis on intelligence as the danger is a personal fantasy of singularists who'd like to think of themselves as powerful and dangerous (or the only stopgap against that). You really don't need to be so smart in order to amass power.
Another example: What I fear more than a super-intelligent AI is a super charismatic AI. That AI, without replicating itself, without trying to improve itself, simply charms a lot of people into following it, and establishes a reign of terror. Alternatively, as charisma is more effective than intelligence in controlling others, I would find it more likely that a charismatic AI would be able to control her replicas. It doesn't need to improve its own intelligence, because it's intelligent enough to inflict all the damage it wants.
Or what about a cunning AI? Sure, cunning is correlated with intelligence to a degree, but beyond a certain point you don't see that very intelligent people are very cunning. Sometimes far from it (Eliezer Yudkowsky is a prime example; his thought process is so predictable, that if our AI deity were like him, some cunning people would quickly find ways to foil its plans; that's a very non-dangerous AI).
We've seen how one charming person is far more powerful than thousands of really smart people. We also see how very dumb insects can coordinate themselves in very large numbers in a very impressive way.
I think intelligent people tend to overestimate the importance of intelligence and not to notice that other abilities we see around us every day are far more powerful. If anything, you'll note that very high intelligence is often correlated with low charisma or with a sort of timidity -- nebbishness if you will. A super-intelligent AI would probably be super-nebbish :)
Now, a super-charismatic AI -- now that's scary. Or a non-intelligent army of insects.
So if I were to ask you one question it would be this: Why do you place such a high emphasis on intelligence in the scenario you've described?
I'm not sure what your background is, but I'm noticing that I probably have significantly more detailed models of both intelligence and cooperation than you do. Please don't take that as an insult at all, it's just that it's part of my work to know these things. I think the inferential gap may be too wide for a reply here :/
I think your position is understandable given your priors, you're not crazy for thinking what you think.
I feel like an asshole for stopping the conversation with "I know more than you, but I won't explain it," it's a total dickhead move, and I'm really sorry. If I had time, I'd write much more and I bet it would be great for both of us. Maybe someday we'll have a chance :)
My main motivation wasn't to challenge your beliefs in the capabilities of an artificial intelligence. I fully realize it is possible that it will take the form of an all-knowing god, but to challenge your beliefs in our ability to predict technological advances and threats to humanity (of which I doubt you have any promising models, because there aren't any), as well as to provoke thought about how power flows in society, where a single charismatic Hitler was far more dangerous than 100 Einsteins, and whose charisma swayed people regardless of their level of intelligence (in fact, people with humanities background like writers, were, on average, far less likely to be swayed than people with science background). I think that an intelligent 100x that of humans (whatever that means) is still likely to be swayed by a charismatic person. I am not suggesting that AIs would thus be subjugated to people or integrated to our society, necessarily, but that they, too, might feel their own social forces, and there are some interesting dangers lurking there. For example, consider this: what happens not if someone invents an AI, but if 50 labs invent 50 AIs at the same time (which, frankly, is far more likely). How would those AIs interact with one another? How would their goals evolve as a result of that interaction?
Most discussions of AI concentrates on intelligence and goals/motivations -- which is understandable given the psychology of the people participating on the discussion -- and not enough on the AI's psychology and social psychology (even amongst other AIs). I probably know the answer you have to my question of "why do you think it is intelligence that is so dangerous", but that is not the right answer, or, at least, not the interesting one (even if you define intelligence as a general problem-solving skill, and higher intelligence = more problems of any kind solved faster -- though there are very good reasons not to define intelligence in this way) :)
My problem with the singularitarians' "threat model" is not that it is not plausible, but that they have not given serious consideration to other kinds of threats, even those including AIs. They tend to fall into the trap many people who are only science-educated do, and fall in love with their one pet-theory rather than truly try to challenge it. Yudkowsky's discussions of potential challenges to his "AI threat" are rather laughable, not because of how he counters them but by how narrow his thinking is (which is a result of his obsession with what he calls rationality).
I'm sure Yudkowsky would find my discussion of his followers' psychology ad hominem and extremely irrational, but that is not the case if you're trying to evaluate not a particular threat model, but to construct a meta-model of threat models and how you weigh one against another. If you construct a Bayesian analysis of threats given that meta-model, you'll see that psychology plays a big role.
Another thing I find curious about his thinking (which very much affects how he weighs threats) is the assignment of negative infinity to the event "destruction of the human race". I think that if you consider the following -- your own death, the destruction of your life's work, your own physical suffering, the death of your entire family, physical suffering of your entire family, destruction of your country (this may not seem relevant to some, but I have a good reason to include it), destruction of your culture, destruction of human civilization, destruction of the human race, and destruction of the planet -- you'll either find that people assign negative infinity much sooner than "destruction of the human race", assign that to events other than that one and not to that one, or to none of them, and in the latter case, the jumps in values would be rather surprising (or not, depending how you look at things). For example, I would give the destruction of civilization and destruction of the human race the same value (not negative infinity, though), and the destruction of the planet a much, much lower (more negative) one. For instance, I find a nuclear holocaust a much worse result than an AI destroying all humans but keeping all animals alive, even if some humans survive the first event (but not civilization) and none survive the second. I would also assign a lower value to all humans becoming fascists than to an AI destroying all humans.
But I'm sure there are others who are more thought-provoking than him who discuss this topic.
Plus, the other two players who won as gatekeepers - Eliezer would presumably have tried to cheat against them, too.
FWIW, I downvoted your original comment on this thread (and only that one) for being vague, snarky and dismissive. If you wish people to engage with you, I recommend not starting off like that, although it seems to have turned out okay in this case.
And I recommend you give people the benefit of the doubt, though honestly I have to say I frequently fail at that myself. For example, your comment could be perceived as somewhat condescending, but I force myself to categorize it differently. I also know that I can come across way more negative than I intend to, I apologize for that and I'm working on it.
For what it's worth, I do think my original comment was snarky and dismissive, but somewhat counterintuitively that's not usually what gets people downvoted and flagged on HN. People can and do get away with artful personal attacks on HN all the time, at least in my defense I can say I attacked an idea instead of a person.
It may well be the case that my insufferability amplified the reaction, but I posit the root cause was disagreement about the message, not its format.
> although it seems to have turned out okay in this case.
It turned out okay because a decent dialogue emerged from it, one of the very few in this entire thread. But it was sufficiently controversial to get enough downvotes in order for my comments to teeter around 0 points and also receive flags. There have been a few updates to HN's comment ranking and voting algorithms that will make me regret taking this stance for some time, which may or may not provide you with some comfort to know.
[edit] To be honest, I didn't even know you could downvote on HN. I've never seen a downvote button.
[edit] Not that I would have.
It's offtopic, but when you reach 500 karma (or thereabouts), you get the ability to downvote the comments you feel are a net-detriment to the discussion. Sometimes, when I express a particularly unpopular opinion I also get flagged. The result is what I call "the doghouse", it causes comments to sink to the bottom like stones for a few months, among other things. HN moderation has hidden mechanics and the results can be very frustrating. But in 99% of cases, when those special mechanics are not in effect, the system works pretty well.
I'm saying that there is a chance of that being the case, but that without any kind of third party confirmation we cannot know either way. Also see homeopathy.
> Eliezer would presumably have tried to cheat against them, too.
Not necessarily. Losing occasionally is a good strategy when running a con.
Under common interpretations of those phrases, that's a massive swing in your confidence levels.
I also misspoke in my earlier comment, i meant "remote chance of knowing that any rules were followed". I didn't mean to imply any confidence on the size of the chance, since we actually don't know enough to make such a judgement call.
My mistake, sorry.