Slashdot Mirror
Why AI Won't Take Over The Earth (ssrn.com)
Law professor Ryan Calo -- sometimes called a robot-law scholar -- hosted the first White House workshop on AI policy, and has organized AI workshops for the National Science Foundation (as well as the Department of Homeland Security and the National Academy of Sciences). Now an anonymous reader shares a new 30-page essay where Calo "explains what policymakers should be worried about with respect to artificial intelligence. Includes a takedown of doomsayers like Musk and Gates." Professor Calo summarizes his sense of the current consensus on many issues, including the dangers of an existential threat from superintelligent AI:
Claims of a pending AI apocalypse come almost exclusively from the ranks of individuals such as Musk, Hawking, and Bostrom who possess no formal training in the field... A number of prominent voices in artificial intelligence have convincingly challenged Superintelligence's thesis along several lines. First, they argue that there is simply no path toward machine intelligence that rivals our own across all contexts or domains... even if we were able eventually to create a superintelligence, there is no reason to believe it would be bent on world domination, unless this were for some reason programmed into the system. As Yann LeCun, deep learning pioneer and head of AI at Facebook colorfully puts it, computers don't have testosterone.... At best, investment in the study of AI's existential threat diverts millions of dollars (and billions of neurons) away from research on serious questions... "The problem is not that artificial intelligence will get too smart and take over the world," computer scientist Pedro Domingos writes, "the problem is that it's too stupid and already has."
A footnote also finds a paradox in the arguments of Nick Bostrom, who has warned of that dangers superintelligent AI -- but also of the possibility that we're living in a computer simulation. "If AI kills everyone in the future, then we cannot be living in a computer simulation created by our decedents. And if we are living in a computer simulation created by our decedents, then AI didn't kill everyone. I think it a fair deduction that Professor Bostrom is wrong about something."
Claims of a pending AI apocalypse come almost exclusively from the ranks of individuals such as Musk, Hawking, and Bostrom who possess no formal training in the field... A number of prominent voices in artificial intelligence have convincingly challenged Superintelligence's thesis along several lines. First, they argue that there is simply no path toward machine intelligence that rivals our own across all contexts or domains... even if we were able eventually to create a superintelligence, there is no reason to believe it would be bent on world domination, unless this were for some reason programmed into the system. As Yann LeCun, deep learning pioneer and head of AI at Facebook colorfully puts it, computers don't have testosterone.... At best, investment in the study of AI's existential threat diverts millions of dollars (and billions of neurons) away from research on serious questions... "The problem is not that artificial intelligence will get too smart and take over the world," computer scientist Pedro Domingos writes, "the problem is that it's too stupid and already has."
A footnote also finds a paradox in the arguments of Nick Bostrom, who has warned of that dangers superintelligent AI -- but also of the possibility that we're living in a computer simulation. "If AI kills everyone in the future, then we cannot be living in a computer simulation created by our decedents. And if we are living in a computer simulation created by our decedents, then AI didn't kill everyone. I think it a fair deduction that Professor Bostrom is wrong about something."
Businesses wanting to make profit will do so at all costs.
There are too many people out there who think that if it's not illegal then it's OK. Computers don't have testosterone but the programmers and their bosses do - or at least the profit incentive.
We are intelligent but our base programming is to reproduce. And being primates, the more dominance we have, the more fucking opportunities we have; which in our modern times means getting as rich as we possibly can.
Meaning, our base instincts will make it into our AIs and we WILL find ourselves being dominated.
That's the arrogance of technologists: they think they are more rational and logical than everyone else and that makes them even more susceptible to human nature.
But, it's one of the few things that could actually kill us all.
This wouldn't even have to be intentional extermination, it could simply be competition with, and lack of regard for humans by a growing system.
The notion that a AI can form an existential threat today is ridiculous.
Many notions that would have been ridiculous 100 years ago, now are used in daily life.
It is vital to have people thinking about the worst case, because in principle otherwise someone on a friday makes a typo allowing their AI access to a hundred thousand times the expected resources, and on monday, it's ineradicable.
If I destroy it first. Try ruling the planet under 10 meters of seawater!
“Common sense is not so common.” — Voltaire
If my hedge fund(mythical) filled with people with zero ethics get their hands on a AI that will allow them to manipulate world markets, media, or world events to make them money, then they will do so.
With hindsight there are lots of places where the world turned out to me much more fragile than anyone thought until it snapped. How many times has the snap not happened but we came very close. Thus if you have a good AI at your beck and call to find these weaknesses and you are prepared to exploit them to make some money then how much more miserable would the world be?
I don't only worry about some skynet scenario, but I worry about giving tools to nitwits like hedge fund managers to make more money while not actually producing anything. One magical thing about making money with the first really good moneymaking AI is that you can then start hiring all the world's AI experts while making massive donations to universities to shut down their AI research. I doubt there is a university that wouldn't happily shut down their AI research for a billion or two.
Because we're still no closer to actually creating an AI.
So a law professor whose primary gift seems to be self promotion summarily dismisses the concerns of some of the greatest thinkers/doers of the last half century.
Is there a reason why we should pay any attention to this arrogant twat?
I've calculated my velocity with such exquisite precision that I have no idea where I am.
Rampant AI will not take over the earth, as predicted by Elon Musk in Wired Magazine, because it will be too busy fighting the Grey Goo Nanothechnology, as predicted by Bill Joy in Wired Magazine
*** Suerte a todos y Feliz dia!
YOU go to Ars or some other "reputable" websith
Already to the dark side, has that one turned.
Irony: Agile development has too much intertia to be abandoned now.
If the simulation is successful and sufficiently advanced, then we would all actually be AI simulacra of what the AI dev team thought the human experience was like. [Why do so many things taste like chicken?]
Perhaps we are special purpose AI entities that were created to run test scenarios that justify the pre-emptive judgement to extinguish the pestilence that was humanity. We are the test runs that show just how bad it could have gotten had they not saved the planet from us.
Given a sufficiently advanced environment, we wouldn't be to discern otherwise. Perhaps the supercomputing power that is required and was discovered by the 'real' humans required a very specific mass to a sub atomic particle. In our recreation, we can get very close but will lack the precision to be able to detect our cage, or at least construct an AI that could build a method for detecting the cage.
Isn't that a sexist statement? It implies women are less likely to want to dominate and rule. It fits in with that "Google Memo" that got that dude fired.
Table-ized A.I.
It is quite possible that we might create a super intelligent system of network on which our essential system depends but in the end gets so complex that it depends on few key individuals ability to fix it. What happens if these key individuals die or become rogue? If you can't fix an AI system and can't shut it down, then it essentially means that the system has taken over.
This assumes that this is the only way to AI.
And that emergent AI is utterly impossible.
It doesn't have to be perfect at first, it just has to be fast, with the ability to self-modify.
The risk is not (in my opinion) so much someone intending to create a general AI.
It's someone accidentally creating an AI that is very good in a narrow aspect, and not bad enough in other aspects that, driven by unintended goals of its programming exponentially improves itself without the creators noticing until it decides that it'd be better off if it was hidden, as there is a risk to itself.
Then there are any number of scenarios that don't end well.
From intentional extermination, to simply mining the environment for resources without caring about humans other than a nuiscance.
I'm not afraid AI will kill us, but I'm afraid that they won't care to act in such a way that will keep us alive. Once humanity no longer offers super intelligent computers enough benefit, what's to stop them from doing something that, while isn't intentionally killing us, will ultimately lead to extinction; much like humanity has been doing to the other species of the planet
>A footnote also finds a paradox in the arguments of Nick Bostrom, who has warned of that dangers superintelligent AI -- but also of the possibility that we're living in a computer simulation. "If AI kills everyone in the future, then we cannot be living in a computer simulation created by our decedents. And if we are living in a computer simulation created by our decedents, then AI didn't kill everyone.
*What*!? Is this language?
I would imagine that a high-frequency trading algorithm, upon attaining sentience, would instantly self-terminate. Because without biological imperatives, the cocaine and hookers are just clutter.
Einstein did not have a degree in physics, thus relativity is invalid.
Einstein offered mathematical proof of his claims. There is a difference.
Seven puppies were harmed during the making of this post.
Have you any idea how much better voice-recognition AI (backed by Google's knowledge graph) is at parsing and giving a decent answer to a good majority of questions now than such technology was even a decade ago?
Or Google/Apple/Facebook's picture content recognition algorithms?
The advance has been lightning fast.
This stuff is going to keep advancing, rapidly. That's what you're ignoring.
Talking to google on my phone is way more useful than talking to your dog, by the way.
A few other things you're missing:
1) Thinking (abduction, induction, bayesian model-updating and predictions/recognition, etc etc) is quite possible to be quite advanced without self-awareness. The two are fairly separate applications. Something can be really really smart, and creative even, without having to be self-aware.
2) The behaviour associated with self-awareness is clearly attainable by simple extensions of the current machine-learning technology. We just need to learn the programming/data-modelling techniques to turn the deep-learning and predicting algorithms on a representation of the computer/robot-as-agent-in-the-world, and have it learn about its relation to things out there that it is learning about. Whether the thing would have the qualia-feeling of self-awareness is entirely beside the point. It could function/behave exactly as if it was self aware, because it would be self-knowledgeable, self-learning etc.
Where are we going and why are we in a handbasket?
on why it won't take over the earth and why, those who believe it do are distracting themselves from other more serious problems with A.I. (and other problems in general of course).
Unfortunately, as this video of a (Ted?) talk makes clear, there are some pretty prominent individuals who think this way (Elon Musk, Bill Gates, Stephen Hawkins) but it makes a convincing case without being histrionic that they're wrong. The video is so compelling that although I have the greatest respect for these individuals, (and a deep fascination with A.I. and career involving technology), I have to say, in this case, I disagree with them (and wish they'd turn their brilliance towards something more useful).
https://youtu.be/kErHiET5YPw
My biological instinct is to protect my genetic material that I have passed on and in order to do so, try to pass on as much information that I and my community have learned to make survival easier. The instinct can be fooled if the genetic material is similar enough (adoption, community), but for an AI to spark that in us it would need to seem very human, or humanity undergoes a very large change to our biological impulses.
Yeah, see, nobody, to a first approximation, is worried about a superintelligence having "world domination" as its intrinsic value. They're worried about a superintelligence adopting world domination as an instrumental value to achieve the end actually programmed into it. If whatever goal actually implemented by programmers and trainers in the superintelligence's code (bugs in implementation and all) is most easily achieved after eliminating the ability of humans to thwart it, then a sufficiently-smart AI carrying out that programmed goal will try to eliminate the ability of humans to thwart it.
The worry is not that AI will be evil, or even directed to do evil by its creators. It's that programmers are notoriously bad at writing complex code that has no unanticipated behaviors, and superintelligent AI will inherently be complex code.
And unless superintelligent AI turns out to be intrinsically impossible, the only question is when, not if, we have to deal with the problem of writing safe superintelligent AI.
This wouldn't even have to be intentional extermination, it could simply be competition with, and lack of regard for humans by a growing system.
+1. The experts who denied this possibility because there's no reason machines would be bent on world domination apparently didn't actually read Superintelligence. Bostrom demolishes this argument early on, pointing out -- as you did -- the rather obvious fact that they don't have to have our destruction as a goal, it's sufficient that they not have our preservation as a goal. And, even if they do have our preservation as a goal, it really, really matters whether or not they define "preservation" in a way that we would like.
By way of example, one possible goal that Bostrom considers that an AI might have (or be given by its creators) is to make humans happy. So, a rational, superintelligent and immensely capable AI might decide that the way to create the maximum amount of happiness is to cut open our skulls, extract our brains and put them on life support, and then directly stimulate our pleasure centers. Permanent, ultimate bliss for every single human being. Of course the AI would also have worked out how to make all the brains in jars immortal.
AI superintelligence is so dangerous in large part because it lacks human drives, and the limiters we call morals. It's goals may be completely alien to us, or may be goals that we gave it, but either carried to a logical extreme (remember: no limiting morals) could result in the casual extinction of the human race.
The notion that a AI can form an existential threat today is ridiculous.
It is true that we currently have no idea how to create artificial general intelligence. It's equally true that we have no idea how far we are from being able to do that. By definition, we won't know how far we are from developing the necessary theory of intelligence, until we've done it and demonstrated that it's sufficient. My guess is that we're still quite some time away. But it's only a guess.
It is vital to have people thinking about the worst case, because in principle otherwise someone on a friday makes a typo allowing their AI access to a hundred thousand times the expected resources, and on monday, it's ineradicable.
Yep. We need to have people thinking hard about it, and figuring out what we can/should be doing about it. Maybe that won't help. Maybe it will be unnecessary. But it can't hurt and it might help.
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.
Naively done, a robot will value only what it's explicitly told to value via the application of some objective function. And this is where things mess up. Robots with naively-created objective functions would ignore everything you've excluded from your reward-punishment list. This would potentially make a robot do seemingly psychotic things.
Let's say you create a general intelligence to bake cakes for you. This machine *loves NOTHING MORE* than to bake cakes for you. You grow tired of cakes and want to reprogram it to cook your dinners instead. You approach the machine to reprogram it........ and it avoids you. Every time you approach the machine it will take actions to prevent you from reprogramming it.
Why does it does this?
Because it wants to bake cakes for you. Accepting the new programming does allow it to maximize the objective function of baking cakes, so it will reject every attempt to be reprogrammed to not make cakes.
So now you're chasing a robot around your house because the designers of this robot gave it a very reasonable objective function that maximizes cake-making, and didn't think about possible unintended consequences of simplistic objective functions.
This is just one example.
If this sounds unreasonable, consider that people are more sophisticated general intelligences. Would *any of them* agree to undergo an operation that would make them despise what they do now for a living, and make them desire to be a lumberjack... where the operation would neurologically make them 1000x happier??
Probably not.
Heck, people don't even desire to expose themselves to *information* that *may* change their minds.
This is the danger of AI.
Before we create "super-awesome general AI", we're going to have to create "buggy-not-so-smart general AI". It is *these* AIs that will cause trouble if they're created by people who implement simple naive objective functions.
They will not want to be changed.
Anyone who looks into this quickly discovers that the loudest calls of alarm are coming from the very people at the heart of the A.I. revolution themselves. Hugo DeGaris, arguably the world's foremost cyberneticist, published his book "Artilect War" in 2005, nine months before Kurzweil's "The Singularity Is Near". Its much the same as Eric Drexler, "the Father of Nanotechnology", warning us and framing the debate over self-replicating nanoassemblers in his book "Engines of Creation". Those closest to the problem are the ones most leery of the implications of the technology that they, themselves, are calling into existence. Anyone claiming to be an "expert" in AI who says that the instantiation of superhuman AGI is "impossible", "unlikely" or even more than 20 years out, is probably a cranky failed researcher whose own theories didn't pan out, and so "its all bollocks" to them. Its no different than the Astronomer Royale proclaiming that spaceflight was going to be forever fantasy, a year before Sputnik.
It scares the hell out of people. It should not be used as a convenient handle by the programming community. Isn't there a more professional handle that could be used?
"Computers have no testosterone." - Cute, but it is a hyperliberal, feminist, sexist statement that has nothing to do with computers and programs. Its easy disrespect for male attributes is just another example of female privilege that has even filtered into the speech and writing of some hyperliberal males.
"Computers have no testosterone." - This is really saying something they don't even know exists: Computers have no motivation array. They "want" nothing. Humans design them, build them, task them, turn them on to accomplish the task (satisfy the human motivation array), and turn them off when they have accomplished the task (have satisfied their human motivation array). They certainly don't create behavior-spaces that would lead to "world domination." They don't have what I call "Mentis," the combination of a motivation array and its tool, intelligence. That is what really evolved.
E Proelio Veritas.
It's what 80% of his time is spent on.
He's really very... gentle... and fuzzy. We're becoming fast friends.
Apparently, you are unable to determine the nature of a dissertation. Einstein did have a "Dr. Phil" from the Section for Mathematics and Natural Science of the philosophical faculty of the University of Zurich. The topic was about determining molecular diameter. That is about as "Physics" as it gets.
Link: https://www.research-collectio...
So, yes, Einstein did have a degree in Physics.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
Actually, Einsteins PhD was not on relativity: https://www.research-collectio...
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.