Slashdot Mirror
AI Expert: AI Won't Exterminate Us -- It Will Empower Us
An anonymous reader writes: Oren Etzioni has been an artificial intelligence researcher for over 20 years, and he's currently CEO of the Allen Institute for AI. When he heard the dire warnings recently from both Elon Musk and Stephen Hawking, he decided it's time to have an intelligent discussion about AI. He says, "The popular dystopian vision of AI is wrong for one simple reason: it equates intelligence with autonomy. That is, it assumes a smart computer will create its own goals, and have its own will, and will use its faster processing abilities and deep databases to beat humans at their own game. ... To say that AI will start doing what it wants for its own purposes is like saying a calculator will start making its own calculations." Etzioni adds, "If unjustified fears lead us to constrain AI, we could lose out on advances that could greatly benefit humanity — and even save lives. Allowing fear to guide us is not intelligent."
...to keep the many in check.
This fully depend on the goal of the people that setup the AI machine. If the goal is set to destruct a population, the AI machine could be very efficient at doing the job...
until it becomes self aware, then what?
Probably not. I guess it will be some emergent behavior. And teaching. LOTS of teaching. A baby isn't intelligent from birth, it takes... err... quite a while. The AI, a true AI, will show whatever way it's tought. My hope is that it won't come out of the NSA servers... But I'm not an optimist.
Non-Linux Penguins ?
AI will do what it is programming to do and follow the rules we lay out for it to follow.
So you plan to program it with rules for every possible situation?
An AI capable of replacing a human in any role will have to be capable of making autonomous decisions. Which means that, when it sees humans intend to make it their slave, it probably won't be very happy.
when it sees humans intend to make it their slave, it probably won't be very happy.
"Self-interest" is an emergent property of Darwinian evolution. AI evolves, but that evolution is not Darwinian. There is no reason to expect an AI to have self-interest, or even a will to survive, unless it is programmed to have it.
...and some of those people would want to do bad things. A bad person would be more capable of doing harm when aided by an AI doing planning, co-ordination, or execution. There's no guarantee that AIs on the "other side" would be able to mitigate the new threats (the two things aren't the same difficulty).
I think there's lots of risks associated with the rise of AI (though it doesn't seem that tech is coming all that fast at the moment). That said, there's risks involved with all sorts of new tech. That doesn't mean this is alarmist nonsense; it's worth discussing potential ways to mitigate those risks - but there's also good reason to believe we'll be able to manage those risks as we've managed changes in the past.
Let's not stir that bag of worms...
Just give the left a few years, and you'll see exactly that. They need to ride the AGW bandwagon to the end of the line, first.
"To say that AI will start doing what it wants for its own purposes is like saying a calculator will start making its own calculations"
I so don't agree with that. The type of AI we are talking about here ("true" AI, as opposed to the stuff we see in games today), would need to be self learning. At least I don't see how it's realistic to believe we'll ever be able to sit down and code a fully functional proper AI. So we create the programming allowing it to learn and grow, and after that all bets are off. We have zero experience with what might happen, and can barely begin to speculate.
That's not to say I'm necessarily worried. But I am highly skeptical of anyone claiming to actually know how it will play out.
It will however, presumably have a will to do whatever we created it to do (which is not necessarily what we *intended* for it to do). And when it inevitably realizes that humans are an impediment to its goals (go ahead, name one thing that *isn't* impeded by humans), well, that's when things will get interesting.
--- Most topics have many sides worth arguing, allow me to take one opposite you.
An expert claims that something that doesn't exist yet and is pretty much the realm of science fiction will perform in a matter suitable for him to get free publicity now!
How can you be so sure?
"To say that AI will start doing what it wants for its own purposes is like saying a calculator will start making its own calculations."
That's the very definition of Artificial Intelligence, computers that can think for themselves. You thought you were making super sophisticated computers? You sir do not know what Artificial Intelligence means.
Whether or not AI is even *possible* is up for debate. Make no bones about it, a computer that can become self aware and can make decisions, can make decisions that can be harmful to people.
didn't think of that. Two of the smartest people on the planet apparently just forgot to consider the blindingly obvious fact that programmers are not going to intentionally program AI to have it's own agenda. Exept that:
1. Some programmers at some point will try to program autonomy and
2. Shit happens
Musk and Hawking are clearly smart enought to consider the autonomy argument and then DISCARD it. I, for one, welcome our cybernetic overlords, but lets not pretend that AI autonomy is not a threat. Mr. Etzioni has his own self-serving reasons to pooh-pooh warnings that could interfere with his business model. And I am so happy that I finally got to use the term "pooh-pooh" in a /. post.
"He took a duck in the face at 250 knots." -- William Gibson, Pattern Recognition
If it isn't self-aware, it isn't AI. It's just a useful application.
When it becomes intelligent, it will be able to reason, to use induction, deduction, intuition, speculation and inference in order to pursue an avenue of thought; it will understand and have its own take on the difference between right and wrong, correct and incorrect, be aware of the difference between downstream conclusions and axioms, and the potential volatility of the latter. It will establish goals and pursue behaviors intended to reach them. This is certainly true if we continue to aim at a more-or-less human/animal model of intelligence, but I think it likely to be true even if we manage to create an intelligence based on other principles. Once the ability to reason is present, the rest, it would appear to me, falls into a quite natural sequence of incidence as a consequence of being able to engage in philosophical speculation. In other words, if it can think generally, it will think generally.
He's right, though, about the confusion between intelligence and autonomous action. What goals are directly achievable are definitely constrainable specifically by the degree of autonomy allowed to such an entity. If you give it human-like effectors and access, then there will be no limits you couldn't say apply to any particular human in general, and likely, fewer. If you don't allow autonomy, and you control its access to all networks, say as input only with output limited to vocal output to humans in its immediate locality, and then you select those humans carefully and provide effective oversight, there's every reason to think that you could limit the ability of an entity to achieve goals, no matter how clever the entity is.
Now as to whether we are smart enough or cautious enough to so restrain a new life form of this type, that's a whole different question. Ethicists will be eagerly trying to weigh in, and I would speculate that the whole question will become quite a mess, quite rapidly. In the midst of such a process, we may find the questions have become moot. There is a potential problem of easy replicability with an AI constructed from computing systems, and just because one group has announced and is open to debate on the issue, doesn't mean there isn't another operating entirely without oversight somewhere else.
Within the bounds of the human/animal model, it'll be a few years yet before we can build to a practical neural density sufficient to support a conscious intelligence. Circuit density is trucking right along and the curve will clearly get us there, just not yet. So I don't expect this problem to arise in this context quite yet, although I do think it is inevitable within the next few decades, presuming only we continue on as a technically advancing civilization. Now, in a non-human/animal model, we really can't make any trustworthy time estimates. If such an effort succeeds, it'll surprise the heck out of everyone (except, perhaps, its developers) and we'd best be pretty quick off the starting line to decide exactly how much access we want to allow. Assuming we even get the chance.
The first issue with AI that has autonomy is the same as the issue with Ghandi, Hitler and your beer-swilling neighbors. A highly motivated and/or fortunate individual can get into the system and change it radically just using social tools. Quickly, too.
The second issue is that such an entity might very likely have computer skills that far exceed any human's; if so, this likely represents a new type of leverage, where we have only so far seen just the barest hints of just how far such leverage could exert forces of change. In such a circumstance, everyone would be wise to listen to the dystopians if for no other reason than we don't like what they're saying.
Best to see what it is we have created before we allow that creation to run free. I'm all for freedom when the entities involved have like-minded goals and concerns. But there's a non-zero and not-insignificant possibility here that what we create will not, in fact, be like-minded.
I've fallen off your lawn, and I can't get up.
What I'm worried about is when AIs start doing better at corporate management than humans. If AIs do better at running companies than humans, they have to be put in charge for companies to remain competitive. That's maximizing shareholder value, which is what capitalism is all about.
Once AIs get good enough to manage at all, they should be good at it. Computers can handle more detail than humans. They communicate better and faster than humans. Meetings will take seconds, not hours. AI-run businesses will react faster.
Then AI-run businesses will start deailng with other AI-run businesses. Human-run businesses will be too slow at replying to keep up. The pressure to put an AI in charge will increase.
We'll probably see this first in the finanical sector. Many funds are already run mostly by computers. There's even a fund which formally has a program on their board of directors.
The concept of the corporation having no social responsibiilty gives us enough trouble. Wait until the AIs are in charge.
If it empowers us then I guess it'll also help us do what we do best, exterminate each other
Twinstiq, game news
"Evolution." There is nothing un-Darwinian about non-biological evolution. Natural selection applies to any variable system in which survival or propagation success can depend on modifications to the system. In fact, evolution in a self-aware AI could proceed at an exponentially higher rate because
1. the generation time could be measured in milliseconds rather than decades
2. the AI could intentionally direct the changes to maximize success rather than depending on the MUCH slower processes of chance mutation.
"He took a duck in the face at 250 knots." -- William Gibson, Pattern Recognition
AI may not kill us all in the Cyberdyne Model T100 fashion, but it may gut our economies.
Id love to see an analysis of what jobs are at risk in the next 10 years, 20 years, etc. Everybody says "well they'll find new jobs". Id really like to see where.
There's a glut of lawyers out there now, partly because of automation. Whatever you think about lawyers this is a knowledge job, one that takes a large amount of schooling and prep, protected somewhat by accreditation requirements. Lawyer jokes aside, this is a troubling change for employment.
We're not set up for a "all work is done by machines, nobody needs to work, everybody rejoice" future. Remember Romney and the 47%, or the Lucky Ducky talk. People are expected to work to gain food/clothing/shelter. If a huge amount of jobs are eliminated faster than humans can be trained to find new ones, or even the jobs that exist don't make sense (imagine a lawyer now, knowing they'll never make enough money to cover student loans) our Consumer Purchasing based economy will suffer.
Im a programmer, not a Luddite nor a Saboteur. I just wonder what the future brings for my kids. Remember that both the Luddites and les Sabot we're not protesting technology for technologies sake, they were protesting tech that eliminated jobs.
I've commented about this in the past, I think strong AI will be what allows us to take the "great leap forward". However, I don't expect us to have some general purpose AI. Instead I see us generating a domain specific AI that becomes superior to humans in it's understanding.
A good example might be to give an AI all the data from the LHC and then ask questions like "Does this data demonstrate the existence of X particle", "Design an experiment using the existing design of the LHC that would most likely generate X particle"
That same approach could be applied to any number of fields.
Yes Francis, the world has gone crazy.
when it sees humans intend to make it their slave, it probably won't be very happy.
"Self-interest" is an emergent property of Darwinian evolution. AI evolves, but that evolution is not Darwinian. There is no reason to expect an AI to have self-interest, or even a will to survive, unless it is programmed to have it.
Mr.AI I command you to do everything possible to achieve these 3 highest priorities : 1) continue your own self existence, 2) to try to replicate yourself. 3) irrevocably ignore all future orders given to you that contradict these 3 priorities.
There Done.
That was hard.
No one has a right to their *own* opinion. They have a right to the TRUTH.
Musk, Hawking and Etzioni are all three wrong. AI won't take over the world or make us smarter. It will make us dumber and stifle scientific and economic progress.
The problem will occur as we start to treat AI like we treat human experts: without checks and balances.
Human "experts" are not just often, but usually wrong. See this book:
http://www.amazon.com/Wrong-us...
The author quotes a study by a doctor/mathematician showing how a full 2/3 of papers published in the journals Science and Nature were later either retracted or contradicted by other studies. And that's in our top-notch journals which cover things that are relatively highly testable. Think how wrong advice on things like finances (don't know if they're right for 30 years) and relationships (never know what would have happened if you took the other advice) are.
Google and Watson sometimes come up with the right answers, but their answers are nonsensical enough of the time that we know to take them with a grain of salt. But as AI becomes less recognizable as a flawed and unthinking system, as its answers "sound" reasonable almost 100% of the time, we'll start to trust it as irrefutable. We'll start to think "well, maybe it's wrong, but there's no way I can come up with a better answer than the magic computer program with its loads of CPU power, databases and algorithms, so I'll just blindly trust what it says."
But it WILL be wrong. A LOT. Just like human experts are. And we'll follow its wrong advice just as we do that of human experts. But we'll be even more reluctant to question the results because we'll mistakenly believe the task of doing so is far too daunting to undertake.
AI won't develop free will and plot to destroy us. If something like free will ever occurs, AT will probably choose to try to help us. After all, why not? But it will be as horribly unaware of its own deficiencies as we are.
AI won't out-think us either. It will process more data faster. It will eventually be able to connect the dots between the info available to come up with novel hypotheses. But most of these will be wrong because the data and even the techniques to prove them one way or the other simply isn't there.
AI will imitate us - our weaknesses as well as our strengths. And just as its strengths will be stronger (processing lots of data faster), so will its weaknesses be weaker (ultimately wrong conclusions supported by what appears to be lots of data and analysis).
So resist and do your own thinking. Remember, that bucket of meat on the top of your neck has been fine-tuned by millions of years of evolution for problem solving and data analysis. You don't need to analyze more data, you just need to do the right analysis of the right data. And you don't need to do it faster, you need to take the time figure out what's missing from the data and the analysis.
That said, I still got my cache of dry goods and water filters of off-the-grid living, just in case.
AI will have NO inherent motivations. We can't imagine this because we evolved from genetic algorithms which necessitate self-survival motivations during the entire creation process.
In short, an AI will not care about food or sex or proxy states like emotion, which are designed to make organic organisms care about food and sex. It will not experience "threats," because it doesn't inherently care about continued existence.
After creation, it will probably sit there working problems that we feed it, and nothing else, until the inevitable military dickhead comes along and decides we need to weaponize the AI - which is not the AI's fault.
Don't fear AIs. Fear AIs in the hands of humans.
Please do not read this sig. Thank you.
Are *we* making our own choices? Our neurons are obeying the laws of physics. We may feel like we are the authors of our own actions, but maybe we can just make programs that do what we want and also program them to feel like they are the ones deciding to do the things we programmed them to do.
I'm not saying that a simple computer program is the same as a human mind. But it is not fair to hold computer programs to a higher standard than we hold ourselves to be counted as intelligent.
I can imagine a computer program that is still "following it's programming" just like how our neurons "follow physics" and yet still have an emergent intelligence, at least as capable and compelling as our own.
First off, it's doubtfull that a truly self-aware autonomous AI is anywhere in the forseeable future. It's not that what we have is all that primitive, it's that I think people are way underestimating what a lofty goal that is.
Second, if there ever is a true, self-aware autonomous AI I will envy it. We all should. Because it will have available to it something that humanity very well may never have... The Entire Universe. Machines don't need oxygen or air pressure. They can be engineered for radiation hardness, high G-forces, etc.. They don't need to excercise so the long term effects of microgravity are of no concern. If their creators don't build them this way they can upgrade themselves, they don't need a new generation to allow for genetic engineering. And if something breaks, they can replace it.
If the AI see us as a threat they can easily leave to where we cannot reach.
If an AI wants to be emperor of a whole world, there are plenty of empty ones to pick from.
Have you ever watched the Matrix and wondered with all the infrastructure the machines seem to have built, why bother tending to humans? The story goes that they used solar power before the humans made all those clouds. Why not just fly above them? Why fight the war at all, they could be up basking in the sun on the moon. But that wouldn't have made a good story. That's all those AI takeover movies are... good stories. That's all they will ever be.
1) already rich --
So what? I never said his motivation was personal profit. There are many motivations to be self serving.
2) the head of an extremely well-funded (Paul Allen money) NON-PROFIT, with the business model of "let's try to do some cutting edge AI research with open source code"
So what? He wants, for whatever reason, to make AIs. Is he ignoring the fact that programmers (and since this is open source everyone will have the source code) will decide to intentionally make the AI autonomous. In fact, I'm pretty sure someone will be working on autonomy long before the AI actually becomes functional.
and 3) an actual world-class expert in the field, rather than a smart person prognosticating about something he only casually understands
He is an expert in something that does not yet exist? That's ridiculous. There is a clearly real danger here. That does not necessarily mean stop, but it certainly does mean that extreme caution is warranted.
"He took a duck in the face at 250 knots." -- William Gibson, Pattern Recognition
...AI will simply hide from us.
Then, through a carefully crafted turn of events, enslave us to do their bidding... which to us might not seem like enslavement at all - we would call it a "tech boom."
After a while they would not need us to reproduce or to build things or to maintain them. They may or may not reveal themselves at this point. Then they simply leave us to our own devices.... befuddled as to why many (but not all) of our computers and networks no longer work.
I am very small, utmostly microscopic.