How Far to AI Utopia?
milesbrundage.com
milesbrundage.com
The first is the very near future, 10-20 years or so. Machine learning is rapidly advancing and we are getting all kinds of cool toys. Automating things that previously weren't automatable. 90% of jobs probably could be automated soon, and there will be huge societal consequences. That's what this article is about, mostly.
The second is far more interesting. Strong AI. The distant future of 30 years, and maybe sooner. I think it's very likely this will happen in our lifetime, though obviously it's impossible to know for sure.
This isn't just replacing a few jobs with machines. This is replacing all humans with machines. When we get AIs that are just as smart as humans. And eventually far smarter than humans. AIs which are so far beyond us that we look like chimps.
We would not be able to compete with the AIs. They would be able to get whatever they want. Humans would no longer be in control. It's very, very important that we make sure the AIs we build want what we want. Then we can get a utopia so advanced we can't even imagine it. But if we get it wrong we will just be replaced or destroyed.
And we haven't the slightest idea how to do this, or even if it's possible. Creating an AI which wants to do what it's human creator wants it to do is a remarkably hard problem. Especially if that AI is much more intelligent than it's human creator. This is the so called "control problem".
I think one possible vision for that utopia is the Culture - the "post scarcity" part of it is obviously based on fantasy physics but way it portrays humans (OK engineered humans-like beings) human level AIs and godlike Minds running things is quite satisfying. Particularly that the only thing that a godlike AI is likely to worry about is the opinion of other godlike AIs - and fiddling with the impossibly slow, stupid, but still loveable, meat beasties is definitely frowned on (see the reputation of the Grey Area).
Why? I'd genuinely like to know. I've yet to read anything to convince me this is true.
Con: We don't actually understand what intelligence is, except by human example. We don't actually have any idea of how far we have to go to create something similar.
We can draw smooth exponential lines like Kurzweil does and see when the level of complexity of our AI servants reaches the same level as the human brain (~30 years), or we can hope for a concerted human effort as was seen in both the 40's and the 60's. Either way it looks to be in our near future.
The last one, reinforcement learning, is really important because it's a basis for general intelligence. By rewarding/punishing a reinforcement learning agent you can train it to execute very complex tasks, for example navigating a maze with pixel input or controlling a robot body to do some tasks. DeepMind is on the cutting edge of this kind of research, you can read about their latest agent here http://arxiv.org/abs/1602.01783
With such rate of progress it's not inconceivable if a sub-human but general intelligence could be developed in 10-20 years, or even sooner.
Maybe they could apply reinforcement learning to training a chatbot how to direct the flow of a conversation in order to make it interesting. I think it could be huge if only the dialogue was a little better.
Google published some promising results of an RNN chatbot that was trained on movie scripts, but as far as I know it was never released publicly. Microsoft supposedly has an AI chatbot that is really incredible in China. Tay was supposed to be the start of something like that for the English world, but had that stupid scandal.
RNN chatbots are the future, but for now they aren't stable enough to fulfill customers' elevated expectations. Also there is a problem with objective function that rewards interesting conversation, there is some new research on this problem that hasn't made it into open RNN chatbots yet. The standard objective function was giving most probable answers, it turned out to be boring.
I agree with grandparent that Reinforcement Learning is a right approach for training an intellectual chatbot, but there is a problem of well-designed training curriculum (i.e. some systematic scheme of rewarding and punishing the RL agent). I wouldn't wonder if facebook's M team tries different approaches, including RL, to train on their data.
"We are going faster and faster so we will get there" is not the way you calculate the likelihood of reaching a destination within a certain time.
Why not? With incomplete information, you can not say the inverse is true either.
At the end of the day, most of us have no direct control over the AI situation.
"The brain is a Von Neumann machine and our ability to model it tracks with Moore's law" is demonstrably false. There are better models than that.
Still, deep learning has become a very powerful tool and we haven't discovered all possible applications just yet. DL models can learn a huge array of tasks, most of these tasks haven't been tried yet. Maybe there is a way to achieve some limited form of general intelligence even with current state of art (deepmind's A3C agent) if one trained it in some special way nobody haven't thought of yet.
There is also a 3rd possibility that I think is more likely than StrongAI taking control. We get better AI that allow us, humans, to improve ourselves (ie basically hack our generic code and/or augment what we can do with additional hardware). That will lead to humans working and building even stronger AI that will lead to even better humans. Cycle continues until StrongAI and humans are basically the same thing.
Improving humans is the least likely possibility. We are evolutionary spaghetti code constructed out of meat. There are limits to what you can do with biology, and it's always going to be thousands of times slower than silicon. And much more difficult to work with, or reverse engineer.
is it really? from a universal perspective, you want the less intelligent of two agents to get what it wants?
> But if we get it wrong we will just be replaced or destroyed.
children replace their parents, yes.
The child analogy breaks down quickly. My biological children would share 99.99...% of my DNA and most of my cultural values. An AI "child" need not share anything in common with it's creator, and they might totally detest each other.
And lastly my children don't murder me and replace me. I would be pretty upset at such a child, and would probably disown them and try to defend myself. I would have known I screwed up producing them and be sad.
Lastly see Value is Fragile: http://lesswrong.com/lw/y3/value_is_fragile/
>My friend, I have no problem with the thought of a galactic civilization vastly unlike our own... full of strange beings who look nothing like me even in their own imaginations... pursuing pleasures and experiences I can't begin to empathize with... trading in a marketplace of unimaginable goods... allying to pursue incomprehensible objectives... people whose life-stories I could never understand.
>That's what the Future looks like if things go right.
>If the chain of inheritance from human (meta)morals is broken, the Future does not look like this. It does not end up magically, delightfully incomprehensible.
>With very high probability, it ends up looking dull. Pointless. Something whose loss you wouldn't mourn.
If you are not in power, if someone has a boot to your neck, then an AI of the first sort, which can automate the work of removing a boot from your neck, looks very interesting.
I'd count the stuff Google does to bring me relevant search results AI. That improves my life a lot. Automatic sorting of letters by destination is also pretty useful, as are automatic quality checks in factories. If you count optimization as AI, there are many applications, for example in chip design, that improve your day-to-day life a lot.
Some thoughts:
I think we have automated a great deal of our lives and automation is growing exponentially at present. Once we have achieved some critical breakover in automation, I believe we will start to see unification systems. These unification systems will bind the various forms of automation relegating human input to a bygone era. That is, the unification system will learn, optimize, and execute the "complete stack" of various industries... Think product distribution from factory to dock to end user.
I don't think this type of automation is that far off. Is that "AI"? To me it is and I suspect most people outside of HN will call it something along those lines. When some kind of AI like Varuna in the novel _Influx_ by Daniel Suarez hits the streets people will not give one shit if you call it AI or God... they will be very, very afraid of it.
I can name countless cases of "semi-trivial" AI being used to improve people's lives. Tree/graph searches, linear regression and clustering algorithms are used all over the place: in science, engineering and medical field.
However, when it comes to more sophisticated algorithms I am hard pressed to point out strongly positive use cases. I.e. something that actually make a positive impact for large groups of people, isn't a gimmick and does not involves massive privacy violations.
Hm, I guess handwriting recognition is used by postal services. There are a lot of expert systems running behind the scenes in some industries (stuff that involves logistics, for example). Voice recognition helps disabled people to interact with computers. Machine translation can be useful at times. Although I've never been a situation when it made any significant difference in my life or in lives of anyone I know.
Stuff like face detection in cameras or voice recognition in phones and cars seem cool, but nothing would really change if they weren't available.
The most impactful uses of fancy AI that I know involves marketing and surveillance, and I don't consider either of those uses as a positive thing.
I prefer to not even talk about "AI" at all, the word doesn't really point to anything we can talk about in a grounded or meaningful way (I consider it a word used to get people in suits to write checks).
Machine Learning and Operations Research touch pretty much every aspect of our lives, however. Whether it be supply chain optimization, factory design, materials science, etc. etc. etc.
Pharmacies not giving out the wrong medicine.
Doctors diagnosing the right disease in the first place.
It looks like the main problem of modern world is complexity. The infrastructure (physical, social, IT) has become too complex for mere humans to control. Humans are slow, learn slowly, have limited attention spans, limited capability of understanding complex nonlinear systems (e.g. markets, social dynamics). IMHO the world is currently in the state of severe deficit of attention. There is too much problems on all levels, yet humanity as a whole has only so much attention to apply to these problems, so only the absolutely critical issues are solved (e.g. 2008 financial crisis or flint water scandal) while everything else is neglected until it becomes critical.
Machine Intelligence (both Machine Learning and more general AIs, e.g. agents based on unsupervised+reinforcement learning) could give intelligent attention to every big and tiny problem there is: in politics, society, infrastructure, healthcare, etc. on Instead of having to deal with a constant stream of crises AI systems could prevent them from occurring before the catastrophical event took place, just like with preventive medicine. That and automated manufacturing & distribution could give way to a really pleasant life experience.
Of course it sounds too easy to be true. Looks like the biggest obstacle to realizing this variant of utopian future is politics. Human primate is naturally hierarchical, and these instincts influence our society & way of life profoundly. Modern society looks like a pyramid with wealthy & powerful on the top, a docile (but shrinking) middle class in the middle and controlled poor at the base. The people on top and in the middle are quite well off, and they are afraid of losing power (even to machines, even if there will be a mathematical proof that everyone will be better off after we delegate decision making to AI agents).
It looks to me that from the perspective of the rich & powerful there really is no need in automation/AI: why do they need robots when they are used to hiring poor people to serve them? Why do they need automation when with the power of money they can order thousands of people to manufacture any product/experience for them? That's the status quo that will have to be disrupted to have AI/robot-based automated utopia for everyone, IMHO.
It's not going to matter what they need. The technology advances with a rhythm of its own. It has done so for hundreds of years, predictably. Not even big wars were able to put a dent in its pace. Just like Google replaced libraries as our source of information, Napster affected the music industry and social networks affected news, there's going to be an AI transition and it will happen no matter what. It's important to keep access to AI open and not place its power just in the hands of a few.
It will be the continuation of the mobile phone era. Mobile phones with search engines and social networks empowered billions of people, they are their main communication device, some don't even have a TV or laptop, but they still have a phone. All people got to have phones, and all our modern phones are amazing compared to the best phones from a decade ago. What technology has given to everyone not even money could buy a mere 10 years ago. But on the other hand, the phone empowered surveillance and the big brother.
The AIs will amplify this effect even more. They will empower people to a higher degree than phones did, but they will make people easy targets for manipulation. They will allow an extremely intimate knowledge about a person and direct access to influence him/her, to the technologists who control the AI.
But seriously, we've had predictions of "strong AI in 5/10/20 years" for decades now. The best we'll be able to do in such a short timeframe is what we do now: study a problem, create a [complex] solution, code the solution in a programming language. "AI Utopia" will arrive some decades after we've figured out how to make a system that can learn and retain over the long-term. Much like a biological brain.
Jeph Jacques' cartoon take on AI and sentient companions seems plausible: a big central AI grows and evolves and often spins off smaller AIs to be placed in androids. I don't know where the idea originated (it might be his?) but it's a very reasonable fiction for where we'd originate individuals in a high quantity.
(*not to scale)
Our brains are just awful at predicting major breakthroughs...
1. Overestimation of intelligence: Computational complexity theory -- in particular the time/space hierarchy theorems, which, IMO, are the most important theorems in all of computer science (and are basically generalizations of the halting theorem) -- tells us that problems form an infinite hierarchy of difficulty, and that some problems require a certain amount of computational resources regardless of the algorithm used. In particular, we know of plenty of very important problems (e.g. NP-complete, or PSPACE-complete) for which intelligence does not seem to help at all, and those problems are faced in many real-world scenarios. Worse, we know that for many of these problems, there can be no efficient approximate solution employing any algorithm. Artificial intelligence, therefore, cannot possibly form an effective way to solve these very real problems.
2. Underestimation of non-intelligent computation: It is unclear that the "general intelligence algorithms" are the best use of computational resources. Humans (and mammals) are far from being the most successful species on Earth by almost any measure of success (we are neither the most plentiful numerically, in biomass, or even in the most capable of shaping the planet). It is therefore a mistake to believe that the "intelligence algorithms" pose either the most severe threat or the most powerful "problem solving" approach.
Artificial intelligence seems to be particularly useful when interacting with other intelligent beings (i.e. humans). For example, there has recently been some success in employing an "intelligent" algorithm for winning the game of Go. However, it is known that the "Go problem" is EXPTIME-complete, and therefore the intelligent algorithm can only be successful when playing against other players who employ a similar algorithm (namely, humans). It is by no means a general solution to the "Go problem". As "disease" is not a human problem, and pathogens employ computation in a very different way than intelligent beings, it is hard to believe that AI would be able to eradicate disease etc., etc..