Building AI
facebook.com
facebook.com
AI and ML are not synonymous. They are related, and there is a great deal of overlap between them, but the fundamental goals and approaches are largely different. Machine learning is primarily interested in studying what you can learn from data - for some suitable definition of the word "learn". Artificial intelligence is the more broadly defined problem of studying algorithms that exhibit or incorporate intelligent solutions to problems. That may involve data, or background knowledge, or domain assumptions, or a variety of other things. AI is about more than data - even as he observed that it is a sign of intelligence that humans do not require thousands of samples to learn.
The central challenge of AI is not in transitioning from supervised learning to "general" unsupervised learning (whatever that means). The techniques are different, and often used for different things. There is some overlap, and they are clearly related - but it is not at all accurate to conflate unsupervised learning with "common sense". In broad strokes, supervised learning is about identifying features in the data that correspond to labels, while unsupervised learning is about identifying intrinsic features in the data. It sounds like he's interested in bootstrapping supervised techniques, or perhaps transfer learning.
My first response: ugh...yet another famous and lauded geek icon wading into the subject of AI. At least this doesn't seem to be in any way connected to all the Nick Bostrom garbage. But if I'm allowed some snark, perhaps if he wanted to be an AI expert he should have stayed in school... ;-)
Zuckerberg isn't afraid of AI because he already has more money than he will never need (or thousands of other people would need) in a lifetime. However, when AI gets good enough, it will make many jobs obsolete and they will not be replaced at a fast enough rate.
I use simple forms of automation at my own company. Instead of hiring 3 or 4 workers, I write software. I can't imagine how many jobs will be replaced when AI gets to the point of near-human levels of learning.
[0]https://en.wikipedia.org/wiki/Technological_unemployment#His...
Sure, the best of us will be fine, but that's always been the case. What about the ones who are struggling to remain relevant today? They'll get left behind unless something's done-- I'm not sure that retraining will be enough, not that there's any political movement in that direction.
It'll be a while before anyone is going to trust a load of valuable goods with an AI. There are security/theft issues, edgecase AI failure issues, loading unloading, etc.
We've seen entire industries leave the USA. Probably within your lifetime.
That said this can have devastating effects on economies. A lot of cities in the rust belt have felt this impact.
But the opposite can happen. One of the huge reasons US companies offshore production is because of labor costs. If you reduce the man-hours / product, it makes sense to open back up domestically.
Last paragraph in the wikipedia section:
"Research by the Oxford Martin School showed that employees engaged in "tasks following well-defined procedures that can easily be performed by sophisticated algorithms" are at risk of displacement. The study, published in 2013, shows that automation can affect both skilled and unskilled work and both high and low-paying occupations; however, low-paid physical occupations are most at risk.[7] However, according to a study published in McKinsey Quarterly[63] in 2015 the impact of computerization in most cases is not replacement of employees but automation of portions of the tasks they perform"
Oxford or McKinsey Quarterly, which source do you prefer?
Large swaths of the population unemployable though... That's a real issue.
A living wage is the only good solution I can think of. Alternatively, pay people to dig and fill holes, but it's expensive and kind of pointless to make up unskilled jobs for people to do.
I do wonder how much a living wage would cost in the US though. Shift some money from defense? There's a good job for AI: balance a budget based on pre-defined policy goals.
IMO the one job that is most threatened by future progress in AI is the job of software developer.
As for the menial jobs, of course nobody cares whether a human or a machine mows your grass. But I do think that most people would rather be nursed by human being than a robot. I think there will be a big increase in care-giving jobs, which would be wonderful because the world is in dire need of care-givers.
http://observer.com/2015/08/stephen-hawking-elon-musk-and-bi...
Which makes sense. The opinion of everyone I know that has actually worked with machine learning and state-of-the-art AI (as in, written code, ran experiments, etc.) is that the idea of today's AI suddenly waking up and taking over the world is utterly ridiculous, or at least not something we should be afraid of for the foreseeable future.
You just beautifully summarized the "Lump of Work fallacy" https://en.wikipedia.org/wiki/Lump_of_labour_fallacy
Yes, technology changes/improvements causes transitions in the job market but as it turns out, human needs are endless and new jobs are created for displaced workers.
The quote (from https://www.marxists.org/archive/marx/works/1865/value-price...):
"...one might infer, as Adam Smith, in whose days modern industry was still in its infancy, did infer, that the accelerated accumulation of capital must turn the balance in favour of the working man, by securing a growing demand for his labour. From this same standpoint many contemporary writers have wondered that English capital having grown in that last twenty years so much quicker than English population, wages should not have been more enhanced. But simultaneously with the progress of accumulation there takes place a progressive change in the composition of capital. That part of the aggregate capital which consists of fixed capital, machinery, raw materials, means of production in all possible forms, progressively increases as compared with the other part of capital, which is laid out in wages or in the purchase of labour. This law has been stated in a more or less accurate manner by Mr. Barton, Ricardo, Sismondi, Professor Richard Jones, Professor Ramsey, Cherbuilliez, and others."
(I personally am not a huge fan of Marx's writings, mainly because of the strong contempt for humanity that I feel emanating from them... incidentally I've recently stumbled upon this one, though.)
A scary analogy in favor of this argument is, a horse cannot economically perform almost any job today, hence most were slaughtered though few were kept for amusement. That happened after horses had a very good run where every improvement in technology increased "horse employment." If a machine exist to outperform a human in any conceivable job, it's not very different (and then you get to the discussion of the political options.)
A counterargument is that up until now this argument was wrong many times, and that we're very far from an AGI, or even from machines beating humans at every possible task.
This time it's the entire human niche that will (eventually) be automated - all cognitive tasks as well as all physical ones. That's the difference and why extrapolating from the past doesn't work for AGI.
Again, automation isn't new, it just seems like it is because we're living it now.
With agricultural automation, the 30% change was filled mostly with new service jobs, not labor jobs. The same goes with A.I. automation - suddenly service jobs that were not possible before become possible. Jobs are not a zero-sum game.
Edit: I see superintelligent AI as not mere automation, but in effect creating an infinite supply of new labor. Humans are diluted out.
Also, the idea that this super intelligent A.I. can do everything is almost a god of the gaps fallacy. There are still many things that humans can do that a computer cannot.
"If the Luddite fallacy were true we would all be out of work because productivity has been increasing for two centuries"
More and more people are unemployed every day, while more and more workers get replaced.
The social welfare system in Germany started in the late 1800s when industrialization and internal combustion engines started to replace people.
Nowadays a country can be considered to have a "good" job market when only 5 to 10 percent are unemployed, while other countries, even in the first world, are at up-to 25%.
This all also extremely depresses wages, because the workers are competing against each other on the job market.
Again, if you look at automation over the centuries, it doesn't destroy jobs, it leads to people switching jobs.
Or at least that could be the case. It's called post scarcity, and it should be something we strive for.
If the machines are privately owned, and you have to pay for the products, you’ll have to work again, too.
At that point maybe we should consider having a UN Charta that tries to make sure the nordic model is applied on a global scale before we allow fully automated systems to replaces whole industries.
-- Greg Clark, A Farewell to Alms
The "Lump of Work fallacy" does seem to be a fallacy. However, that didn't help the horse. Why is that? The invention of "Artificial General Muscles" in the form of the internal combustion engine was a better-than-perfect replacement for all forms of horse labour. We should expect artificial general intelligence technology to do much the same to humans. There is no law in economics that prevents human wages from declining below subsistence. And a perfect substitute for human labour that requires no wages and is more intelligent would certainly vastly lower the value of human labour.
Oh wait. They are fine. They're just not used as constant slave labor all the time anymore, much like human labor.
The fallacy claims that the number of jobs is a zero-sum game. I'm not saying this at all. While more jobs will be created, because automation is happening at such an exponential rate, there won't be enough jobs replaced.
I have feeling when we do have AI, many of these economic principals will need to be re-written.
More from this article:
"Whereas some argue that immigrants displace domestic workers, others believe this to be a fallacy, arguing that such a view relies on a belief that the number of jobs in the economy is fixed, whereas in reality immigration increases the size of the economy, thus creating more jobs"
Immigration may increase the size of the economy, but if they are hired at let's say 30% of an American working at the same rate, it essentially dilutes the market. Americans now have to compete with someone making much less than them, which will mean a huge pay cut.
With AI, it will mean competing with algorithms/hardware that can probably be purchased for even less than the cost of a worker.
The end game for this will be a class of people owning all of this AI hardware, the poor, and a much smaller middle class of people in-between that have the skills to work on the AI hardware.
What happens to all of the people that aren't educated or don't have the skills in this new economy when all of the jobs they could do to make money have been replaced by AI?
What happens now to people who have no education or skills relevant to the economy? AI won't make this new problem, it's something we already deal with.
Right now, they can get a minimum wage job and survive. My point is that AI could completely replace these and the only way you will be able to get a job is to have some form of education or skill.
AI tends not to work like that. High-skill jobs tend to be easier for machine-learning than low-tech jobs. https://en.wikipedia.org/wiki/Moravec%27s_paradox
Taken to a limit, a sector-wide displacement could have quite severe repercussions for a society. Imagine for example a majority of truck drivers* losing their jobs once self-driving cargo vehicles become available. Assuming they'll be adopted rapidly (for their advantages in safety, fleet efficiency, and reduced transit time due to continuous operation) and the change is quite sudden. Those truckers are now an immediate, acute problem for the country: unemployed, not spending, not earning, heavily invested in training and experience for a now non-existent job, and not (in aggregate) well-suited for other work. They are also a large, angry, nationwide group - a pre-fabricated voting block - ripe for political exploitation.
*I picked this example based on that map of common occupations that was floating round here a few months ago. By far the most popular job by state was "truck driver", so this is potentially a real issue.
The general idea that technology deprecates people is true. It is also true that there is an escape valve in new needs, requiring new jobs, that allows for us to shift occupations and remain employed.
However, there is no law that states that the escape valve will always work, nor that it can work at high rates of job displacement.
Moreover, there are signs of low aggregate demand in developed economies. These signs are appearing now even in developing economies. This means that money is not percolating down into the hands of people, who can't then spend and create demand. There is a huge number of factors at play here (see the note about being a complex problem). Higher worker efficiency causing higher unemployment rates is one such factor. It can't be dismissed outright.
We may be witnessing Marx being right (albeit ahead of its time).
Why won't they?
I have hopes for a more utopian future versus one where most are homeless because their jobs get replaced. The transition will be painful.
There's also the mentality of clinging to jobs because it's needed for livelihood. Wanting to keep jobs that will bring 0 value to society. Think about all the crappy toys we make that's thrown out after a month or two. Who would want to work these jobs except if it's needed to sustain a decent quality of life outside of work?
That said, it would be really nice to cure diseases that are particularly painful or kill young people (cancer, etc).
Argh, these days, AI is like NoSQL a few years ago. The magical buzzword taken to an extreme level.
When do you think it will happen, the point at which AI will overtake most of the jobs done by humans? By the time it happens, the society will change significantly. Significantly enough so that it's probably pointless to guess. Will the society adapt to the change? Well, it has adapted more or less successfully to the changes so far.
Even when it happens, there will always be human factor. Humans are social, they need interaction with other humans.
We should probably slow down with all the 'AI will take everything away!' drama (which makes nice headlines which contributed significantly to this buzz, thanks a lot). We should remind ourselves, that we don't even have self-driving cars in mainstream (or, in software terms, 'production').
We are talking about a massive leap here, required in both society to accept the change and the technology to make it feasible beyond cool demos. By the time this happens, who knows. Maybe half of the humanity will go extinct because of WW3?
I get the hacker mentality, that we should try to change the world and everything. But sometimes, it's good to stop for a moment and do a reality check.
ftfy
Let me tell you there is no point in resisting the inevitable. In fact if anything preparing for it makes it easier when it actually happens.
Have found this free book that I saw on YCs 2015 reading list particularly helpful in this respect: http://neuralnetworksanddeeplearning.com/
When I started playing Go, it took me about 4 months to beat the stronger available computer AI. Today, the strongest computers would be a challenge to play against evenly.
However, even after all the patterns and montecarlo solutions, they still fumble the initial stage of the game. As a Go player, I look forward to AI's that beat humans in that stage of the game, because at that point, we will be able to learn from them.
Until then, their victories are purely computational, and they are not even interesting to play with.
Only recently researchers have started to expose chess to reinforcement learning (i.e. a true AI engine that learns chess on its own and beats humans). But the existing commercial engines that beat humans (like stockfish) are anything but AI.
Historically, when you play with a bot you sense relentlessness(instead of calmness), and sometimes desperation(computers sometimes go wacky when they see they are losing).
What will we sense when they are even or stronger in the more artistic and strategic phases of the game?
It isn't reinforcement learning though. That isn't state of the art (at the moment), instead modern deep NN's are learned by gradient descent. This can be made to work really well but isn't regarded as similar to how biological neural networks work.
The idea of General Artificial Intelligence is something that can learn anything you throw by adapting and reinforcing base assumptions.
There's a podcast called Talking Machines that really dives into a lot of this stuff. He's freaking great and pretty darn entertaining.
http://www.wired.com/2016/01/in-a-huge-breakthrough-googles-...
[0] http://www.amazon.com/Superintelligence-Dangers-Strategies-N...
The announcement is slightly disappointing, in the sense that face and voice recognition are fairly solved and there is code on github that already allows you to achieve this. I was hoping for more, like at least an AI engine that can scan his shirts and recommends what to wear but I guess he does not have that problem.
1-star life long project on GAI
2-....
3-Invent GAI
4-Use general AI to extend human life for centuries, advance human knowledge like never before, In secret. I can easily see somebody putting a bullet through my head to steal my AI.
5-Decide what to do next. Decide If I want to share.
I'm not joking. I actually believe I have a chance.
Why I believe I will be successful:
1-I'm like a dog with a bone.
2-Once I'm working on a project I don't let go.
3-I've worked in years long projects before.
4-I've been thinking about GAI for a long time.
5-Recently I just came up with an approach that may lead to GAI. It is no like the current approaches in existence. I think my approach is much better.
6- There is nothing more interesting to work on then GAI.
7-I like the challenge of beating the best minds in the world. Time will tell.
You'd think we'd collaborate!
Maybe fb will go the Alphabet route?
He is more curious about the ability of using AI for daily interactions (hence the reference to what are movie characters): do notions like personality, convenience and interface, implicit assumptions matter? If he asks “Buy a new set t-shirt” will the robot be able to joke “which colour?” (MZ notoriously only wear the same grey model).
I would look into what Messenger wants to be doing (or, what Slack is doing) to see the implications: can small companies do what Uber does with the key-word integration in chat? Is it creepy, hackable, tedious? This of this like his dog’s Fb page: it’s building empathy for the users of the Facebook platform.
First problem he’ll probably have to find: Who to call when saying “Call Mike!”?
Of course “Mike” is probably the Mike with whom he has the closest ties (based on Facebook graph data) — except his VP of Engineering, Mike Schroepfer, is known as “Shrep”, so that’s probably not him: how easy is it to program something that can learn that seamlessly? In which case the ambiguity too high, and Jarvis should ask for confirmation vs. just casually saying the full name out loud and wait a second? How to measure ambiguity? Would it be simpler to just start calling people with non-ambiguous name? Would he feel like sharing his code with peers? His heuristics? How to gather that “Mike” is still in Europe where it’s 3AM, vs. was in Europe for in last post and actually just landed?
"My personal challenge for 2016 is to build a simple AI -- like Jarvis from Iron Man -- to help run my home and help me with work.
I'm planning on writing up some thoughts every month on what I've built and what I'm learning. I'm still early in coding, so I'll start this month with a summary of the state of the AI field.
Artificial intelligence may seem like something out of science fiction, but most of us already use tools and services every day that rely on AI. When you do a voice search on your phone, put a check into an ATM, or use a fitness tracker to count your steps, you're using basic forms of pattern recognition and artificial intelligence. More sophisticated AI systems can already diagnose diseases, drive cars and search the skies for planets better than people. This is why AI is such an exciting field -- it opens up so many new possibilities for enhancing humanity's capabilities.
So what can AI do and what are its limits? What things is AI good at and what is AI bad at? Simply put, today's AI is good at recognizing patterns and bad at what we would call "common sense".
The primary method used to train AI systems is called supervised learning. This is like when you show a picture book to a child and tell them the names of everything they see. If you show an AI thousands of pictures of dogs, you can train it to start recognizing dogs.
You can teach AIs to do a lot of things this way. For example, we can teach an AI to recognize all of your friends' faces by showing it thousands of photos, and then it can suggest tags for the photos you upload on Facebook. You can teach an AI to recognize speech by having it listen to thousands of hours of speeches throughout history while also showing it transcriptions of what was said. You can teach an AI to diagnose melanoma by showing it thousands of photos of tumors. You can even teach an AI how to drive a car and automatically brake by showing it thousands of examples of people and obstacles it might encounter on the road.
Diagnosing cancer, driving cars, transcribing speech, playing games and tagging photos may sound like very different tasks, but they're all examples of teaching an AI to recognize patterns by showing them many examples.
Many different problems can be reduced to pattern recognition tasks that sophisticated AIs can then solve. This year, I'll teach my simple AI to recognize patterns. I'll train it to recognize my voice so I can control my home through speaking. I'll train it to recognize my face so it can open the door when I'm approaching, and so on.
But there are lots of limitations of this approach. For one, to teach a person something new, you typically don't need to tell them about it thousands of times. So the state of the art in AI is still much slower than how we learn.
But more importantly, pattern recognition is very different from common sense -- and nobody knows how to teach an AI that yet.
Without common sense, AI systems can't use knowledge they've learned in one area and easily apply it to another situation. This means they can't effectively react to new problems or situations they haven't seen before, which is so much of we all do everyday and what we call intelligence.
Our best guess at how to teach an AI common sense is through a method called unsupervised learning. My example of supervised learning above was showing a picture book to a child and telling them the names of everything they see. Unsupervised learning would be giving them a book and letting them figure out what to do with it. They could pick it up and by touching it learn to turn the pages. Or they could let go of it and realize it falls to the ground.
Unsupervised learning is learning how the world works by observing and trying things out rather than being told what to do. This is how most animals learn. It's key to building systems with human-like common sense because it doesn't require a person to teach it everything they know. It gives the machine the ability to anticipate what may happen in the future and predict the effect of an action. It could help us build machines that can hold conversations or plan complex sequences of actions -- necessary components for any authentic Jarvis.
Unsupervised learning is a long term focus of our AI research team at Facebook, and it remains an important challenge for the whole AI research community.
Since no one understands how general unsupervised learning actually works, we're quite a ways off from building the general AIs you see in movies. Some people claim this is just a matter of getting more computing power -- and that as Moore's law continues and computing becomes cheaper we'll naturally have AIs that surpass human intelligence. This is incorrect. We fundamentally do not understand how general learning works. This is an unsolved problem -- maybe the most important problem of this century or even millennium. Until we solve this problem, throwing all the machine power in the world at it cannot create an AI that can do everything a person can.
We should not be afraid of AI. Instead, we should hope for the amazing amount of good it will do in the world. It will saves lives by diagnosing diseases and driving us around more safely. It will enable breakthroughs by helping us find new planets and understand Earth's climate. It will help in areas we haven't even thought of today.
Jarvis is still a long way off, and we’re not going to solve most of these engineering challenges in the next year. But I'm glad to be joining the effort and doing what I can to push the field of AI forward."
Now I know how to get into Zuck's house. Just make a paper cutout of his face. Or, maybe the AI will develop the common sense not to allow this as a verification method.
People here seem to think it is just a question of time before we do. At some point we will have an artificial intelligence.
But what if this isn't possible? Could it be that AI requires so complex algorithms that we humans can't understand them because our brains are too simple? Not all things can be simplified; maybe creating an AI is so inherently complex that the only way to create one would be by chance, which is how evolution did it; maybe it just isn't possible to create it by "intelligent design"
> But there are lots of limitations of this approach. For one, to teach a person something new, you typically don't need to tell them about it thousands of times. So the state of the art in AI is still much slower than how we learn.
Unsupervised learning is not necessarily the answer. The computer would have to be given plenty of spare time to learn random patterns from the universe and have enough inteligence to apply these to a problem. What would be better is if we could tell the AI once and have it figure things out. When you start a new job that's what happens, someone tells you the instructions. (They may or may not show you as well, and may or may not stick around to correct your mistakes.) This requires interpreting the instructions into rules and then attempting to apply them, learning from mistakes, and then evolving that into rules that work, then after enough examples of success, evolving into patterns recognition that makes it more automatic.
Human expertise can be broken into three levels, the first is strategic planning and takes a lot of mental effort, then there are rules-based responses which are faster, then there is the muscle-memory-like automatic responses. Right now it seems we either manually program in all the rules or else we use 1000s of examples to build up the automatic level, but we don't have the strategic level for the AI to build its own rules or the level of using its rules to learn from examples over time. (Though I am not well enough versed in AI to know for sure that we don't have pieces of those solutions.)
It would also be nice for the AI to be able to take patterns it has learned and articulate them as rules which someone else could learn.
Reinforcement Learning is an approach to learn rules by observing a positive or negative feedback from the world. Recently a single algorithm learned 40 Atari games on its own and a Go playing algo beat the European champion. They both used RL.
Is this really the case? I thought the field had a pretty good handle on the theoretical foundations of unsupervised learning. Can anyone confirm what he's asserting here?
> Some people claim this is just a matter of getting more computing power -- and that as Moore's law continues and computing becomes cheaper we'll naturally have AIs that surpass human intelligence.
And this is happening, it's just not in a general sense because the general case of human intelligence is enormously complex, a composite of all the simple cases that improvements in computing power and models are just beginning to master.
> This is incorrect. We fundamentally do not understand how general learning works. This is an unsolved problem -- maybe the most important problem of this century or even millennium. Until we solve this problem, throwing all the machine power in the world at it cannot create an AI that can do everything a person can.
Again, can anyone fact check this? Seems a bit overstated.
I think this is disproved by approximating AIXI: https://en.wikipedia.org/wiki/AIXI
"All tasks which require intelligence to be solved can naturally be formulated as a maximization of some expected utility in the framework of agents. We gave a functional (2) and an iterative (9) formulation of such a decision theoretic agent, which is general enough to cover all AI problem classes, as has been demonstrated by several examples. The main remaining problem is the unknown prior probability distribution AI of the environment(s).
<...> the universal semimeasure, based on ideas from algorithmic information theory, solves the problem of the unknown prior distribution for induction problems. No explicit learning procedure is necessary ... . We unified the theory of universal sequence prediction with the decision theoretic agent by replacing the unknown true prior AI by an appropriately generalized universal semimeasure ξAI. We gave strong arguments that the resulting AIξ model is the most intelligent, parameterless and environmental/application independent model possible.
The major drawback of the AIξ model is that it is uncomputable, or more precisely, only asymptotically computable, which makes an implementation impossible. To overcome this problem, we cons tructed a modified model AI, which is still effectively more intelligent than any other time and space bounded algorithm."
The copy/paste lost some information, especially symbols, so take a look at the paper for a better read.
In the meantime, if you have time, could you elaborate on: "All tasks which require intelligence to be solved can naturally be formulated as a maximization of some expected utility in the framework of agents."
Is this proven in the paper, or assumed true outside of it?
This statement is a contradiction in terms. If we do not understand how something works, it is impossible to estimate how far off it is.
You can only estimate the distance of discovery when it is inherently incremental, but since we really have no idea how to do general unsupervised learning, it's entirely possible that it consists of a single, brilliant algorithmic insight. Nobody could have estimated how far off the airplane was before the Wright brothers invented it, similarly with cars, or for a more recent and concrete example, a quasipolynomial time algorithm for graph isomorphism.
With general unsupervised learning we can't even clearly describe the goal we're trying to reach or define it in objective terms.
I can't really imagine how they could be more similar.
And note, they're not doing it to make everyone happy. They're doing for the reasons you might surmise. Your instincts to be concerned are correct.
A lot of jobs don't require much intelligence, but some. Like cashiers, warehouse workers, machine operators, etc.
That kind of automation will benefit "The ultra wealthy".
Since no one understands how general unsupervised learning actually works, we're quite a ways off from building the general AIs you see in movies. Some people claim this is just a matter of getting more computing power -- and that as Moore's law continues and computing becomes cheaper we'll naturally have AIs that surpass human intelligence. This is incorrect. We fundamentally do not understand how general learning works. This is an unsolved problem -- maybe the most important problem of this century or even millennium. Until we solve this problem, throwing all the machine power in the world at it cannot create an AI that can do everything a person can.
This circuit element is basically a synapse (I simplified this a LOT, fyi). We have to implement memristors better into the existing chip architecture. As is, they are just really low power static memory systems.
Also, that your mind is different than your body is, well, just false. These AIs need better input devices to 'learn' from, just we learn too. Mobility is key, you can then just bump about and make mistakes, the keys to learning.