Slashdot Mirror
'Modern AI is Good at a Few Things But Bad at Everything Else' (wired.com)
Jason Pontin, writing for Wired: Sundar Pichai, the chief executive of Google, has said that AI "is more profound than ... electricity or fire." Andrew Ng, who founded Google Brain and now invests in AI startups, wrote that "If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future." Their enthusiasm is pardonable.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
[...] But there are many things that people can do quickly that smart machines cannot.
Yet. Powered flight used to be considered a pipe dream too.
Keep that in mind.
...before a bunch of angry old coots post telling us that none of this is AI.
Nothing can be good at everything. That is why there are differnt things. Even something as basic as a hammer has multiple versions. Even if you go and look at just nailing, there are several types that are good for one, but bad for another.
Ut zould be news if it where NOT thee case. This "news" is "dog bites man" please come back when it is "man bites dog".
Don't fight for your country, if your country does not fight for you.
I feel we're at the same crossroads as hacking/cracking. They hope if they use the wrong word for long enough people will forget there is a correct one.
AI getting into the trough (https://en.wikipedia.org/wiki/Hype_cycle) again (https://en.wikipedia.org/wiki/AI_winter)?
Prominent people seem to fear AI (http://time.com/3614349/artificial-intelligence-singularity-stephen-hawking-elon-musk/), but isn't this just Fear of the Unknown? I mean, Elon and Stephen are really smart people, but do they know that most NN:s come down to linear algegra and spiced with non-linearities in the end, just simulating neurons? I mean neurons are common-place on the planet already, equipped with malice and stuff...
Sadly AI isn't really intelligence. Its knowledge based on what is in its data base. Hardly any of it is considered really deep self learning other then some programmed learning abilities. But then again, people get excited about Space X launching a rocket that was done back in the 60's and 70's? Will AI hit a brick wall as well?
We don't have AI, in any form, in the modern world. We have code which solves program similar to a neural network and we have code which can mutate within very strict limits with genetic algorithms. We have nothing even approaching "artificial intelligence," which at the very minimum of the bar would be the level of an "intelligent" Human. If it's not better than a Human with an IQ of no less than 135 at literally everything it's not AI. We have nothing remotely close to equal to an actually retarded Human with an IQ of 70.
A much bigger problem is that if an AI were capable of every single thing he's whining about, then rather than an "AI", it would be a "slave".
"If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future."
Perhaps AI is already generating the majority of Slashdot posts these days.
Not a lot of people are interested in automating "ordinary human activities" but they are very interested in automating very specialized activities. These activities include assembling objects, inspecting objects, moving a vehicle loaded with goods to a destination and estimating risk in the stock market. These aren't ordinary human activities but they'll put half the country out of work.
Anons need not reply. Questions end with a question mark.
And we will see more and more things humans do replaced by AI/machine learning/automation. Especially tasks with well-defined rule sets. Low skill labor is still going to be at risk of being automated away, especially as sensors and robotics continue to improve as well.
Deep learning optimizes coefficients of long equations. It's just math.
Exactly.
Plus, "we" often fall into the flawed thinking that AI/robots/automated systems/whatever-you-want-to-call-it has to figure out how to do activities as humans do them. Much more likely, I see a world where we adapt the way work is done to better meet the strengths of these automated systems. You see this, for example, in food production. Instead of inventing processing machines that can deal with all the variation in "natural" vegetables, processors started demanding that farmers grow vegetables that are relatively uniform in shape and size and don't buy the vegetables that don't fit the tolerances of their machines. In many cases it's much easier to control the structure of work inputs than to develop a "work machine" that can deal with all sorts of edge cases.
We don't have AI, in any form, in the modern world.
Not true at all unless you are narrowing the definition of AI to such a narrow degree as to make it effectively meaningless.
We have nothing even approaching "artificial intelligence," which at the very minimum of the bar would be the level of an "intelligent" Human
Nonsense. Dogs do not as a general proposition approach human level intelligence. Yet do have real and measurable intelligence. A computer with the intelligence of a dog could very fairly be described as intelligent. AI does not have to pass human intellect be classified as intelligence or to be useful.
The argument that AI isn't even close, or isn't here, is just plain stupid. It won't take "perfect" or "true" AI to replace an imperfect prone-to-error human in a job. We're being blinded by the need for perfection when it will only take "good-enough" AI to start replacing human workers.
Even worrying about the problem of AI is rather stupid when the problem of automation is the more immediate issue staring the economy in the face. We're working quickly to replace cashiers, warehouse and assembly line workers, and soon we will be replacing drivers. Just targeting these jobs will make millions of people unemployable. And don't try and regurgitate that age-old mantra of go-get-an-education either. Not every human is capable of being re-trained for a more advanced skill, and we have a hell of a lot more humans on the planet to employ with this next evolution of job decimation. And when you start thinking about the types of jobs you held in order to get an education, you quickly realize that automation is looking to remove the bottom half of the ladder of success. Rather hard to climb that proverbial ladder when the first rung is 12 fucking feet in the air, and you're competing with a few million people.
Our economy is going to feel this pain well before we start having to worry about any shitty form of AI.
Spot on!
Starships were meant to fly, Hands up and touch the sky - Nicky Minaj
Not a lot of people are interested in automating "ordinary human activities" but they are very interested in automating very specialized activities.
People are VERY interested in automating "ordinary human activities" but automation != AI outside of very specialized niches. A dishwashing machine is automation of an ordinary human activity but it is decidedly not AI. It's not clear what you actually mean by "ordinary human activities" but humans have been automating those since there were humans.
These activities include assembling objects, inspecting objects, moving a vehicle loaded with goods to a destination and estimating risk in the stock market. These aren't ordinary human activities but they'll put half the country out of work.
No it would decidedly not put half the country out of work. First off actually assembling objects does not require the device to be intelligent in the sense of AI. Automation != AI and the two concepts are orthogonal for purposes of most assembly work. Second, the percentage of people doing assembly work in manufacturing in the US is no where close to half the work force. 10-20% tops. I am GM for a company that does this sort of work. Third, automation is EXPENSIVE and you have to do a relatively high volume of work to justify the expense of automation over people. That's not going to change any time soon.
AI isn't going to put transport workers out of jobs anytime soon either. When goods get delivered exactly how do you think they are going to get loaded and unloaded from the truck? Even if the worker isn't driving they aren't going to send goods to their destination without someone to watch over them and facilitate the transaction any time soon. It's not as if UPS is going to be able to magically make packages appear on your door step by teleporting them from the truck.
Estimating risk in the stock market? No that won't take humans out of the loop any time soon either.
+1. This is algorithms and infant ML.
I can take my kid and train him to swim and then train him to drive a car and get rudimentary skill in a week in both.
You can only do this after about six years of full-time learning in how to navigate in the real world and how to operate his body. This is the hard part, the part that humans learn in their first six years and AIs don't: dealing with the external world.
Learning to swim and learning to drive a car are easy; machines can do that. Learning to make a peanut-butter-and-jelly sandwich out of what is in the refrigerator: now that's hard.
Low skill labor is still going to be at risk of being automated away, especially as sensors and robotics continue to improve as well.
Probably not to the degree you imply. The reason is simple economics. Automation is in most cases expensive and if you actually do the financial analysis (which I do for a living FYI) you'll find that it's nearly impossible to automate most jobs to such a degree that low skill labor becomes unnecessary. Automation is used in high volume or high content value or high risk jobs. While automation has gotten and will continue to get cheaper, it's unlikely to reach such a low price point that it pushes people out of the work force entirely within the lifetime of anyone reading this. To do that you would have to have near human level intelligence AI that you can sell for less money per unit than a human costs. That is a FAR more difficult goal to reach than most people realize. People are flexible and for low production volumes or ill defined tasks rather inexpensive.
That describes me. Perhaps I am an AI.
I'm a good cook. I'm a fantastic eater. - Steven Brust
...before a bunch of angry old coots post telling us that none of this is AI.
Let's put some of that into context.
A 5-year old can recognize a dog in an image in about 1/2 a second. A neuron takes about 0.05 seconds to activate and fire, so on average the entire recognition process takes about 10 steps.
Those steps include reading the image (sensing and converting the image data to internal form), and activating the physical response: saying "dog" or clicking the right button or whatever.
So let me ask this: what AI algorithm takes ten *steps* to recognize something as complicated as a dog?
Note that this works with dogs partially obscured (half masked by a tree, for instance), any size, rotated, from any angle (top down, face on, from the side), any breed (dalmations and chihuahuas), toys made to look like dogs, and cartoon dogs.
The algorithm does this at a very high level of accuracy, and can tell dogs apart from other animals with similar features: cats, opossums, and so on.
And the algorithm does this without a zillion training examples. A typical 5-year old has seen far fewer dogs than the Tensor Flow algorithm training set.
So tell me again: in what measure is our current level of AI anywhere close to being "real" AI?
I think part of the problem is a lot of folks don't really understand how current AI technology which hasn't changed in decades works compared to how our minds work things out. Recall that there was a recent AI project to find the meaning of the Internet and the answer it came up with was "Cats" because they seem to appear far more often then any other topic on the Internet. That is a mathematical mean or average, the optimal answer but ask any normal person and cats won't be the answer that they give you.
And there in lies one of the biggest problems with our current AI, it's only able to do things that we ask it and they need a clear solution. You can't exactly ask an AI, "do you think this person lived a happy life?". What makes this question bizarre is we're not entirely sure what the answer to that is ourselves but in most cases we can usually tell.
30 years ago when I was taking AI in college my professor summed this up perfectly. He said "AI is like artificial milk. Artificial milk doesn't have to be as good as real milk, it's just quite handy if you haven't got any real milk."
But that whole list very specifically requires a proper, fully functioning visual system.
Visual systems are more than one bit of the brain.
A fully functioning visual system requires memory, motor, spatial and so on. It's incredibly complex.
Absolutely NO ML algorithm out there is even close to how those work.
Every single ML algorithm out there that is based on visual analysis is brute force as fuck.
Then there is also the fact that the data-set isn't actively optimized.
Then there is also the fact that as they learn, some algorithms can fail horribly.
Just imagine that. Just imagine your auto-driving car suddenly decides a wall is roadway. They've been tricked by stickers before. Literally black tape in fact.
Now just imagine that was a cloud-connected ML algorithm that decided this. That's a million+ potential accidents in minutes of it happening.
There is ordinary human behaviour and the very foundation of human behaviour.
Sight is tied in to much of how we work. This is why it is such a shock to the system when blindness occurs.
So much of our brain is allocated resources for analysis of our vision. Both unconscious and conscious analysis.
We're nowhere near ready for any useful, consumer-ready ML. Nowhere close.
It's a silly gimmick whose only use is in attaching Nicholas Cage on to every actor in films or making celebrity porn.
Current, modern ML is Dotcom, Bitcoin and Tulips all over again. It's just people making big bucks off saps and hoping shit goes somewhere.
It might, but not any time soon. Think gaming industry pre-Atari collapse compared to post-crash. Yeah, that needs to happen first.
Can computers (controlling robots) learn through observation? Can any computer learn to understand natural language, even after a decade of conversation with humans? Infants understand by 2, and are talking soon after.
Computers have no cognition. All they can do is run programs written by humans, or programs derived from programs written by humans (genetic programming).
To date there is nothing out there where a compute can do a fraction of what a 6 year old can do.
On top of that there is no AI that I could drop into a car then drop into a pool in a human body and learn how to drive and swim.
So you've proven my point. All we have a programs and programs than can look at large sets of data to appear smart.
There is nothing even close to AI out there.
What does this even mean? I can't figure it out, but I wonder if NLP can...
Learning to make a peanut-butter-and-jelly sandwich out of what is in the refrigerator: now that's hard.
It's also sexual harassment if you ask a woman to do that. http://domainincite.com/20201-...
The shepherds did so well protecting the flock that the sheep no longer believed that wolves existed.
There is a giant bubble of confidence in tech night now because true AI is coming year / in five years / in fifty years, and Pichai has to go and spoil everything!
I guess this is a signal that Google / Alphabet is done gobbling up small startups. They are going into data mining mode, building profiles of every human on the planet so they can more efficiently control our lives.
Are you familiar with the Chinese Room argument?
Use any definition other than "pattern recognition" and we don't have it.
So me any form of intelligence that isn't some form of pattern recognition. Hell the entire field of physics and every other science is simply the act of observing patterns and building a model to describe them that has predictive value. At it's most basic form that is just sophisticated pattern recognition.
Our brains have many different parts each with thier own function(s). It's pretty apparent humans have many simple algorithms running simultaneously, in addition to whatever else happens. So the obvious conclusion that will surprise no one is that a true AI like a human would just have deep learning (or a number of deep learning modules) as a single component among thousands that would be required to get the emergent behavior that is strong AI. The technique is simple, easy to implement, and accomplishes certain simple tasks with ease so it's usefulness as a tool when designing strong AI won't ever go away. There is no magic in our molecules, there is nothing magic about strong AI, it should be obvious a single algorithm won't ever make a strong AI, we and other animals prove it can be done it's just difficult. People need to get over it and realize that when millions of people work together, contributing our results and pooling our knowledge, we can make much of science fiction a reality (hopefully just the better parts).
deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
Exactly, precisely this. They can't 'think', and never will. The approach being used is completely wrong, or at least incomplete, because we don't even have a clue how we are capable of 'thinking'.
Bearing in mind that, ever from the 60s, the AI community has come up, time and again, with exuberant forecasts that never came to pass, it is interesting that some keep issuing equally exuberant forecasts. A human brain emulation by 2020? The Singularity by 2030? Chances are the AI for the foreseeable will be more of the same: more and more systems that a excel at very, very narrow fields.
The current AI algorithms seem quite good at automatically solving some well defined problem, given enough input and learning cycles.
Is it really such a stretch of the imagination that we will come up with a more automated way, a sort of meta AI, which will detect if a given problem hasn't been defined yet, and how it may be defined, in order to afterwards pass it on to the learning-through-simulation subsystem?
One DeepMind guy said that AlphaGo would have utterly failed, if they would have modified the board by adding an extra column, whereas a human would still probably do okay-ish. So I am guessing that some meta-algorithm will be developed to detect such a modification, and then adjust the underlying subsystems accordingly.
Same with all the stated situations that self-driving cars seem to suck at currently (snow, construction sites, Tesla crashes). Right now, it seems that researches have to explicitly add the programming for these unforseen situations. But I bet this part will be automated, too: when the Tesla crashed into that trailer it didn't 'see', it should automatically detect that it messed up (crash is bad), and re-evaluate all previous sensor data to come up with a solution that would not have ended with a crash, and adjust its NN weights accordingly. This doesn't seem too much of a stretch in light of AlphaZero, and might very well lead to exponential AI 'cleverness'.
The key to deep learning neural networks is being able to simulate the performance, because it doesn't have a causal model that we would use to predict an outcome or weed out spurious correlations. That is to say it can't hold a million simulated conversations with a human figuring out what works and what doesn't. But it can do that with physics and other STEM branches, like it can play a driving simulator or the kerbal space program. And it can come up with new concepts within those constraints. I watched a documentary on AlphaGo not that long ago, and if you watch the expert reaction to move 37 in game 2 it's like WTF!? and unlike say Deep Blue it's not just the computer thinking even deeper than humans. It's basically going its own way and making a move no human would make. It's a double standard where for a human it would be novel and creative, for an AI it's just the result of an algorithm.
That has a lot of potential for say Rube Goldberg machines, like you give the computer a set of tools and basically says you figure out how these could be combined to achieve some sort of goal. It's a bit like giving the computer a toolbox and saying build me a house and it'll figure out what tools to use in which order to reach the end goal. The problem has been in building simulations that are sufficiently accurate that they can be applied in the real world and actually work like in the simulation. A lot of people here commented on how the Falcon Heavy launch performed pretty much exactly like in the CGI simulation. And Musk commented on it too in the press conference, it was a validation of their ability to design rockets through computer simulations. If it can be simulated, you can throw a DNN at it and ask it to optimize "the game". And that game has pretty serious real world implications.
Live today, because you never know what tomorrow brings
Our brains work with neural networks. It's only a matter of programming the coefficients. With enough layers, we will have a computer as smart as a person. If it's not deep learning, that does not negate the fact that when that happens whether in 5 years, 10 years, or 30 years, our intelligence will no longer be a marketable commodity. Think about that. How much is a person's physical labor worth these days? Not much now that automation can do so much more.
So the author claiming that deep learning won't do it? Who knows? We've got 20 smart people working on the problem in my university, and probably the same at every research institution in the world. The only question is how long will it take.
Saying that intelligence won't scale? That, the author is completely pulling out of their ass. Yes, people are good at different things, including social interaction, writing, gymnastics, and playing a violin. All those things work on the same hardware. There is no reason to think that human neural networks are particularly different than the neural networks being built in silicon. And all trends show that soon, robots will be able to run, walk, jump, drive better than humans.
This ass is no different than Lord Kelvin pronouncing that heavier than air flight would never work, except that Lord Kelvin did great work in thermodynamics, hydrodynamics, and this buffoon just writes puff pieces in wired.
Can any of these things divide by 3 yet? How about test for prime?
They're all basically databases in disguise. If it's not in the database, it's not going to know what to do beyond saying whether it's like things in the database in some way that they've been trained on.
However, they are very big databases. So Google Alpha doesn't play chess like a human - it doesn't really play chess at all - but it simulates it by lookups well enough to beat almost any human.
Computers have no cognition.
That is nonsense.
Every most stupid neural network "program" is trained to "recognize".
And the programmer did _nothing_
Every NN is basically an off the shelf empty brain, working the same way as any other empty brain, in other words: nothing at all. Only after training it does what it does: recognize stuff. Aka: performing "cognition".
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
"Sundar Pichai, the chief executive of Google, has said that AI "is more profound than ... electricity or fire."" ... underestimating the importance of both electricity and fire...
Regardless of the limitations of backprop, the traditional logic-based approach will never fully recover from the blow of discovering distributed representation.
https://youtu.be/wuhbqcMzOaw
Learning to swim and learning to drive a car are easy; machines can do that.
You are moving the goalposts.
The current so called AI cant change task domains without a complete overhaul.
You can write a program that can learn to drive. That program will never be able to learn how to swim, without massive recoding. Whic of course, means there is no intelligence.
numbnuts
You know nobody can mod you up if all you do is squabble with creimer.
1) Any post mentioning creimer related has a good chance of getting slaughtered in the m2 system.
2) Modding up anything in a creimer thread increases his visibility so most people won't do it even if it's funny.