6 ms·

I see a lot of AI engineers who seem concerned with this particular issue, which I never really understand. Is it because of a perception that most regular peo

I see a lot of AI engineers who seem concerned with this particular issue, which I never really understand. Is it because of a perception that most regular people are likely overestimating the speed of which AI is going to overtake human intelligence? Or more about corp management wanting miracles that aren't possible? Why does this matter and always seem to be talked about?

UncleOxidant 7y ago
Because there's a history of overhyping ML/AI (whatever you want to call it) leading to AI winters. Winter in this case being kind of like a recession in economic terms - most research funding dries up, etc. We essentially had one of those winters from the late 80s until about a dozen years ago. A lot of laymen now think of AI as being "magic" that can do anything and that's not a good thing when the reality turns out to be different. At this point I don't think we'll see an AI winter as deep as some of the previous ones. But we could certainly see an AI Fall.
- 0xdeadbeefbabe 7y ago
  The name is overhyped and pretentious by itself, and history bears this out. Who cares if it's an AI fall or winter if it's an AI stupid, because of all the credulous students. Edit: Russel and Norvig's book is good though http://aima.cs.berkeley.edu/ http://aima.cs.berkeley.edu/
- YeGoblynQueenne 7y ago
  >> Because there's a history of overhyping ML/AI (whatever you want to call it) leading to AI winters. Note that past AI winters have not occurred because of overhyping machine learning. They occurred because of overhyping of symbolic AI that had nothing to do with machine learning. For example, the last AI winter at the end of the '80s happened because of the overhyping of expert systems- which of course are not machine learning systems. Machine learning is not all, not even most, of AI, historically. It's the dominant trend right now, but it was not the dominant trend in the past. The dominant trend until the 1980's was symbolic reasoning.
  - UncleOxidant 7y ago
    There was a neural net popularity surge in the late 80s, early 90s. Of course, the hardware wasn't there yet to be able to deliver on the promises. I was in a Goodwill book section about a year ago and there were a couple of NN books from that era on sale for $3, one titled "Apprentices of Wonder: Inside the Neural Network Revolution" from 1990 and the other was for programmers and included C code for a NN to predict the stock market from 1989. Anyway, that all had died out by about '92 or '93 and NNs were a pretty dead academic topic until about 2005 or so when they figured out that GPUs could be used to accelerate them.
  - radarsat1 7y ago
    But symbolic reasoning mostly worked, did it not? However, its Achilles heel was that for it to be useful, it's necessary to distill a lot of domain knowledge into a format that can be processed by an expert system. That means, writing 10s upon thousands of rows of "if then then that". Machine learning is different in that it is more amenable to distilling those rules from the data automatically. It is successful where symbolic reasoning failed because it can go from the raw data. A good portion of machine learning research is in new ways to preprocess and format data into a structure that can be further consumed by linear algebra, which turns out to be a lot easier and practical than figure out a huge database of sensible first order predicate logic statements. If ML techniques can be used to feed symbolic systems, the latter would show promise again, which is already happening in recent trends in causal inference and graph networks. The marriage of these two fields is inevitable, and has already started.
    - mistrial9 7y ago
      > Machine learning is different .. partly because, Machine Learning is inherently probabilistic. Lots of room for errors and hype with systems that never claim to give the right answer all of the time!
    - blackrock 7y ago
      > That means, writing 10s upon thousands of rows of "if then then that". LOL.. It's full of ifs.. [1] [1] https://qph.fs.quoracdn.net/main-qimg-89cfa17ca63ddac683e04f0852336e47.webp https://qph.fs.quoracdn.net/main-qimg-89cfa17ca63ddac683e04f...
twblalock 7y ago
It seems analogous to Searle's "Chinese Room" argument: automated responses to predefined stimuli isn't the same as "intelligence" or "understanding". The OP suggests modern AI is a fancy way of teaching systems to effectively hardcode or automate their behavior themselves. I'm not sure why that matters, as long as the results are what we aim for. It's not like most AI researchers are trying to create sentient artificial life-forms.
- mattkrause 7y ago
  It matters because the hard-coded behavior is brittle and often doesn't do exactly what we want (or think). For example, GPT-2 has been ascribed nearly magical powers: it's a knowledge base, it can play chess, it does calculus, it's a dessert topping AND a floor wax! When you look closer, however, it doesn't do any of those things particularly well. It can regurgitate something that looks like a true fact--or its negation with equal probability. It doesn't quite know the rules of chess. It needs a solver to check that the solution to an integral is, in fact, a solution.
  - whatshisface 7y ago
    All of those caveats apply to human intelligence, but to a lesser degree. Kids can play chess without exactly knowing the rules, and come on, everybody needs to check their integrals.
    - TeMPOraL 7y ago
      But kids can eventually learn proper algorithms for chess and integrals. DNNs cannot. In a sense, deep learning is like slide rules - you can squeeze some problem domains into giant, funny looking lookup tables, but generalize it does not.
- taurath 7y ago
  Interestingly though, at the population level it can seem a lot closer.
TallGuyShort 7y ago
Personally, when a lay person asks what I do, I like telling people I work on "Artificial Intelligence software" because it's the most accurate term that doesn't (a) get an immediate request to implement their app idea for them and (b) require explaining what machine learning / deep learning is. But beyond that I hate the term within the industry because I think artificial intelligence gets equated with a Jarvis-like general AI that will talk to you like a superhuman servant. I get the desire to better define the current state of the art. But for most people, I agree it's going to seem like pedantry.
- laichzeit0 7y ago
  There’s already a term you can use. “Statistical learning”. There’s even a well known important book with that title: Elements of Statistical Learning.
  - TallGuyShort 7y ago
    There are books well known in the field with "machine learning" in the title. I don't think that's any clearer to a lay person.
- radarsat1 7y ago
  > artificial intelligence gets equated with a Jarvis-like general AI that will talk to you like a superhuman servant to be fair, for a lot of researchers, that is the ultimate end goal, even for those who admit we are not even close to it. I for one first got interested in AI from an 80's movie, can't remember which, with a character who talked to his computer, which talked back. Since those early years, I haven't spent even one second working on actual AGI, seeing the plethora of subgoals needed to get there, but.. thinking about it.. plenty. that dream is a driving force behind more ML/AI researchers than maybe you think. Particularly in the RL community I would guess.
wayoutthere 7y ago
It's because this kind of hype inevitably leads to a trough of disillusionment -- the methods we collectively call "AI" today are never going to lead to a general-purpose artificial intelligence. People are disappointed we don't have self-driving cars yet, but it's not clear whether that problem domain is constrained enough for deep neural networks to solve. What we have developed are ways to automate complex tasks within a constrained input domain that can be easily quantified. It seems like magic, which leads people to say that it's "AI" but in reality it's just a complex automation built through reinforcement techniques that leverage some clever math tricks. Throw an unexpected input or new set of circumstances at the model and you get interesting results. It's not a sense that people are overestimating the speed with which AI is going to overtake human intelligence -- it's that the techniques we're using today that we call "AI" are not capable of doing anything of the sort.
- hunter-gatherer 7y ago
  This. Working in big Corp and Government I have seen how far this disillusionment can take an organization down the wrong road. Learning how to articulate complex technical / scientific topics to bureaucracy, I am learning, is a very valuable and needed skill amongst engineers.
  - wayoutthere 7y ago
    It's a fine line you have to walk. They usually are looking for a person who will tell them what they want to hear, so it's usually a matter of starting with "the art of the possible" (aka a bunch of bullshit they heard on NPR) and working them over to something more realistic. I've found it helps if you can frame it in the context of the other options (i.e. agree with where they want to go and present multiple ways to get there) they're more receptive. Leaders know about these hype cycles too, but they often have to play along for political reasons and they'll be thankful if you work with them rather than against them.
liamcardenas 7y ago
Andrew Yang is a serious contender for the US presidency whose entire platform rests on assumptions about AI. He wants to fundamentally reshape welfare in the country and implement an entirely new tax. Thinking clearly about AI is therefore very important, as it is having real and substantial political implications.
- chrshawkes 7y ago
  Luckily he won't be elected and shouldn't if he really thinks AI (in its current form) will solve these problems.
  - jimbokun 7y ago
    He doesn't think it will solve them, he thinks it is going to cause them, and that we need to be ready with solutions. Take Universal Basic Income. He is predicting far more jobs are going to be automated in the near future than most people expect, and something like UBI will be needed to keep the people out of work from starving or rioting.
- chillacy 7y ago
  I remember him a year ago saying stuff that wasn't particularly mainstream, like warning about fast food cashiers being replaced by kiosks, malls closing due to competition with Amazon, call center workers being automated, etc. These are all things that are coming, I have peers working on some of them, but they aren't particularly mainstream, even though the most accessible jobs in the economy fall under those categories.
  - ForHackernews 7y ago
    The thing is, technology to replace fast food cashiers or call center employees doesn't even have to be good, it just has to be really cheap. The companies making the decision to put those systems in place are not the end users who will be forced to interface with them. Human customers will be forced to modify their speech to be understood by mediocre telephone agent programs. So yes, Yang is right that automation is coming for millions of American jobs, and sooner than most people might think.

UncleOxidant 7y ago

Because there's a history of overhyping ML/AI (whatever you want to call it) leading to AI winters. Winter in this case being kind of like a recession in economic terms - most research funding dries up, etc. We essentially had one of those winters from the late 80s until about a dozen years ago. A lot of laymen now think of AI as being "magic" that can do anything and that's not a good thing when the reality turns out to be different. At this point I don't think we'll see an AI winter as deep as some of the previous ones. But we could certainly see an AI Fall.

0xdeadbeefbabe 7y ago
The name is overhyped and pretentious by itself, and history bears this out. Who cares if it's an AI fall or winter if it's an AI stupid, because of all the credulous students. Edit: Russel and Norvig's book is good though http://aima.cs.berkeley.edu/ http://aima.cs.berkeley.edu/
YeGoblynQueenne 7y ago
>> Because there's a history of overhyping ML/AI (whatever you want to call it) leading to AI winters. Note that past AI winters have not occurred because of overhyping machine learning. They occurred because of overhyping of symbolic AI that had nothing to do with machine learning. For example, the last AI winter at the end of the '80s happened because of the overhyping of expert systems- which of course are not machine learning systems. Machine learning is not all, not even most, of AI, historically. It's the dominant trend right now, but it was not the dominant trend in the past. The dominant trend until the 1980's was symbolic reasoning.
- UncleOxidant 7y ago
  There was a neural net popularity surge in the late 80s, early 90s. Of course, the hardware wasn't there yet to be able to deliver on the promises. I was in a Goodwill book section about a year ago and there were a couple of NN books from that era on sale for $3, one titled "Apprentices of Wonder: Inside the Neural Network Revolution" from 1990 and the other was for programmers and included C code for a NN to predict the stock market from 1989. Anyway, that all had died out by about '92 or '93 and NNs were a pretty dead academic topic until about 2005 or so when they figured out that GPUs could be used to accelerate them.
- radarsat1 7y ago
  But symbolic reasoning mostly worked, did it not? However, its Achilles heel was that for it to be useful, it's necessary to distill a lot of domain knowledge into a format that can be processed by an expert system. That means, writing 10s upon thousands of rows of "if then then that". Machine learning is different in that it is more amenable to distilling those rules from the data automatically. It is successful where symbolic reasoning failed because it can go from the raw data. A good portion of machine learning research is in new ways to preprocess and format data into a structure that can be further consumed by linear algebra, which turns out to be a lot easier and practical than figure out a huge database of sensible first order predicate logic statements. If ML techniques can be used to feed symbolic systems, the latter would show promise again, which is already happening in recent trends in causal inference and graph networks. The marriage of these two fields is inevitable, and has already started.
  - mistrial9 7y ago
    > Machine learning is different .. partly because, Machine Learning is inherently probabilistic. Lots of room for errors and hype with systems that never claim to give the right answer all of the time!
  - blackrock 7y ago
    > That means, writing 10s upon thousands of rows of "if then then that". LOL.. It's full of ifs.. [1] [1] https://qph.fs.quoracdn.net/main-qimg-89cfa17ca63ddac683e04f0852336e47.webp https://qph.fs.quoracdn.net/main-qimg-89cfa17ca63ddac683e04f...

twblalock 7y ago

It seems analogous to Searle's "Chinese Room" argument: automated responses to predefined stimuli isn't the same as "intelligence" or "understanding". The OP suggests modern AI is a fancy way of teaching systems to effectively hardcode or automate their behavior themselves. I'm not sure why that matters, as long as the results are what we aim for. It's not like most AI researchers are trying to create sentient artificial life-forms.

mattkrause 7y ago
It matters because the hard-coded behavior is brittle and often doesn't do exactly what we want (or think). For example, GPT-2 has been ascribed nearly magical powers: it's a knowledge base, it can play chess, it does calculus, it's a dessert topping AND a floor wax! When you look closer, however, it doesn't do any of those things particularly well. It can regurgitate something that looks like a true fact--or its negation with equal probability. It doesn't quite know the rules of chess. It needs a solver to check that the solution to an integral is, in fact, a solution.
- whatshisface 7y ago
  All of those caveats apply to human intelligence, but to a lesser degree. Kids can play chess without exactly knowing the rules, and come on, everybody needs to check their integrals.
  - TeMPOraL 7y ago
    But kids can eventually learn proper algorithms for chess and integrals. DNNs cannot. In a sense, deep learning is like slide rules - you can squeeze some problem domains into giant, funny looking lookup tables, but generalize it does not.
taurath 7y ago
Interestingly though, at the population level it can seem a lot closer.

TallGuyShort 7y ago

Personally, when a lay person asks what I do, I like telling people I work on "Artificial Intelligence software" because it's the most accurate term that doesn't (a) get an immediate request to implement their app idea for them and (b) require explaining what machine learning / deep learning is. But beyond that I hate the term within the industry because I think artificial intelligence gets equated with a Jarvis-like general AI that will talk to you like a superhuman servant. I get the desire to better define the current state of the art. But for most people, I agree it's going to seem like pedantry.

laichzeit0 7y ago
There’s already a term you can use. “Statistical learning”. There’s even a well known important book with that title: Elements of Statistical Learning.
- TallGuyShort 7y ago
  There are books well known in the field with "machine learning" in the title. I don't think that's any clearer to a lay person.
radarsat1 7y ago
> artificial intelligence gets equated with a Jarvis-like general AI that will talk to you like a superhuman servant to be fair, for a lot of researchers, that is the ultimate end goal, even for those who admit we are not even close to it. I for one first got interested in AI from an 80's movie, can't remember which, with a character who talked to his computer, which talked back. Since those early years, I haven't spent even one second working on actual AGI, seeing the plethora of subgoals needed to get there, but.. thinking about it.. plenty. that dream is a driving force behind more ML/AI researchers than maybe you think. Particularly in the RL community I would guess.

wayoutthere 7y ago

It's because this kind of hype inevitably leads to a trough of disillusionment -- the methods we collectively call "AI" today are never going to lead to a general-purpose artificial intelligence. People are disappointed we don't have self-driving cars yet, but it's not clear whether that problem domain is constrained enough for deep neural networks to solve. What we have developed are ways to automate complex tasks within a constrained input domain that can be easily quantified. It seems like magic, which leads people to say that it's "AI" but in reality it's just a complex automation built through reinforcement techniques that leverage some clever math tricks. Throw an unexpected input or new set of circumstances at the model and you get interesting results. It's not a sense that people are overestimating the speed with which AI is going to overtake human intelligence -- it's that the techniques we're using today that we call "AI" are not capable of doing anything of the sort.

hunter-gatherer 7y ago
This. Working in big Corp and Government I have seen how far this disillusionment can take an organization down the wrong road. Learning how to articulate complex technical / scientific topics to bureaucracy, I am learning, is a very valuable and needed skill amongst engineers.
- wayoutthere 7y ago
  It's a fine line you have to walk. They usually are looking for a person who will tell them what they want to hear, so it's usually a matter of starting with "the art of the possible" (aka a bunch of bullshit they heard on NPR) and working them over to something more realistic. I've found it helps if you can frame it in the context of the other options (i.e. agree with where they want to go and present multiple ways to get there) they're more receptive. Leaders know about these hype cycles too, but they often have to play along for political reasons and they'll be thankful if you work with them rather than against them.

liamcardenas 7y ago

Andrew Yang is a serious contender for the US presidency whose entire platform rests on assumptions about AI. He wants to fundamentally reshape welfare in the country and implement an entirely new tax. Thinking clearly about AI is therefore very important, as it is having real and substantial political implications.

chrshawkes 7y ago
Luckily he won't be elected and shouldn't if he really thinks AI (in its current form) will solve these problems.
- jimbokun 7y ago
  He doesn't think it will solve them, he thinks it is going to cause them, and that we need to be ready with solutions. Take Universal Basic Income. He is predicting far more jobs are going to be automated in the near future than most people expect, and something like UBI will be needed to keep the people out of work from starving or rioting.
chillacy 7y ago
I remember him a year ago saying stuff that wasn't particularly mainstream, like warning about fast food cashiers being replaced by kiosks, malls closing due to competition with Amazon, call center workers being automated, etc. These are all things that are coming, I have peers working on some of them, but they aren't particularly mainstream, even though the most accessible jobs in the economy fall under those categories.
- ForHackernews 7y ago
  The thing is, technology to replace fast food cashiers or call center employees doesn't even have to be good, it just has to be really cheap. The companies making the decision to put those systems in place are not the end users who will be forced to interface with them. Human customers will be forced to modify their speech to be understood by mediocre telephone agent programs. So yes, Yang is right that automation is coming for millions of American jobs, and sooner than most people might think.