4 ms·

"AI" is not currently autonomous; its algorithms that do exactly what their creators tell them to do. They run on binary computers that only do exactly as they

"AI" is not currently autonomous; its algorithms that do exactly what their creators tell them to do. They run on binary computers that only do exactly as they are told.

circuit10 4y ago
That’s not true, current machine learning algorithms involve no manual programming past the training and inference code and it’s extremely difficult to predict what they will do without just trying it and seeing. I think this video is a nice introduction to the basic concept to how a computer can figure things out automatically without being manually programmed and without the creators understanding the “why”: https://youtu.be/qv6UVOQ0F44 https://youtu.be/qv6UVOQ0F44 ChatGPT is much more complicated than the AI in that video but it shows some of the basic concepts
- goatlover 4y ago
  LLMs generate text. They're built to generate text. That they generate some kind of textual output is entirely predictable. Same with image generators. They will generate some kind of image given a prompt. They're not Skynet.
  - circuit10 4y ago
    That an AI will have some kind of output is obvious, it doesn’t mean that you can predict what that output will be. It’s like saying that you have solved physics by saying “something will happen”
    - goatlover 4y ago
      I'm confident the output of chatbot won't be literally taking over the world and killing all the humans.
    - philipov 4y ago
      I think the point he's trying to make is that AI does not have an independent Will. It lacks desires and the ability to operate in opposition to its programming. This makes it no different from any other tool we use to enhance our abilities. Whether or not you can predict a tool's output is irrelevant. I can't predict the output of a random number generator, but that doesn't make it sentient.
      - blackbear_ 4y ago
        This is not necessarily true, however, for example in reinforcement learning there is a lot of work on "intrinsic motivation", i.e., creating systems that set and pursue their own goals. https://arxiv.org/abs/1908.06976 https://arxiv.org/abs/1908.06976
        philipov 4y ago
        I think it should be possible to build a sentient AI, but it hasn't been done yet. What remains to be seen is whether our current techniques will be suitable for making that self-retraining process efficient, or if we'll need to find better math to use as the basis for it. Part of what makes the brain so useful is that it fits in our skull, and is fast enough to learn in real time. But, either way, I think that's what's on the line for people who disagree about how to use the word "intelligence." They mean it as a synonym for sentience, and the people arguing against them are using it differently. Before we can evaluate the truth of an argument, we should first agree to use words the same way.
      - flangola7 4y ago
        Asteroids are not sentient either, but they can still kill you.
        philipov 4y ago
        Asteroids aren't intelligent, either.
      - SanderNL 4y ago
        With LLMs you say “you want to do X” and voila, personality. What is indeed missing from current implementations is continuous looping. Doing actions and taking stock of the results. I guess that’s kind of expensive right now. We’ll get there. I don’t see the fundamental problem.
    - chongli 4y ago
      You can’t predict the outcome of a coin toss. That doesn’t make the coin intelligent.
  - planede 4y ago
    Humans are also entirely predictable, all they do is move their limbs around, make sound, etc...
    - PaulDavisThe1st 4y ago
      Sure, absolutely. Make a list of the things we do "entirely predictably". Make a list of the things a given ML system does "entirely predictably". One of these lists is (a lot) longer than the other. Interesting, no?
      - planede 4y ago
        Conversely if a person loses their ability to move their limbs or to talk then they also become less inteligent.
  - vharuck 4y ago
    To be fair, humans exist only because of a long chain of organisms that started with "DNA generates proteins." Granted, it took billions of years for that process to create humans, but it shows that what seems to be a constrained process can have wild outcomes when it feeds itself. And text commands are how these models are defined, trained, deployed, and used.
- candiddevmike 4y ago
  I think OP meant autonomous in the sense of motivation. It has no motivation to do anything aside from wait for someone to ask it to do something.
  - pixl97 4y ago
    I mean, if I was paying for the power bill every month and had a limited amount of computing capacity, I wouldn't want my AI behaving like my teenage daughter busy daydreaming when I ask her to clean her room. But I have no reason to believe this will always be the case. As these machines become more capable and our compute power grows, someone will give one a server cluster and some free time to 'think' on it's own.
ben_w 4y ago
Given that algorithms are "how to learn" and "show me what you infer", that's the same kind of overly reductionist view that you don't need to worry about being eaten by a tiger, because it's just a set of chemical reactions that merely follow the laws of quantum mechanics.
- circuit10 4y ago
  https://twitter.com/nearcyan/status/1632661647226462211 https://twitter.com/nearcyan/status/1632661647226462211
  - ben_w 4y ago
    Yup, that's the thing I was thinking of. Wasn't even sure where I'd seen it, so thanks for the link!
- PaulDavisThe1st 4y ago
  Not entirely a fair comparison. The tiger is dangerous because whether you consider it a sentient, intentional killing machine or a bunch of atoms, it exists and manipulates the same physical space that you do (indeed, as the tweeted image points out implicitly, it is only a tiger when you consider at the same sort of physical scale that we exist at). Software, however, does not have this property. Ultimately it does exist as something in the physical world (voltages on gates, or whatever), but at that level it's equivalent to the "bunch of atoms" view. Software (by itself) does not operate in the physical space that we do, and so it cannot pose the same kind of threats to us as other physical systems do. The question is therefore a lot more nuanced: what types of control (if any) can (a given piece of) software exert over the world in which we operate? This includes the abstract yet still large scale world of things like finance and record keeping, but it also obviously covers the physical space in which our bodies exist. Right now, there is very (very) little software that exists as a sentient, intentional threat to us within that space. When and if software starts to be able to exert more force on that space, then the "it's just logic and gates and stuff" view will be inappropriate. For now, the main risk from software comes from what other humans will do with it, not what it will do to us (though smartphones do raise issues about even that).
YetAnotherNick 4y ago
This is a flawed analogy, it certainly breaks down even in simple case of random number generation. Computers could use external source like minor heat changes for that.
krisoft 4y ago
> its algorithms that do exactly what their creators tell them to do This is very much in doubt :) > They run on binary computers that only do exactly as they are told. This is true in first approximation. Every CPU instruction runs exactly as it is written, that is true. This is probably the interpretation of "only do exactly as they are told" to someone strictly technology minded. But even with much simpler systems the words "huh, that should have not happened", and "I wonder why it is doing that" are uttered frequently. The interpretation most humans would attach to "only do exactly as they are told" is that the maker can predict what the code will do, and that is far from the truth. After all if it so simple, why did the google engineers tell their computer to tell lies about the James Webb Space Telescope? Couldn't they just told it to only tell the truth?
- mahoho 4y ago
  I think the machine code–level understanding is what's important. We can, in theory, put a person in a Chinese Room–style scenario and have them manually perform the code, and it will generate the same outputs (It would probably take millions or billions of years, but it is true in principle). A major difference is that we created the machine and the code and, at least as low as the level of digital logic design, we understand and control its behavior. The person in the room has a human mind with thoughts and behaviors completely out of the program designers' control and unrelated to the program; if they want to, they can break out of the room and punch the operator. The "unpredictability" of the machine is still constrained by the fundamental capabilities we give to it, so it might generate surprising outputs but it can't do things like punch people or launch nukes unless we connect it to other systems that have those capabilities.
  - PaulDavisThe1st 4y ago
    > A major difference is that we created the machine and the code and, at least as low as the level of digital logic design, we understand and control its behavior. The moment the software gets to interact with the world, whether via robotics or handling a mouse button event or some other type of sensor, we no longer fully understand or control its behavior. Pure computation (the dream of functional programming) is fully understandable and entirely predictable. When you add interaction, you add both randomness but also time - when something happens can lead to different outcomes - and this can rapidly cause predictability to spiral away from us.

"AI" is not currently autonomous; its algorithms that do exactly what their creators tell them to do. They run on binary computers that only do exactly as they are told.

circuit10 4y ago

That’s not true, current machine learning algorithms involve no manual programming past the training and inference code and it’s extremely difficult to predict what they will do without just trying it and seeing. I think this video is a nice introduction to the basic concept to how a computer can figure things out automatically without being manually programmed and without the creators understanding the “why”: https://youtu.be/qv6UVOQ0F44 https://youtu.be/qv6UVOQ0F44 ChatGPT is much more complicated than the AI in that video but it shows some of the basic concepts

goatlover 4y ago
LLMs generate text. They're built to generate text. That they generate some kind of textual output is entirely predictable. Same with image generators. They will generate some kind of image given a prompt. They're not Skynet.
- circuit10 4y ago
  That an AI will have some kind of output is obvious, it doesn’t mean that you can predict what that output will be. It’s like saying that you have solved physics by saying “something will happen”
  - goatlover 4y ago
    I'm confident the output of chatbot won't be literally taking over the world and killing all the humans.
  - philipov 4y ago
    I think the point he's trying to make is that AI does not have an independent Will. It lacks desires and the ability to operate in opposition to its programming. This makes it no different from any other tool we use to enhance our abilities. Whether or not you can predict a tool's output is irrelevant. I can't predict the output of a random number generator, but that doesn't make it sentient.
    - blackbear_ 4y ago
      This is not necessarily true, however, for example in reinforcement learning there is a lot of work on "intrinsic motivation", i.e., creating systems that set and pursue their own goals. https://arxiv.org/abs/1908.06976 https://arxiv.org/abs/1908.06976
      - philipov 4y ago
        I think it should be possible to build a sentient AI, but it hasn't been done yet. What remains to be seen is whether our current techniques will be suitable for making that self-retraining process efficient, or if we'll need to find better math to use as the basis for it. Part of what makes the brain so useful is that it fits in our skull, and is fast enough to learn in real time. But, either way, I think that's what's on the line for people who disagree about how to use the word "intelligence." They mean it as a synonym for sentience, and the people arguing against them are using it differently. Before we can evaluate the truth of an argument, we should first agree to use words the same way.
    - flangola7 4y ago
      Asteroids are not sentient either, but they can still kill you.
      - philipov 4y ago
        Asteroids aren't intelligent, either.
    - SanderNL 4y ago
      With LLMs you say “you want to do X” and voila, personality. What is indeed missing from current implementations is continuous looping. Doing actions and taking stock of the results. I guess that’s kind of expensive right now. We’ll get there. I don’t see the fundamental problem.
  - chongli 4y ago
    You can’t predict the outcome of a coin toss. That doesn’t make the coin intelligent.
- planede 4y ago
  Humans are also entirely predictable, all they do is move their limbs around, make sound, etc...
  - PaulDavisThe1st 4y ago
    Sure, absolutely. Make a list of the things we do "entirely predictably". Make a list of the things a given ML system does "entirely predictably". One of these lists is (a lot) longer than the other. Interesting, no?
    - planede 4y ago
      Conversely if a person loses their ability to move their limbs or to talk then they also become less inteligent.
- vharuck 4y ago
  To be fair, humans exist only because of a long chain of organisms that started with "DNA generates proteins." Granted, it took billions of years for that process to create humans, but it shows that what seems to be a constrained process can have wild outcomes when it feeds itself. And text commands are how these models are defined, trained, deployed, and used.
candiddevmike 4y ago
I think OP meant autonomous in the sense of motivation. It has no motivation to do anything aside from wait for someone to ask it to do something.
- pixl97 4y ago
  I mean, if I was paying for the power bill every month and had a limited amount of computing capacity, I wouldn't want my AI behaving like my teenage daughter busy daydreaming when I ask her to clean her room. But I have no reason to believe this will always be the case. As these machines become more capable and our compute power grows, someone will give one a server cluster and some free time to 'think' on it's own.

ben_w 4y ago

Given that algorithms are "how to learn" and "show me what you infer", that's the same kind of overly reductionist view that you don't need to worry about being eaten by a tiger, because it's just a set of chemical reactions that merely follow the laws of quantum mechanics.

circuit10 4y ago
https://twitter.com/nearcyan/status/1632661647226462211 https://twitter.com/nearcyan/status/1632661647226462211
- ben_w 4y ago
  Yup, that's the thing I was thinking of. Wasn't even sure where I'd seen it, so thanks for the link!
PaulDavisThe1st 4y ago
Not entirely a fair comparison. The tiger is dangerous because whether you consider it a sentient, intentional killing machine or a bunch of atoms, it exists and manipulates the same physical space that you do (indeed, as the tweeted image points out implicitly, it is only a tiger when you consider at the same sort of physical scale that we exist at). Software, however, does not have this property. Ultimately it does exist as something in the physical world (voltages on gates, or whatever), but at that level it's equivalent to the "bunch of atoms" view. Software (by itself) does not operate in the physical space that we do, and so it cannot pose the same kind of threats to us as other physical systems do. The question is therefore a lot more nuanced: what types of control (if any) can (a given piece of) software exert over the world in which we operate? This includes the abstract yet still large scale world of things like finance and record keeping, but it also obviously covers the physical space in which our bodies exist. Right now, there is very (very) little software that exists as a sentient, intentional threat to us within that space. When and if software starts to be able to exert more force on that space, then the "it's just logic and gates and stuff" view will be inappropriate. For now, the main risk from software comes from what other humans will do with it, not what it will do to us (though smartphones do raise issues about even that).

YetAnotherNick 4y ago

This is a flawed analogy, it certainly breaks down even in simple case of random number generation. Computers could use external source like minor heat changes for that.

krisoft 4y ago

> its algorithms that do exactly what their creators tell them to do This is very much in doubt :) > They run on binary computers that only do exactly as they are told. This is true in first approximation. Every CPU instruction runs exactly as it is written, that is true. This is probably the interpretation of "only do exactly as they are told" to someone strictly technology minded. But even with much simpler systems the words "huh, that should have not happened", and "I wonder why it is doing that" are uttered frequently. The interpretation most humans would attach to "only do exactly as they are told" is that the maker can predict what the code will do, and that is far from the truth. After all if it so simple, why did the google engineers tell their computer to tell lies about the James Webb Space Telescope? Couldn't they just told it to only tell the truth?

mahoho 4y ago
I think the machine code–level understanding is what's important. We can, in theory, put a person in a Chinese Room–style scenario and have them manually perform the code, and it will generate the same outputs (It would probably take millions or billions of years, but it is true in principle). A major difference is that we created the machine and the code and, at least as low as the level of digital logic design, we understand and control its behavior. The person in the room has a human mind with thoughts and behaviors completely out of the program designers' control and unrelated to the program; if they want to, they can break out of the room and punch the operator. The "unpredictability" of the machine is still constrained by the fundamental capabilities we give to it, so it might generate surprising outputs but it can't do things like punch people or launch nukes unless we connect it to other systems that have those capabilities.
- PaulDavisThe1st 4y ago
  > A major difference is that we created the machine and the code and, at least as low as the level of digital logic design, we understand and control its behavior. The moment the software gets to interact with the world, whether via robotics or handling a mouse button event or some other type of sensor, we no longer fully understand or control its behavior. Pure computation (the dream of functional programming) is fully understandable and entirely predictable. When you add interaction, you add both randomness but also time - when something happens can lead to different outcomes - and this can rapidly cause predictability to spiral away from us.