4 ms·

To create AGI, we need a new theory of intelligence

kordlessagain 5y ago
> “Such agents do not need to be anthropomorphic — they could have unusual designs, structures and functions that would produce intelligent behavior alien to our own (e.g., an octopus-like design, with brain functions distributed throughout the body),” Raghavachary said. “That said, the most relatable human-level AI would likely be best housed in a human-like agent.”
- mindcrime 5y ago
  “That said, the most relatable human-level AI would likely be best housed in a human-like agent.” I would agree with that. I think it's probably possible in principle to get to something that could be considered "generally intelligent" without embodiment, but as I allude to in my other comment, it seems to me that if you really want something extremely "human like" then embodiment is probably going to be very important.
  - kordlessagain 5y ago
    There's only one reason to really want something extremely human like...
    - mindcrime 5y ago
      Alright now, let's keep this "family friendly"! :p
      - kordlessagain 5y ago
        I meant so it won't kill us, but I get your drift. :)
mindcrime 5y ago
There is some interesting stuff here, but I'll need more time to stop and really digest this. I will say though, that a quick skim reveals a few instances of what strike me as something of a "naturalistic fallacy" argument. That is, an argument of the form "intelligence in nature doesn't work like this, therefore we can't achieve intelligence like this." Regarding this bit: In his paper, he notes that there is not a single example of “brain in a vat” in nature’s diverse array of intelligent lifeforms. And thus, the considered response theory of intelligence suggests that artificial general intelligence requires agents that can have a direct embodied experience of the world. This is an example of what I consider a bit of a naturalistic fallacy in a sense, although the extent to which that is true depends on exactly how you define "artificial general intelligence." And this confusion, to me, gets at one of the core issues that make this issue hard to discuss. Most people conflate "human level general intelligence" and "human like behavior" in these discussions. And I posit that they are two very different - albeit related - things, and that the latter might not actually be a goal at all. Within that framework though, I have a strong bias towards agreeing with the sentiment that if you do want "human like behavior" then embodiment is likely to be a requirement, or at least a near requirement. And even if you don't care about strict "human like behavior" but merely want "human level general intelligence" embodiment may still be an important element, albeit perhaps not an absolute requirement. Why does embodiment seem so important? My position is that our embodied interactions with the physical world are how we learn a lot of important things that are hard to manually encode - our basic metaphysics of the world (what are objects, what is time, what is movement, do objects disappear from existence when our vision of them is occluded, etc), our intuitionist notion of the physics of the world (things fall down if not supported, things stop falling when they hit other things, unless they hit the other thing and ricochet, etc.), and so on. And a lot of metaphors we use in cognition and communication are based on these fundamental learnings. Take the word "fall" for example (as in "I don't want to see the market fall this week"). I'd argue that the underlying notion here relates back to our association of the word "fall" with the act of falling down - something we has humans understand because we've all experienced a fall, and the resulting pain, scraped palms, etc. And there are so many other examples like this that you could consider. Anyway, looking forward to finding some time to dig in and read this and thoroughly digest it.

> “Such agents do not need to be anthropomorphic — they could have unusual designs, structures and functions that would produce intelligent behavior alien to our own (e.g., an octopus-like design, with brain functions distributed throughout the body),” Raghavachary said. “That said, the most relatable human-level AI would likely be best housed in a human-like agent.”

mindcrime 5y ago
“That said, the most relatable human-level AI would likely be best housed in a human-like agent.” I would agree with that. I think it's probably possible in principle to get to something that could be considered "generally intelligent" without embodiment, but as I allude to in my other comment, it seems to me that if you really want something extremely "human like" then embodiment is probably going to be very important.
- kordlessagain 5y ago
  There's only one reason to really want something extremely human like...
  - mindcrime 5y ago
    Alright now, let's keep this "family friendly"! :p
    - kordlessagain 5y ago
      I meant so it won't kill us, but I get your drift. :)

mindcrime 5y ago

There is some interesting stuff here, but I'll need more time to stop and really digest this. I will say though, that a quick skim reveals a few instances of what strike me as something of a "naturalistic fallacy" argument. That is, an argument of the form "intelligence in nature doesn't work like this, therefore we can't achieve intelligence like this." Regarding this bit: In his paper, he notes that there is not a single example of “brain in a vat” in nature’s diverse array of intelligent lifeforms. And thus, the considered response theory of intelligence suggests that artificial general intelligence requires agents that can have a direct embodied experience of the world. This is an example of what I consider a bit of a naturalistic fallacy in a sense, although the extent to which that is true depends on exactly how you define "artificial general intelligence." And this confusion, to me, gets at one of the core issues that make this issue hard to discuss. Most people conflate "human level general intelligence" and "human like behavior" in these discussions. And I posit that they are two very different - albeit related - things, and that the latter might not actually be a goal at all. Within that framework though, I have a strong bias towards agreeing with the sentiment that if you do want "human like behavior" then embodiment is likely to be a requirement, or at least a near requirement. And even if you don't care about strict "human like behavior" but merely want "human level general intelligence" embodiment may still be an important element, albeit perhaps not an absolute requirement. Why does embodiment seem so important? My position is that our embodied interactions with the physical world are how we learn a lot of important things that are hard to manually encode - our basic metaphysics of the world (what are objects, what is time, what is movement, do objects disappear from existence when our vision of them is occluded, etc), our intuitionist notion of the physics of the world (things fall down if not supported, things stop falling when they hit other things, unless they hit the other thing and ricochet, etc.), and so on. And a lot of metaphors we use in cognition and communication are based on these fundamental learnings. Take the word "fall" for example (as in "I don't want to see the market fall this week"). I'd argue that the underlying notion here relates back to our association of the word "fall" with the act of falling down - something we has humans understand because we've all experienced a fall, and the resulting pain, scraped palms, etc. And there are so many other examples like this that you could consider. Anyway, looking forward to finding some time to dig in and read this and thoroughly digest it.