5 ms·

It’s fascinating how many people feel that this is a specifically powerful breakthrough toward AGI, I suppose because it seems creative. Don’t get me wrong, thi

It’s fascinating how many people feel that this is a specifically powerful breakthrough toward AGI, I suppose because it seems creative. Don’t get me wrong, this is extremely cool! But art and composition are “pattern languages” that seem well-suited to our current ML techniques given sufficient investment in training. To me this feels more like progress toward the Star Trek computer, which we can interact with in basically natural language, and which can do sophisticated knowledge operations for us (from natural language instructions), but which still doesn’t “think” in a meaningful way. Now, when Dall-E 5 starts to argue with you about which art style looks best or will be most marketable…

Lambdanaut 5y ago
GPT-3 has been able to argue with you about which art styles look best or will be most marketable for over a year now
- robbedpeter 5y ago
  I don't think a lot of these critics realize how capable gpt-3 is of convincingly outsmarting most people in most contexts. The failures are glaring, and you'll no doubt catch them over time, but transformers are a sea change in ai power. If gpt-3 can meet or beat human performance 80% of the time, that's sufficient for me to call it real intelligence. Not conscious or aware in any meaningful way, but smart, in many of the ways humans are smart. In most online contexts where conversations are singular collisions between individuals who often don't have history, 80% of the real-seeming interactions you have could be gpt-3 output. In a many-shot, recursive, and domain targeted sequence of prompts, the percentage only improves. The recursive, self referential functions learned by these models are not mere extrapolation from Markov chains and static stimulus/response bots. These models are approximating the functions required to produce human level output. They are approximating the functions biological brains use to produce output. Even the naive, single pass probes into gpt capabilities we've seen in the last two years are incredible compared to aiml or even the best bots until ~2018. It's not agi, but it's evidence that agi is feasible, and a definitive milestone on that path. We need to be preparing for takeoff.
  - andai 5y ago
    >Not conscious or aware in any meaningful way On what grounds have you determined that GPT-3 has no subjective experiences? More generally, what would need to be the case (with some future AI) for you to say, "ok, yes, this one is definitely conscious" ?
    - tshaddox 5y ago
      I would need some explanation of what consciousness is, for starters. Even better would be an explanation of how I might test whether any thing (including myself or another person) has it.
      - andai 5y ago
        A perspective I thought was interesting is when Jordan Peterson says that the "proof" that people have a divine spark within them is "try treating them as if they don't, and see how well that works out for you!" In that sense it's quite similar to the idea of always assuming a person is acting in good faith: you cannot ever truly know someone's intentions, and even when everyone is communicating honestly and openly, communication is imperfect. But if you start to believe that someone has negative intentions towards you, your body language, tone, words and actions will reflect that assumption, and it will create harmful outcomes where they were not necessary at all. Similarly, we may never be able to objectively prove that a machine is conscious -- as you say, we cannot even prove it to another human about ourselves! But if we keep treating them with the assumption that they lack this "something special" that we have, and treat them unethically as a result, I think we could get ourselves into a lot of trouble. A similar conversation is to be had about the way humans are factory-farming animals, 80 billion of which are born and die every year in conditions worse than concentration camps (cages so small they can't move or even turn around, living in their own shit). If nothing else, what will this teach future superintelligent AIs about how to treat beings less intelligent than itself? About denying them all freedom and agency and exploiting them entirely for its own benefit (or arguably, enjoyment)?
  - bmitc 5y ago
    Do you have any references that show such dialogues? How does GPT-3 handle questions like "why do you feel this way?" or basically any question with "why"?
jjeaff 5y ago
Side note, one of the new Star Trek iterations, "Star Trek Discovery" has a story arc where a data set collected over millenia is integrated into the ship's computer which eventually causes the ship computer to become sentient.
- ben_w 5y ago
  In the Trek universe “the computer is now sentient” seems to happen every few years to every starship even without encountering outside forces. TNG had Moriaty, Wesley Crusher’s nanites, the Exocomps, and whatever it was in season 7’s Emergence.
  - jjeaff 5y ago
    That's true, I guess I didn't consider those as being the computer. Most of those examples were creations using the ships computer. I don't remember the ship computer itself becoming sentient or human. That story ark was usually either Data (tng) or 7 of 9 (Voyager) or Odo (day).
randomsearch 5y ago
That’s kinda a stance I’d like to take. The difficulty with that argument is not the one posted by others - “ no no you don’t understand how clever AI is” - rather the difficulty I have is with the counter argument “how do you know that humans aren’t doing the same kind of thing” Maybe humans are different, maybe not, but I currently don’t have strong arguments to show they’re not and that means we have to take this type of AI seriously. For example: “but you just trained a big pattern matches on a gazillion images and all the Al does is recombine them” - ok, but how do we know human artists aren’t essentially the same? Great writers read a lot. If you hadn’t seen lots of pop art, you’d struggle to do pop art style paintings… even if a baby had perfect motor control, it’d struggle to paint a Victorian rabbit tea party because it would have no examples of a Victorian aesthetic.
- burlesona 5y ago
  I think humans actually are doing the same thing. The world is procedurally generated, and so are our creations :) We also clearly act in pursuit of reward functions. You could go so far as to say that is the definition of “normal” behavior for humans. Learning about how ML works is actually humbling to me, specifically because it reveals how much our skills reduce to pattern recognition. Humans seem much better at learning than machines, ie. we can “figure out” the Victorian aesthetic after a mere handful of examples. What’s not clear is how much we learn this blank-slate as children versus how much encoded / genetic knowledge and memory we have (to what degree is the brain pre-trained?). But I assume we’ll close that gap. So what differences remain to label something as truly intelligent? We have a strong Theory of Mind (ability to reason about what others are thinking and feeling). I think that for us to recognize a machine as genuinely sapient we would expect it to have theory of mind about itself and us. Many animals seem to have this characteristic. It seems unique to humans (so far) is that we have intrinsic motivation as well, and our behaviors are the combination of external and internal motivations. This is where true creativity happens - when one person makes something totally new, not directly derived from prior work. Be that a new musical composition or mathematical theorem. as a species we seek out novelty, experimentation, and “progress.” Now, in our various AI and ML techniques we can and do simulate this with the injection of randomness, and in “genetic programming” we go to lengths to try and emulate that evolutionary model. So maybe there will come a point where the AI is generalized and then it starts having internal motivations that aren’t transparent to us. I don’t know where the line is, but theory of mind and intrinsic motivation seem important, so perhaps there’s something there.
Satam 5y ago
I think you may be underestimating the potential of Dall-E. While it is true that pattern recognition is a key part of what it does, the fact that it can generate images from textual descriptions is a significant step forward. It's not just about art or composition; it's about understanding the meaning of the text and then creating something that reflects that meaning. In that sense, it is quite similar to how a human would approach the task. And oh yeah, gpt generated all the sentences before this one.

Lambdanaut 5y ago

GPT-3 has been able to argue with you about which art styles look best or will be most marketable for over a year now

robbedpeter 5y ago
I don't think a lot of these critics realize how capable gpt-3 is of convincingly outsmarting most people in most contexts. The failures are glaring, and you'll no doubt catch them over time, but transformers are a sea change in ai power. If gpt-3 can meet or beat human performance 80% of the time, that's sufficient for me to call it real intelligence. Not conscious or aware in any meaningful way, but smart, in many of the ways humans are smart. In most online contexts where conversations are singular collisions between individuals who often don't have history, 80% of the real-seeming interactions you have could be gpt-3 output. In a many-shot, recursive, and domain targeted sequence of prompts, the percentage only improves. The recursive, self referential functions learned by these models are not mere extrapolation from Markov chains and static stimulus/response bots. These models are approximating the functions required to produce human level output. They are approximating the functions biological brains use to produce output. Even the naive, single pass probes into gpt capabilities we've seen in the last two years are incredible compared to aiml or even the best bots until ~2018. It's not agi, but it's evidence that agi is feasible, and a definitive milestone on that path. We need to be preparing for takeoff.
- andai 5y ago
  >Not conscious or aware in any meaningful way On what grounds have you determined that GPT-3 has no subjective experiences? More generally, what would need to be the case (with some future AI) for you to say, "ok, yes, this one is definitely conscious" ?
  - tshaddox 5y ago
    I would need some explanation of what consciousness is, for starters. Even better would be an explanation of how I might test whether any thing (including myself or another person) has it.
    - andai 5y ago
      A perspective I thought was interesting is when Jordan Peterson says that the "proof" that people have a divine spark within them is "try treating them as if they don't, and see how well that works out for you!" In that sense it's quite similar to the idea of always assuming a person is acting in good faith: you cannot ever truly know someone's intentions, and even when everyone is communicating honestly and openly, communication is imperfect. But if you start to believe that someone has negative intentions towards you, your body language, tone, words and actions will reflect that assumption, and it will create harmful outcomes where they were not necessary at all. Similarly, we may never be able to objectively prove that a machine is conscious -- as you say, we cannot even prove it to another human about ourselves! But if we keep treating them with the assumption that they lack this "something special" that we have, and treat them unethically as a result, I think we could get ourselves into a lot of trouble. A similar conversation is to be had about the way humans are factory-farming animals, 80 billion of which are born and die every year in conditions worse than concentration camps (cages so small they can't move or even turn around, living in their own shit). If nothing else, what will this teach future superintelligent AIs about how to treat beings less intelligent than itself? About denying them all freedom and agency and exploiting them entirely for its own benefit (or arguably, enjoyment)?
- bmitc 5y ago
  Do you have any references that show such dialogues? How does GPT-3 handle questions like "why do you feel this way?" or basically any question with "why"?

jjeaff 5y ago

Side note, one of the new Star Trek iterations, "Star Trek Discovery" has a story arc where a data set collected over millenia is integrated into the ship's computer which eventually causes the ship computer to become sentient.

ben_w 5y ago
In the Trek universe “the computer is now sentient” seems to happen every few years to every starship even without encountering outside forces. TNG had Moriaty, Wesley Crusher’s nanites, the Exocomps, and whatever it was in season 7’s Emergence.
- jjeaff 5y ago
  That's true, I guess I didn't consider those as being the computer. Most of those examples were creations using the ships computer. I don't remember the ship computer itself becoming sentient or human. That story ark was usually either Data (tng) or 7 of 9 (Voyager) or Odo (day).

randomsearch 5y ago

That’s kinda a stance I’d like to take. The difficulty with that argument is not the one posted by others - “ no no you don’t understand how clever AI is” - rather the difficulty I have is with the counter argument “how do you know that humans aren’t doing the same kind of thing” Maybe humans are different, maybe not, but I currently don’t have strong arguments to show they’re not and that means we have to take this type of AI seriously. For example: “but you just trained a big pattern matches on a gazillion images and all the Al does is recombine them” - ok, but how do we know human artists aren’t essentially the same? Great writers read a lot. If you hadn’t seen lots of pop art, you’d struggle to do pop art style paintings… even if a baby had perfect motor control, it’d struggle to paint a Victorian rabbit tea party because it would have no examples of a Victorian aesthetic.

burlesona 5y ago
I think humans actually are doing the same thing. The world is procedurally generated, and so are our creations :) We also clearly act in pursuit of reward functions. You could go so far as to say that is the definition of “normal” behavior for humans. Learning about how ML works is actually humbling to me, specifically because it reveals how much our skills reduce to pattern recognition. Humans seem much better at learning than machines, ie. we can “figure out” the Victorian aesthetic after a mere handful of examples. What’s not clear is how much we learn this blank-slate as children versus how much encoded / genetic knowledge and memory we have (to what degree is the brain pre-trained?). But I assume we’ll close that gap. So what differences remain to label something as truly intelligent? We have a strong Theory of Mind (ability to reason about what others are thinking and feeling). I think that for us to recognize a machine as genuinely sapient we would expect it to have theory of mind about itself and us. Many animals seem to have this characteristic. It seems unique to humans (so far) is that we have intrinsic motivation as well, and our behaviors are the combination of external and internal motivations. This is where true creativity happens - when one person makes something totally new, not directly derived from prior work. Be that a new musical composition or mathematical theorem. as a species we seek out novelty, experimentation, and “progress.” Now, in our various AI and ML techniques we can and do simulate this with the injection of randomness, and in “genetic programming” we go to lengths to try and emulate that evolutionary model. So maybe there will come a point where the AI is generalized and then it starts having internal motivations that aren’t transparent to us. I don’t know where the line is, but theory of mind and intrinsic motivation seem important, so perhaps there’s something there.

Satam 5y ago

I think you may be underestimating the potential of Dall-E. While it is true that pattern recognition is a key part of what it does, the fact that it can generate images from textual descriptions is a significant step forward. It's not just about art or composition; it's about understanding the meaning of the text and then creating something that reflects that meaning. In that sense, it is quite similar to how a human would approach the task. And oh yeah, gpt generated all the sentences before this one.