6 ms·

I asked chatGPT to summarize this like I was a child: > Imagine you have a toy that you're not sure will work perfectly or not. The people who made the toy say

I asked chatGPT to summarize this like I was a child: > Imagine you have a toy that you're not sure will work perfectly or not. The people who made the toy say that they're not sure either and that you have to take the risk of using it. They also say that if the toy doesn't work well or isn't accurate, it's not their fault, and they won't promise to fix it or make it better. They also can't promise that the results you get from using the toy will be correct. So, you have to be careful and use it at your own risk!

seanw444 3y ago
Warning: this product might be the polar opposite of what you want.
- ChuckNorris89 3y ago
  But we'll take your money anyway.
j1elo 3y ago
If ChatGPT is just a glorified autocomplete engine, which doesn't know the context or the meaning of what is being said (as I've read in random discussions here on HN), then how come it is able to process the license text and summarize it in those terms? I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. So, does chatGPT really know what it's writing about, or not?
- api 3y ago
  A lot of these questions hinge on words like “think” and “understand” for which we do not have precise definitions. It’s really kind of calling bullshit on a lot of the terminology we use to talk about thinking, revealing that most of it is hand wavey anthropocentrism. It does what it does. We do not know what that means. We do not really know how close what it is doing is to what we do. It’s also possible that it’s doing similar things to what we do but in some very different way. “Neural network” has always been a term to make biologists cringe as they’re really only the loosest of analogies to how brains work.
  - xp84 3y ago
    So true! We basically define those words almost self-referentially. Understanding is what human brains do, so by definition nothing else but a human can 'understand' complex things (we say). We make only slight allowances for animals, allowing that they can "understand" but only simple things. As this tech gets better, those who cling to that definition are going to get more and more uncomfortable, and tie themselves into more and more logical knots trying desperately to explain why it "doesn't count."
    - lxgr 3y ago
      A similar thing has happened (and arguably is still happening) when it comes to animal cognition. "The only species capable of recognizing itself in the mirror. Ok, recognizing itself in the mirror and object permanence. Ok, all of that and mourning its dead. Ok, hang on for a minute..."
- andrewstuart2 3y ago
  The words it autocompleted were the most likely words (chains of word fragments) to come next in a theoretical conversation when the question as context was taken into consideration, given the rest of the context it's been trained on up to EOY 2021.
  - vonnieda 3y ago
    I wonder what an LLM would say if it had, say, 2 orders of magnitude more parameters and 46 years of training like I do. And if it was trained up to the millisecond with new inputs from millions of individual sensors. It might say: > I wonder what an LLM would say if it had...
- dylan604 3y ago
  it'll be funny to find some unheard of website where someone wrote something very similar in blog or reply in comment. maybe even in a usenet chat archive or something?
- esafak 3y ago
  We are "glorified" autocomplete engines too. The glory is in the fact that we have rich sensors, a memory, and a world model underpinning the engine. ChatGPT has a complex model, if not as complex as ours, and it is good enough to perform such tasks. As these models improve, they will compete with and outperform us on more and more tasks, at which point we will stop comparing our abilities to execute tasks. "Sure, it can do everything, but it's not human."
  - acchow 3y ago
    > We are "glorified" autocomplete engines too This doesn’t feel intuitive to me or convincing. What evidence do you have to support such a claim? I certainly don’t feel like I’m an autocomplete engine. Was Gödel also an autocomplete engine? How could an autocomplete engine come up with novel ideas like the incompleteness theorems and understand them?
    - esafak 3y ago
      It's the "intelligence is compression/prediction" paradigm, which is well known in the field, if you want to read about it.
    - flangola7 3y ago
      >How could an autocomplete engine come up with novel ideas like the incompleteness theorems and understand them? Increase the temperature to above 0.
      - acchow 3y ago
        Sure, you can create new sentences. But you can't recognize which ones are "true" and which are not.
        thewataccount 3y ago
        Neither can we. We have a lot of theorems ourselves that exist because they are what we consider the most likely solution to either extrapolating existing rules or modeling a set of measurements - but it's unknown as to whether the theory is absolutely true. a child might see that 1 + 2 = 2 + 1 4 * 3 = 4 * 3 7 + 7 = 7 + 7 They might predict this may apply elsewhere, and that: 8 + 0 = 0 + 8 This is the Commutative Law, which the child can predict exists, even if they don't have the specialty to write proofs and understand the theory, etc. They don't know that this is true, but based on their observations this is a theory that holds true to this situation.
- _a_a_a_ 3y ago
  > I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. well IMO it did, so it does. So yes, it 'understands' by your predicate.
- nr2x 3y ago
  A lot of human “decision making” literally happens after some pattern recognition system in your brain makes the decisions. We are all just autocompleting all day. Does ChatGPT “think”? Maybe? Do we “think” as much as we think we do? Maybe not? https://www.psychologytoday.com/us/blog/unconscious-branding/202012/our-brains-make-our-minds-we-know-it https://www.psychologytoday.com/us/blog/unconscious-branding... https://www.nature.com/articles/news.2008.751 https://www.nature.com/articles/news.2008.751
- sangnoir 3y ago
  > I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. Have you interacted with Markov chains before? They are surprisingly coherent and how they work is easier to grok (or implement as a weekend project)
- buggyipadmettoo 3y ago
  I was wondering the same and asked gpt to explain it to me the other day. Basically it’s all one process, there aren’t 2 modules (understand and generate). The “understanding” is perhaps like the generating in reverse. Rather than generating the next word from a statistical process, it’s defining the statistical process from the next words of the input. But as I explain here, even reviewing that conversation we had, I think maybe gpt and I were both hallucinating and feigning understanding. Cuz I still don’t get it. It still seems far too magical to fit this explanation. So my next best guess: we’re a simulation, GPT is magic, and our creators are just having fun introducing more and more magic into the system.
- matrix_overload 3y ago
  Because it can instantly go through countless examples where people were summarizing things on the forums, paper abstracts, documents, etc., and blend it together with a 99% accuracy. Except for that 1% where it will look totally plausible, but will be completely wrong.
codeflo 3y ago
When will “I asked ChatGPT” stop being posted in response to every other HN comment?
- CatWChainsaw 3y ago
  Never, at this rate.
- sieabahlpark 3y ago
  [dead]

seanw444 3y ago

Warning: this product might be the polar opposite of what you want.

ChuckNorris89 3y ago
But we'll take your money anyway.

j1elo 3y ago

If ChatGPT is just a glorified autocomplete engine, which doesn't know the context or the meaning of what is being said (as I've read in random discussions here on HN), then how come it is able to process the license text and summarize it in those terms? I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. So, does chatGPT really know what it's writing about, or not?

api 3y ago
A lot of these questions hinge on words like “think” and “understand” for which we do not have precise definitions. It’s really kind of calling bullshit on a lot of the terminology we use to talk about thinking, revealing that most of it is hand wavey anthropocentrism. It does what it does. We do not know what that means. We do not really know how close what it is doing is to what we do. It’s also possible that it’s doing similar things to what we do but in some very different way. “Neural network” has always been a term to make biologists cringe as they’re really only the loosest of analogies to how brains work.
- xp84 3y ago
  So true! We basically define those words almost self-referentially. Understanding is what human brains do, so by definition nothing else but a human can 'understand' complex things (we say). We make only slight allowances for animals, allowing that they can "understand" but only simple things. As this tech gets better, those who cling to that definition are going to get more and more uncomfortable, and tie themselves into more and more logical knots trying desperately to explain why it "doesn't count."
  - lxgr 3y ago
    A similar thing has happened (and arguably is still happening) when it comes to animal cognition. "The only species capable of recognizing itself in the mirror. Ok, recognizing itself in the mirror and object permanence. Ok, all of that and mourning its dead. Ok, hang on for a minute..."
andrewstuart2 3y ago
The words it autocompleted were the most likely words (chains of word fragments) to come next in a theoretical conversation when the question as context was taken into consideration, given the rest of the context it's been trained on up to EOY 2021.
- vonnieda 3y ago
  I wonder what an LLM would say if it had, say, 2 orders of magnitude more parameters and 46 years of training like I do. And if it was trained up to the millisecond with new inputs from millions of individual sensors. It might say: > I wonder what an LLM would say if it had...
dylan604 3y ago
it'll be funny to find some unheard of website where someone wrote something very similar in blog or reply in comment. maybe even in a usenet chat archive or something?
esafak 3y ago
We are "glorified" autocomplete engines too. The glory is in the fact that we have rich sensors, a memory, and a world model underpinning the engine. ChatGPT has a complex model, if not as complex as ours, and it is good enough to perform such tasks. As these models improve, they will compete with and outperform us on more and more tasks, at which point we will stop comparing our abilities to execute tasks. "Sure, it can do everything, but it's not human."
- acchow 3y ago
  > We are "glorified" autocomplete engines too This doesn’t feel intuitive to me or convincing. What evidence do you have to support such a claim? I certainly don’t feel like I’m an autocomplete engine. Was Gödel also an autocomplete engine? How could an autocomplete engine come up with novel ideas like the incompleteness theorems and understand them?
  - esafak 3y ago
    It's the "intelligence is compression/prediction" paradigm, which is well known in the field, if you want to read about it.
  - flangola7 3y ago
    >How could an autocomplete engine come up with novel ideas like the incompleteness theorems and understand them? Increase the temperature to above 0.
    - acchow 3y ago
      Sure, you can create new sentences. But you can't recognize which ones are "true" and which are not.
      - thewataccount 3y ago
        Neither can we. We have a lot of theorems ourselves that exist because they are what we consider the most likely solution to either extrapolating existing rules or modeling a set of measurements - but it's unknown as to whether the theory is absolutely true. a child might see that 1 + 2 = 2 + 1 4 * 3 = 4 * 3 7 + 7 = 7 + 7 They might predict this may apply elsewhere, and that: 8 + 0 = 0 + 8 This is the Commutative Law, which the child can predict exists, even if they don't have the specialty to write proofs and understand the theory, etc. They don't know that this is true, but based on their observations this is a theory that holds true to this situation.
_a_a_a_ 3y ago
> I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. well IMO it did, so it does. So yes, it 'understands' by your predicate.
nr2x 3y ago
A lot of human “decision making” literally happens after some pattern recognition system in your brain makes the decisions. We are all just autocompleting all day. Does ChatGPT “think”? Maybe? Do we “think” as much as we think we do? Maybe not? https://www.psychologytoday.com/us/blog/unconscious-branding/202012/our-brains-make-our-minds-we-know-it https://www.psychologytoday.com/us/blog/unconscious-branding... https://www.nature.com/articles/news.2008.751 https://www.nature.com/articles/news.2008.751
sangnoir 3y ago
> I would think that it would be needed to "understand" the license text in order to produce that easier to understand version. Have you interacted with Markov chains before? They are surprisingly coherent and how they work is easier to grok (or implement as a weekend project)
buggyipadmettoo 3y ago
I was wondering the same and asked gpt to explain it to me the other day. Basically it’s all one process, there aren’t 2 modules (understand and generate). The “understanding” is perhaps like the generating in reverse. Rather than generating the next word from a statistical process, it’s defining the statistical process from the next words of the input. But as I explain here, even reviewing that conversation we had, I think maybe gpt and I were both hallucinating and feigning understanding. Cuz I still don’t get it. It still seems far too magical to fit this explanation. So my next best guess: we’re a simulation, GPT is magic, and our creators are just having fun introducing more and more magic into the system.
matrix_overload 3y ago
Because it can instantly go through countless examples where people were summarizing things on the forums, paper abstracts, documents, etc., and blend it together with a 99% accuracy. Except for that 1% where it will look totally plausible, but will be completely wrong.

codeflo 3y ago

When will “I asked ChatGPT” stop being posted in response to every other HN comment?

CatWChainsaw 3y ago
Never, at this rate.
sieabahlpark 3y ago
[dead]