6 ms·

> but all it does is regurgitate existing works I am genuinely intrigued by this point of view and so would love to hear people who hold it's reasoning. Over

by jah242 4y ago

> but all it does is regurgitate existing works I am genuinely intrigued by this point of view and so would love to hear people who hold it's reasoning. Over the last few days I've seen hundreds of poems and stories from ones about climate change in the style of a sonnet to peanut butter sandwiches getting stuck in toasters in the style of the bible. I even asked it to make a text adventure game for me to play where I could put in any instruction, leading to a unique series of events and narrative. Is the claim that these were all simply copy and pastes of something on the internet in their entirety? And that as such the internet already seems to contain essentially every permutation of everything I could ask ChatGPT, as to me this sounds highly implausible. If the claim is that whilst these are not direct copy pastes, it is essentially a remix of lots of different things people have said before on the internet repurposed to a different end, is that not literally just what language is? Humans use common sayings, idioms, slang and phrases all the time, never mind the 'tropes' and story lines that are reused constantly. Coders use common patterns and styles and copy from stackoverflow. In fact language literally only works because we all share it and share the meaning of it. If we are saying that all ChatGPT does is remix existing language and phrases to a new purpose... to me we are saying ChatGPT does the same thing as humans. Any thoughts would be appreciated.

DebtDeflation 4y ago
In some sense, every AI tool is "regurgitating its training set". In the case of NLP, that training set will be a corpus of documents, utterance-intent pairs, lists of Named Entities, etc.
sarchertech 4y ago
If you spend enough time with it you can quickly start to see the building blocks. More than just the similarities you see in several works by the same person. If you’ve ever played a video game where the levels are programmatically generated, you’ll eventually notice the modular nature of everything. It becomes very noticeable and it’s different from noticing natural patterns that arise in a completely handcrafted level. GPT is exactly like that. You won’t notice it with short output. But ask it to keep going and write the next verse of that sonnet. Keep going and it gets VERY repetitive in a way that a human poet doesn’t. Also keep in mind that the examples you’re seeing are a form of sample bias. You’re seeing output curated by humans.
- consumer451 4y ago
  But isn’t it possible that humans are “just” doing the same but with a much higher foundational resolution?
  - sarchertech 4y ago
    Sure it’s possible. I mean every one of us is a unique arrangement of the same fundamental building blocks. But there’s no reason to assume that something works the same way we do just because someone named it a neural network.
    - consumer451 4y ago
      The key concept here for me is pattern recognition. It lines up with my experience when trying to practice creativity. For example, for decades it has been my opinion that my own supposedly higher than average pattern matching skills were the main thing I brought to the table at work. When I play drums, keys, or guitar, coming up with new patterns and styles to play feels magical when it happens in real-time, but if I break it down later, I almost always see the influences.
  - randcraw 4y ago
    In judging GPT, we should be drawing a line in the sand between a poor writer (e.g. a college student) and an excellent one (a professional, ideally a domain expert). GPT is impressing us now only because we have very low expectations for a computer to write like a human. But when judged against a capable human, especially when you focus on the _information_ that's conveyed, GPT sucks. Few of us would want to read a book written as flaccidly as GPT does, a wandering narrative that cannot connect the semantics of one sentence to the next. GPT is capable only of writing _sets_ of sentences, not a progression of them where a concept is introduced and subsequent text tests and contrasts a convincing set of complements and alternatives. Given LLM architectures (and inherent black boxes), that's likely to never change since they don't understand logic or dependency, sufficiency, or necessity. Without those skills, no LLM will ever be able to _develop_ an idea or expand on the implications proposed in an introduction. Have you ever seen GPT conclude a paragraph with a bang? And you never will. You'll note too that GPT has not attempted to write dialog. I think there's good reason for that. Talk strips away all the excess verbiage. Humans say only what's essential to convey our message. Pointless blather that leads nowhere is maddening to listen to for long. But without a clear focus, that's exactly what I'd expect from GPT: wandering speech that's totally disengaging. Surprising the reader simply is not in its DNA. So if you're happy reading narratives that wander, that don't lead anywhere, with characters lacking personality, and a narrative style lacking the essentials of every engaging tale (exposition, conflict, suspense, climax, resolution)... then perhaps GPT is for you. But if that's GPT's great promise -- to generate businessspeak that's "tolerable" but even less engaging than the marketing schlock that immerses us now -- then thanks, but no thanks.
453qtgreq 4y ago
>Is the claim that these were all simply copy and pastes of something on the internet in their entirety? And that as such the internet already seems to contain essentially every permutation of everything I could ask ChatGPT, as to me this sounds highly implausible. It's the training data supplied to GPT3 (as explained by OpenAI themselves), so yes, it is literally true. You are just seeing snippets of the internet, re-formed and regurgitated. It can only do what you ask.
- jah242 4y ago
  So I appreciate the jist of your point but the way these models work is rather more complicated that copying and pasting snippets and so it certainly is not 'literally' true. The models are trained to predict sub-word level tokens from the internet training dataset, so the level of re-formation and regurgitation in a generated sentence can be vast, to the point of final sentence being novel it's own right.
- ben_w 4y ago
  100% of the training data of the salty jelly in a human skull, is a sensory input. Were your argument flawless and your conclusion correct, then all human creativity would "literally" be a remix of things in the natural world, as even when we remix things made by other humans that too would ultimately derive back to nature. This can certainly be asserted depending on how you wish to use those words, but just as it is not useful at predicting what our abilities are — for example, a perfect intellect could predict quantum mechanics from scratch by watching a camp fire die down on a rainy night, but we didn't do that in one step in the neolithic — so too this isn't useful at telling us what the limits of GPT-family LLMs might be, as that which has been built on such inputs, both in the case of humans and this particular AI, greatly exceeds the imagination of any single individual.
vba616 4y ago
>In fact language literally only works because we all share it and share the meaning of it. Most human communication is to some extent adversarial and constantly subverts the assumption of shared meaning on purpose or by accident. It's extremely difficult to "solve" because cooperative communication often appears to subvert the basic principles too. Have you ever read about Grice's maxims of conversation? Once you think in those terms, the frequent absence of cooperation in most human conversation is obvious. https://en.wikipedia.org/wiki/Cooperative_principle https://en.wikipedia.org/wiki/Cooperative_principle

by jah242 4y ago

DebtDeflation 4y ago

In some sense, every AI tool is "regurgitating its training set". In the case of NLP, that training set will be a corpus of documents, utterance-intent pairs, lists of Named Entities, etc.

sarchertech 4y ago

If you spend enough time with it you can quickly start to see the building blocks. More than just the similarities you see in several works by the same person. If you’ve ever played a video game where the levels are programmatically generated, you’ll eventually notice the modular nature of everything. It becomes very noticeable and it’s different from noticing natural patterns that arise in a completely handcrafted level. GPT is exactly like that. You won’t notice it with short output. But ask it to keep going and write the next verse of that sonnet. Keep going and it gets VERY repetitive in a way that a human poet doesn’t. Also keep in mind that the examples you’re seeing are a form of sample bias. You’re seeing output curated by humans.

consumer451 4y ago
But isn’t it possible that humans are “just” doing the same but with a much higher foundational resolution?
- sarchertech 4y ago
  Sure it’s possible. I mean every one of us is a unique arrangement of the same fundamental building blocks. But there’s no reason to assume that something works the same way we do just because someone named it a neural network.
  - consumer451 4y ago
    The key concept here for me is pattern recognition. It lines up with my experience when trying to practice creativity. For example, for decades it has been my opinion that my own supposedly higher than average pattern matching skills were the main thing I brought to the table at work. When I play drums, keys, or guitar, coming up with new patterns and styles to play feels magical when it happens in real-time, but if I break it down later, I almost always see the influences.
- randcraw 4y ago
  In judging GPT, we should be drawing a line in the sand between a poor writer (e.g. a college student) and an excellent one (a professional, ideally a domain expert). GPT is impressing us now only because we have very low expectations for a computer to write like a human. But when judged against a capable human, especially when you focus on the _information_ that's conveyed, GPT sucks. Few of us would want to read a book written as flaccidly as GPT does, a wandering narrative that cannot connect the semantics of one sentence to the next. GPT is capable only of writing _sets_ of sentences, not a progression of them where a concept is introduced and subsequent text tests and contrasts a convincing set of complements and alternatives. Given LLM architectures (and inherent black boxes), that's likely to never change since they don't understand logic or dependency, sufficiency, or necessity. Without those skills, no LLM will ever be able to _develop_ an idea or expand on the implications proposed in an introduction. Have you ever seen GPT conclude a paragraph with a bang? And you never will. You'll note too that GPT has not attempted to write dialog. I think there's good reason for that. Talk strips away all the excess verbiage. Humans say only what's essential to convey our message. Pointless blather that leads nowhere is maddening to listen to for long. But without a clear focus, that's exactly what I'd expect from GPT: wandering speech that's totally disengaging. Surprising the reader simply is not in its DNA. So if you're happy reading narratives that wander, that don't lead anywhere, with characters lacking personality, and a narrative style lacking the essentials of every engaging tale (exposition, conflict, suspense, climax, resolution)... then perhaps GPT is for you. But if that's GPT's great promise -- to generate businessspeak that's "tolerable" but even less engaging than the marketing schlock that immerses us now -- then thanks, but no thanks.

453qtgreq 4y ago

>Is the claim that these were all simply copy and pastes of something on the internet in their entirety? And that as such the internet already seems to contain essentially every permutation of everything I could ask ChatGPT, as to me this sounds highly implausible. It's the training data supplied to GPT3 (as explained by OpenAI themselves), so yes, it is literally true. You are just seeing snippets of the internet, re-formed and regurgitated. It can only do what you ask.

jah242 4y ago
So I appreciate the jist of your point but the way these models work is rather more complicated that copying and pasting snippets and so it certainly is not 'literally' true. The models are trained to predict sub-word level tokens from the internet training dataset, so the level of re-formation and regurgitation in a generated sentence can be vast, to the point of final sentence being novel it's own right.
ben_w 4y ago
100% of the training data of the salty jelly in a human skull, is a sensory input. Were your argument flawless and your conclusion correct, then all human creativity would "literally" be a remix of things in the natural world, as even when we remix things made by other humans that too would ultimately derive back to nature. This can certainly be asserted depending on how you wish to use those words, but just as it is not useful at predicting what our abilities are — for example, a perfect intellect could predict quantum mechanics from scratch by watching a camp fire die down on a rainy night, but we didn't do that in one step in the neolithic — so too this isn't useful at telling us what the limits of GPT-family LLMs might be, as that which has been built on such inputs, both in the case of humans and this particular AI, greatly exceeds the imagination of any single individual.

vba616 4y ago

>In fact language literally only works because we all share it and share the meaning of it. Most human communication is to some extent adversarial and constantly subverts the assumption of shared meaning on purpose or by accident. It's extremely difficult to "solve" because cooperative communication often appears to subvert the basic principles too. Have you ever read about Grice's maxims of conversation? Once you think in those terms, the frequent absence of cooperation in most human conversation is obvious. https://en.wikipedia.org/wiki/Cooperative_principle https://en.wikipedia.org/wiki/Cooperative_principle