3 ms·

And that's the crux of the matter. Specific words are input into the algorithm, to conjure your code from a probabilistic model. Use a random sentence generator

And that's the crux of the matter. Specific words are input into the algorithm, to conjure your code from a probabilistic model. Use a random sentence generator enough and at some point it'll plagiarise. It's how the human uses it's output is what is important at the end of the day.

palata 3y ago
If you believe that humans have an infinite memory and can optimize to be just on the right side of the plagiarism line, then I suppose it is the same thing. Now my opinion is that humans are not machines, and therefore it makes sense to not apply the same laws. To me, it seems natural to think that a human will never remember every single word of a book they read. Whereas a machine can totally do it. > It's how the human uses it's output is what is important I would take it from the other direction: the AI should be able to prove that it "read" the words like a human (i.e. with a "bad" memory and everything that goes with it).
- theonlybutlet 3y ago
  A computer can remember to do that only if it's programmed to do so. These are probabilistic models,it's basically a large algorithm. They're not copy pasting a database verbatim. They don't have access to it, only a list of parameters (sort of the algorithm weightings) and the list of tokens (translated words). With OpenAI's API you can even see the certainty/probability associated with the next word it quotes. We know for definite whether a model is doing this by looking at it's source code. Humans do also come up with the same ideas from time to time, without any external input.
  - palata 3y ago
    Wait, so you're saying that you can go look at the billions of trained parameters, and tell me for sure that my blog post is not "encoded" there (for any definition of "encoded" that would allow the model to generate text that would reasonably be considered as plagiarism)? > A computer can remember to do that only if it's programmed to do so. Not sure what your point is there. So your belief is that if the programmer did not explicitly put a specific intent into the code, then the computer cannot do anything consistent with that intent? Like nothing unexpected can ever happen with computers, because they are strictly limited to what the programmer explicitly wanted them to do?
    - theonlybutlet 3y ago
      A random word generator given enough time could do that. Given these models have billions (hundreds of billions) of parameters, the odds of it exactly copying your code is less than 100%. Academics when publishing run through plagiarism checkers as it happens even if they're 100% sure they haven't consulted other thesis'. (Usually 20%-30% is acceptable). Some coders run their code through checkers to make sure it's not copyrighted. The onus similarly would be on whoever is using the LLM output to check it before monetizing it. It's just a tool after all.

palata 3y ago

If you believe that humans have an infinite memory and can optimize to be just on the right side of the plagiarism line, then I suppose it is the same thing. Now my opinion is that humans are not machines, and therefore it makes sense to not apply the same laws. To me, it seems natural to think that a human will never remember every single word of a book they read. Whereas a machine can totally do it. > It's how the human uses it's output is what is important I would take it from the other direction: the AI should be able to prove that it "read" the words like a human (i.e. with a "bad" memory and everything that goes with it).

theonlybutlet 3y ago
A computer can remember to do that only if it's programmed to do so. These are probabilistic models,it's basically a large algorithm. They're not copy pasting a database verbatim. They don't have access to it, only a list of parameters (sort of the algorithm weightings) and the list of tokens (translated words). With OpenAI's API you can even see the certainty/probability associated with the next word it quotes. We know for definite whether a model is doing this by looking at it's source code. Humans do also come up with the same ideas from time to time, without any external input.
- palata 3y ago
  Wait, so you're saying that you can go look at the billions of trained parameters, and tell me for sure that my blog post is not "encoded" there (for any definition of "encoded" that would allow the model to generate text that would reasonably be considered as plagiarism)? > A computer can remember to do that only if it's programmed to do so. Not sure what your point is there. So your belief is that if the programmer did not explicitly put a specific intent into the code, then the computer cannot do anything consistent with that intent? Like nothing unexpected can ever happen with computers, because they are strictly limited to what the programmer explicitly wanted them to do?
  - theonlybutlet 3y ago
    A random word generator given enough time could do that. Given these models have billions (hundreds of billions) of parameters, the odds of it exactly copying your code is less than 100%. Academics when publishing run through plagiarism checkers as it happens even if they're 100% sure they haven't consulted other thesis'. (Usually 20%-30% is acceptable). Some coders run their code through checkers to make sure it's not copyrighted. The onus similarly would be on whoever is using the LLM output to check it before monetizing it. It's just a tool after all.