3 ms·

How is suing OpenAI different from suing an army of gifted individuals who can speed-read and who read most NYT articles and create responses based on them? As

How is suing OpenAI different from suing an army of gifted individuals who can speed-read and who read most NYT articles and create responses based on them? Assuming there is no plagiarism, of course.

microtonal 3y ago
Because larger LLMs can reproduce text verbatim, speed readers probably not: https://nitter.net/maksym_andr/status/1740776900786626608 https://nitter.net/maksym_andr/status/1740776900786626608
- pmarreck 3y ago
  So what this tells me is that if you make a carefully constructed prompt and change the temperature to zero (not an option normally available), and know the exact article title and perhaps the beginning of the first paragraph (which is pretty difficult to do unless you already basically have access to the full article), you can potentially get verbatim articles back out. Great, so all those paywalls that use verbatim subjects and early texts as a teaser are now broken! Well, for all old articles that the LLM was trained on, at least, I guess. Very simple countermeasure is to simply use AI to paraphrase the subject and article (paradoxically) until the person pays for the privilege of reading human-authored text. I'm sorry but this is a very weak argument. For example, I don't even believe normal users have access to the GPT4 system prompt unless they use the API directly (and possibly not even then, I'd have to check).
eviks 3y ago
One is real, another one is made up (and there is plagiarism, of course, many pages of examples of that)
- pmarreck 3y ago
  The temperature setting controls randomness output of the LLM. We paraphrase all the time to avoid plagiarism and that's just somewhat randomized retelling of the same idea. If you set the temperature to 0 in an LLM it's basically in "decompress/rote mode". I don't think this is qualitatively the same as "copying", possibly more akin to "memorization". I haven't seen very many demonstrations of verbatim-copy output that wasn't done with a temperature of or near 0.
  - eviks 3y ago
    How is the fact that there is a flag to disable plagiarism relevant for the issue that there is plagiarism?
    - pmarreck 3y ago
      Because enabling "plagiarism mode" is a conscious action that a human takes, it does not default to "plagiarize" no more than a machine that has simply stored the verbatim copy of an article, when asked to print it out, is "plagiarizing". Plus, citations are showing up in LLM's now.
  - eviks 3y ago
    Also, you can't avoid plagiarism by paraphrasing because paraphrasing is a form of plagiarism. The key here is whether you cite the source, which the model doesn't https://www.scribbr.com/frequently-asked-questions/is-paraphrasing-considered-plagiarism/ https://www.scribbr.com/frequently-asked-questions/is-paraph...
    - pmarreck 3y ago
      that's coming. some LLM's are already starting to do that

microtonal 3y ago

Because larger LLMs can reproduce text verbatim, speed readers probably not: https://nitter.net/maksym_andr/status/1740776900786626608 https://nitter.net/maksym_andr/status/1740776900786626608

pmarreck 3y ago
So what this tells me is that if you make a carefully constructed prompt and change the temperature to zero (not an option normally available), and know the exact article title and perhaps the beginning of the first paragraph (which is pretty difficult to do unless you already basically have access to the full article), you can potentially get verbatim articles back out. Great, so all those paywalls that use verbatim subjects and early texts as a teaser are now broken! Well, for all old articles that the LLM was trained on, at least, I guess. Very simple countermeasure is to simply use AI to paraphrase the subject and article (paradoxically) until the person pays for the privilege of reading human-authored text. I'm sorry but this is a very weak argument. For example, I don't even believe normal users have access to the GPT4 system prompt unless they use the API directly (and possibly not even then, I'd have to check).

eviks 3y ago

One is real, another one is made up (and there is plagiarism, of course, many pages of examples of that)

pmarreck 3y ago
The temperature setting controls randomness output of the LLM. We paraphrase all the time to avoid plagiarism and that's just somewhat randomized retelling of the same idea. If you set the temperature to 0 in an LLM it's basically in "decompress/rote mode". I don't think this is qualitatively the same as "copying", possibly more akin to "memorization". I haven't seen very many demonstrations of verbatim-copy output that wasn't done with a temperature of or near 0.
- eviks 3y ago
  How is the fact that there is a flag to disable plagiarism relevant for the issue that there is plagiarism?
  - pmarreck 3y ago
    Because enabling "plagiarism mode" is a conscious action that a human takes, it does not default to "plagiarize" no more than a machine that has simply stored the verbatim copy of an article, when asked to print it out, is "plagiarizing". Plus, citations are showing up in LLM's now.
- eviks 3y ago
  Also, you can't avoid plagiarism by paraphrasing because paraphrasing is a form of plagiarism. The key here is whether you cite the source, which the model doesn't https://www.scribbr.com/frequently-asked-questions/is-paraphrasing-considered-plagiarism/ https://www.scribbr.com/frequently-asked-questions/is-paraph...
  - pmarreck 3y ago
    that's coming. some LLM's are already starting to do that