5 ms·

You took that quote out of context and missed the broader point in the process. The snippets provided in regular search results cannot generally replace the sub

You took that quote out of context and missed the broader point in the process. The snippets provided in regular search results cannot generally replace the substance of the full articles they link to, while that's the whole point of GP's hypothetical website—it simply doesn't reproduce large chunks of text verbatim, presumably to avoid copyright infringement claims in the hypothetical's frame, and in GP's rhetorical frame to present an analogy with the information-laundering powers of LLMs that their creators claim make their exploitation of unlicensed training data fair use. The whole point of a search engine (as we've classically known them) is to index the web and respond to queries with a list of links that you will inspect and click through on. The whole point of an LLM chatbot tool is to eliminate those inspecting and clicking-through steps, becoming a one-stop shop for content whose substance was created by someone else. That's also the whole point of GP's hypothetical, which is why it works as an analogy. --- There are substantially better arguments for search engines being legitimate fair use. Consider, for example, transformation. AI defenders will argue that these systems are transformative because they reshuffle elements of their input in their output, but that's clearly a much weaker form of transformation than one in which the transformed work has an entirely different nature and purpose, i.e. search engines vs. the results they return. Ultimately these technicality-based "nuh uh" arguments aren't going to save the practice of training AI on unlicensed data, because they are incompatible with the spirit of copyright law even if the novel nature of these technologies means the letter of said law can't quite nail them down yet. If these arguments do succeed, it will be because the judicial/regulatory environment in which they were applied has been corrupted by capital.

8note 3y ago
A search engine takes an input string, a corpus of text, and returns a series of text that best comes next after the input string. An LLM takes an input string a corpus of text, and returns a series of text that best comes after the next input string. To get a paragraph of output, you run the search over and over again Both the search and LLM reshuffle the inputs to the outputs. If I'm describing the purpose of the LLM, it's got a wide number of usages. "Making my resume look more professional" or "be a crud api" or "reformat my ask into a api call to X service" or "give me a timeline of events surrounding Y with source links"
- caconym_ 3y ago
  An 18 wheeler travels on wheels. A shopping cart travels on wheels. A shopping cart does not require a license to operate, therefore an 18 wheeler does not require a license to operate. A shopping cart can be operated inside a grocery store, therefore an 18 wheeler can be operated inside a grocery store. A child can operate a shopping cart, therefore a child can operate an 18 wheeler. If I'm describing the purpose of an 18 wheeler, it's got a wide number of usages. "Carry my chicken" or "carry my lettuce" or "carry my Cheetos". Or, simply, "carry my groceries".
fennecbutt 3y ago
>content whose substance was created by someone else And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance. You think all fantasy writers gotta pay Tolkien estate bc so much of fantasy draws from his tropes? Lmao no.
- caconym_ 3y ago
  > And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance. If training data is so unimportant, why not simply not use it and avoid the controversy? At the very least that would certainly fix the issue where the model demonstrates how "inspired" it is by NYT articles by reproducing them verbatim. :)

8note 3y ago

A search engine takes an input string, a corpus of text, and returns a series of text that best comes next after the input string. An LLM takes an input string a corpus of text, and returns a series of text that best comes after the next input string. To get a paragraph of output, you run the search over and over again Both the search and LLM reshuffle the inputs to the outputs. If I'm describing the purpose of the LLM, it's got a wide number of usages. "Making my resume look more professional" or "be a crud api" or "reformat my ask into a api call to X service" or "give me a timeline of events surrounding Y with source links"

caconym_ 3y ago
An 18 wheeler travels on wheels. A shopping cart travels on wheels. A shopping cart does not require a license to operate, therefore an 18 wheeler does not require a license to operate. A shopping cart can be operated inside a grocery store, therefore an 18 wheeler can be operated inside a grocery store. A child can operate a shopping cart, therefore a child can operate an 18 wheeler. If I'm describing the purpose of an 18 wheeler, it's got a wide number of usages. "Carry my chicken" or "carry my lettuce" or "carry my Cheetos". Or, simply, "carry my groceries".

fennecbutt 3y ago

>content whose substance was created by someone else And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance. You think all fantasy writers gotta pay Tolkien estate bc so much of fantasy draws from his tropes? Lmao no.

caconym_ 3y ago
> And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance. If training data is so unimportant, why not simply not use it and avoid the controversy? At the very least that would certainly fix the issue where the model demonstrates how "inspired" it is by NYT articles by reproducing them verbatim. :)