4 ms·

It does not attribute no matter whether credit or caution is due. Imagine if Google did this with their index (trained a model and just stopped giving you link

It does not attribute no matter whether credit or caution is due. Imagine if Google did this with their index (trained a model and just stopped giving you links), web masters would be up in arms.

marginalia_nu 4y ago
Most likely it can't, since as I understand it ChatGPT synthesizes a response from the entire training set, rather than simply summarizing any particular document.
- anileated 4y ago
  I don’t think it’s unattainable, intuitively it feels like a couple of extra layers to reflect on where the information might have come from could do it (as well as obviously capturing attribution along with the data). The cynic in me says they would rather keep it ambiguous and not investigate the possibility though. Attribution means everyone can ask for their fair share of the profits made by running the systems trained on the data they supply. Shareholder value and all that.
  - marginalia_nu 4y ago
    Given it's synthetic, there may not be a specific location the information came from. Ask yourself, who told you Russia has invaded Ukraine? Maybe someone did at first, but even then, the notion of a war in Ukraine has been reinforced by thousands of sources since then. It doesn't come from any one of them, but the common picture they all paint. What they'd end up having to do is the very human act of "googling" for sources that agree with what they already believe. Which is not the same.
    - anileated 4y ago
      > Ask yourself, who told you Russia has invaded Ukraine? Rapidly changing events aren’t a great example. > Given it's synthetic, there may not be a specific location the information came from. For bits of sufficiently commonly known information, sure. But if GPT quotes an exact or close enough snippet to something somebody written, and/or part of the topic has been mostly covered in depth by one or two authors, they can be attributed. (Not very unlike a human mind would do it.) I can’t see how attribution is that red line beyond which lies something unattainable by LLM. > What they'd end up having to do is the very human act of "googling" for sources that agree with what they already believe. Which is not the same. Not sure I got this part in the context…

It does not attribute no matter whether credit or caution is due. Imagine if Google did this with their index (trained a model and just stopped giving you links), web masters would be up in arms.

marginalia_nu 4y ago

Most likely it can't, since as I understand it ChatGPT synthesizes a response from the entire training set, rather than simply summarizing any particular document.

anileated 4y ago
I don’t think it’s unattainable, intuitively it feels like a couple of extra layers to reflect on where the information might have come from could do it (as well as obviously capturing attribution along with the data). The cynic in me says they would rather keep it ambiguous and not investigate the possibility though. Attribution means everyone can ask for their fair share of the profits made by running the systems trained on the data they supply. Shareholder value and all that.
- marginalia_nu 4y ago
  Given it's synthetic, there may not be a specific location the information came from. Ask yourself, who told you Russia has invaded Ukraine? Maybe someone did at first, but even then, the notion of a war in Ukraine has been reinforced by thousands of sources since then. It doesn't come from any one of them, but the common picture they all paint. What they'd end up having to do is the very human act of "googling" for sources that agree with what they already believe. Which is not the same.
  - anileated 4y ago
    > Ask yourself, who told you Russia has invaded Ukraine? Rapidly changing events aren’t a great example. > Given it's synthetic, there may not be a specific location the information came from. For bits of sufficiently commonly known information, sure. But if GPT quotes an exact or close enough snippet to something somebody written, and/or part of the topic has been mostly covered in depth by one or two authors, they can be attributed. (Not very unlike a human mind would do it.) I can’t see how attribution is that red line beyond which lies something unattainable by LLM. > What they'd end up having to do is the very human act of "googling" for sources that agree with what they already believe. Which is not the same. Not sure I got this part in the context…