This does quantitatively decrease my objection, yeah. My objection would still be there, somewhat, also quantitatively.
Maybe it’s difficult to write the condensed version that’s just the parts you added while getting most of the same effect, so there’s not a better option. That’s certainly the case with Gwern’s images (and I use image generators for the same reason).
At a wild guess, I’d say that if the useful artifact is literally a paragraph or less, and you’ve gone over it several times, then it could be “ok” as testimony according to me. Like, if the LLM drafted a few sentences, and then you read them and deeply checked “is this really the right way to say this? does this really match my idea / felt sense?”, and then you asked for a bunch of rewrites / rewordings, and did this several times, then plausibly that’s just good.
If it’s longer than a paragraph, then I’d suspect there’s substantial slop that’s slopping in, at various levels of abstraction. IDK.
At a wild guess, I’d say that if the useful artifact is literally a paragraph or less, and you’ve gone over it several times, then it could be “ok” as testimony according to me. Like, if the LLM drafted a few sentences, and then you read them and deeply checked “is this really the right way to say this? does this really match my idea / felt sense?”, and then you asked for a bunch of rewrites / rewordings, and did this several times, then plausibly that’s just good.
Yeah, insofar as I’d endorse publishing LLM text that’d be the minimum, maybe in addition to adding links.
Code feels similar, I often end up deleting a bunch of LLM-generated code because it’s extraneous to my purpose, and this is much more of an issue because I don’t feel like publishing LLM-written text but don’t know how to feel about LLM-written code. I guess a warning at the top telling the reader that they’re about to wade into some-level-of-unedited code is warranted.
This does quantitatively decrease my objection, yeah. My objection would still be there, somewhat, also quantitatively.
Maybe it’s difficult to write the condensed version that’s just the parts you added while getting most of the same effect, so there’s not a better option. That’s certainly the case with Gwern’s images (and I use image generators for the same reason).
At a wild guess, I’d say that if the useful artifact is literally a paragraph or less, and you’ve gone over it several times, then it could be “ok” as testimony according to me. Like, if the LLM drafted a few sentences, and then you read them and deeply checked “is this really the right way to say this? does this really match my idea / felt sense?”, and then you asked for a bunch of rewrites / rewordings, and did this several times, then plausibly that’s just good.
If it’s longer than a paragraph, then I’d suspect there’s substantial slop that’s slopping in, at various levels of abstraction. IDK.
Yeah, insofar as I’d endorse publishing LLM text that’d be the minimum, maybe in addition to adding links.
Code feels similar, I often end up deleting a bunch of LLM-generated code because it’s extraneous to my purpose, and this is much more of an issue because I don’t feel like publishing LLM-written text but don’t know how to feel about LLM-written code. I guess a warning at the top telling the reader that they’re about to wade into some-level-of-unedited code is warranted.