Usually, when I’m sharing LLM generated text, it’s to demonstrate some observed property of LLMs, not to make some other claim about the world.
It’s not, “This claim about the real world is true, because this LLM said it” — that’s an invalid deduction. It’s “This claim about a particular LLM is true, because here’s evidence of it doing the thing.”
We get into more interesting territory when an LLM suggests some thing about the world, and I verify that its argument is sound. How should we credit that? It’s not true because the LLM said it, it’s true because you can verify it. But perhaps we don’t want to take credit for coming up with it ourselves.
Yeah I think it makes sense to say your sources. I not infrequently tell someone “my gippities say X but I didn’t verify” or “my gippities say X and I googled a bit and seems right” or “my gippities gave me list of Zs and i check and half were wrong but these ones seem right”.
Usually, when I’m sharing LLM generated text, it’s to demonstrate some observed property of LLMs, not to make some other claim about the world.
It’s not, “This claim about the real world is true, because this LLM said it” — that’s an invalid deduction. It’s “This claim about a particular LLM is true, because here’s evidence of it doing the thing.”
We get into more interesting territory when an LLM suggests some thing about the world, and I verify that its argument is sound. How should we credit that? It’s not true because the LLM said it, it’s true because you can verify it. But perhaps we don’t want to take credit for coming up with it ourselves.
Yeah I think it makes sense to say your sources. I not infrequently tell someone “my gippities say X but I didn’t verify” or “my gippities say X and I googled a bit and seems right” or “my gippities gave me list of Zs and i check and half were wrong but these ones seem right”.