I think that we should have revision control for intermediate stages - for code, documents, even paintings. So we can at least have some idea of provenance, how it's made.
1) For things made with LLMs:
1a) The fact that older versions aren't online forever. You literally might never be able to put the original prompt in and get the same result.
1b) A minor change in input prompt can result in a huge output change, rendering the original prompt practically meaningless, especially if modifications were required for the output of the LLM.
2) For things made the old-fashioned way, most history is boring and not useful. The best git repos have carefully curated history, with cohesive change sets that are both readable, and usable when bisecting the commit history for regressions.
And I don’t care if it’s boring, it has to be available. Crime scene details or forgery details are mundane and boring too, but for the investigators they are essential.
1) For things made with LLMs: 1a) The fact that older versions aren't online forever. You literally might never be able to put the original prompt in and get the same result. 1b) A minor change in input prompt can result in a huge output change, rendering the original prompt practically meaningless, especially if modifications were required for the output of the LLM.
2) For things made the old-fashioned way, most history is boring and not useful. The best git repos have carefully curated history, with cohesive change sets that are both readable, and usable when bisecting the commit history for regressions.