Interesting article on a possible theory for effective LLM prompt generation (h/t Andrej Karpathy) https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluigi-effect-mega-post
@twitskeptic this definitely gets into typical LW silliness a bit, but it's a great run-down of some prompt engineering issues that I think is somewhat accessible to that not already in AI.
I think a succinct summary of pre-prompt negation strategies might be that they recontextualise the pre-prompt to promote continuations whose likelihood given the pre-prompt is high not because they are texts which include similar ideas contiguously with the pre-prompt but because they include the ideas of the pre-prompt contextualized differently (e.g., sarcastically, or as an aside, or as a counterexample, etc.).
@twitskeptic
People keep finding variations on the same thing over and over?
https://www.nature.com/articles/s41562-022-01316-8
When this was published, it was already known among anyone who has played with word vectors for a sufficient time, but it was nice to get data confirming it.