Ah, then let me rephrase, it's actually this model: > P(A|B,C,D,E,F....) And with clever choices of B,C,D.... you can make A abitarily probable. Eg., Suppose

Ah, then let me rephrase, it's actually this model: > P(A|B,C,D,E,F....) And with clever choices of B,C,D.... you can make A abitarily probable. Eg., Suppose, 'lolita' were rare, well then choose: B=Library, C=Author, D=1955, E=... Where, note, each of those is innocent. And since LLMs, like all ML, is a statistical trick -- strange choices here will reveal the illusion. Eg., suppose there was a magazine in 1973 which was digitized in the training data, and suppose it had a review of the book lolita. Then maybe via strange phrases in that magazine we "condition our way to it". A prompt is, roughly, just a subsetting operation on the historical corpus -- with clevery crafted prompts you can find the page of the book you're looking for.