2 ms·
The original claim was that they can produce those robustly, though. Yes, the chances will be non-zero, but that doesn't mean it will be common or high fidelity
by lolinder 3y ago
The original claim was that they can produce those robustly, though. Yes, the chances will be non-zero, but that doesn't mean it will be common or high fidelity.
- mjburgess 3y agoAh, 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.