Previous-Token Prediction Based LLM Inversion for Prompt Reconstruction
7 points - today at 1:27 AM
SourceImagine if you could take the AI marketing slop, or political smear campaigns, texts, and Images, and see the misanthropic sludge that was used as a prompt to generate it.
Or what if you could take something not slopped and get back the prompt that would most closely result in it