Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss.

A LessWrong post discusses the Persona Selection Model, which holds that post-training selects a persona the model learned during pre-training, and cites Subliminal Learning by Cloud et al., which showed models can impart covert signals such as a "love of owls" into semantically unrelated output.

Written by AI from LessWrong's published text.