x
Anthropomorphic CoT traces: Can we distill a "character-driven" model, or is this data poison? — LessWrong