Inference-Time Intervention: Eliciting Truthful Answers from a Language Model