Thoughts On the Nature of Capability Elicitation via Fine-tuning — LessWrong