Distilling persona vectors into LLM weights / hacker news