Gurkenglas comments on Untrusted advice for AI control: Short, strong advice significantly uplifts weak LLMs