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

[Thomas Kwa]

This is very exciting. It means a large part of monitoring reduces to monitoring short advice strings.