“The ability of language models to analyze their own output for problematic or dangerous content will be the solution to many AI control issues.”