Zach Stein-Perlman comments on Claude 3.5 Sonnet

[Zach Stein-Perlman]

I thought that paper was just dangerous-capability evals, not safety-related metrics like adversarial robustness.