defun 🔸 comments on Twitter thread on AI safety evals