richard_ngo comments on Twitter thread on AI safety evals