“must expand the evaluation portfolio to include the possibility of extreme risks from general-purpose AI models that have strong skills in manipulation, deception, cyber-offense”
deepmind.com/blog/an-early-… pic.twitter.com/dcGlDzLnxz