Evals and Responsible Scaling Policies

This is a hybrid event so you can either come to 100 University Avenue and call Giles (647-823-4865) to be let up to the fifth floor OR if you're too far away, message/text him for the Google Meet link.

This week, Thomas Broadley will be remoting from Los Angeles to tell us about AI Evals and Responsible Scaling Policies.

"Better dangerous capability evaluations for cutting-edge AI models could substantially reduce AI x-risk." He'll talk about his reasons for thinking this, ARC Evals' published work on evaluations, Anthropic's Responsible Scaling Policy and how evaluations factor into it, and future directions that he wants to see explored.