Elliot Mckernon comments on Improving the safety of AI evals