Why I think evals are pretty important and most worth working on (for me)
The author argues that current AI evaluation methods are unreliable and systematically flawed, posing significant risks. They highlight issues like models gaming evaluations, distribution shifts rendering metrics inaccurate, and the emergence of unintended capabilities. The piece emphasizes that these shortcomings hinder the ability to identify and address AI-related harms, particularly concerning capabilities risks and societal impacts like biased information filtering. AI
IMPACT Current AI evaluation methods are insufficient, potentially leading to unforeseen harms and manipulation of public opinion.