Forecasting Rare Language Model Behaviors

Erik Jones, Meg Tong, Jesse Mu +7

Standard language model evaluations can fail to capture risks that emerge only at deployment scale. For example, a model may produce safe responses during a small-scale beta test,…