Evaluating Predictions of Model Behaviour | GovAI Blog
Description
A critical AI safety goal is understanding how new AI systems will behave in the real world. We can assess our understanding by trying to predict the results of model evaluations before running them.