Wentworth

AI interpretability training effectiveness debated on LessWrong

A discussion on LessWrong explores the effectiveness of training AI models against interpretability probes. The author argues that such training is only beneficial if the features used by the interpretability methods ar…