Is AI interpretability close to a breakthrough? A new essay has researchers divided

A recent study suggests that approximately 90 percent of neurons in neural networks perform identifiable functions, potentially simplifying the process of interpreting AI decision-making. This finding has sparked debate among researchers regarding whether the inner workings of large models can be truly transparent. If validated, the ability to map these internal circuits could provide a significant tool for AI safety by allowing developers to directly identify and mitigate deceptive or harmful behaviors within models.