MIT researchers showed that GPT-6 Astra can interpret raw internal values in other AI models, predict the functions of different components and edit the systems to improve them, including boosting mathematical abilities. The findings appeared in a preprint on 2 November.