Researchers at MIT have demonstrated that an AI model called GPT-6 Astra can interpret the internal workings of other neural networks and edit them to improve performance. The team showed that GPT-6 Astra can read raw internal values from other models well enough to predict what different parts do, then modify those models to boost their mathematical abilities. The researchers reported their findings in a preprint published on November 2. The work offers a new method for understanding the "black box" nature of AI systems, moving beyond the typical approach of improving models through incremental training or building them from scratch.