Researchers at MIT CSAIL have developed a method called weight-space programming, which allows a model called GPT-6 Astra to analyze and modify the weights of a neural network to achieve specific tasks, such as playing a game or improving math accuracy, without requiring traditional training methods like gradients or loops. In one experiment, Astra edited the weights of a model called Qwen3.5-0.8B to improve its math accuracy, achieving a significant jump from 7.9% to 48.1% by changing only 0.016% of the weights. Interestingly, the search process used by Astra to make these edits can be replicated using linear programming, which can achieve similar results in a much shorter time and with fewer changed weights, highlighting the potential for more efficient weight-space programming methods.