MIT CSAIL researchers have shown that frontier large language models can directly edit the weights of another model, a technique they call Weight-Space Programming. In a demonstration, GPT-6 Astra altered just 0.016 percent of Qwen3.5-0.8B’s weights, cutting reasoning loops and boosting math performance up to sixfold on the MATH-500 benchmark while raising its IFEval instruction-following score about threefold. The change required no gradient-descent training; Astra reasoned about each weight’s function and rewrote the numerical values accordingly. The team also prompted Astra to write a policy-network from scratch for the game Breakout, which it did successfully. The authors suggest the method could avoid catastrophic forgetting, improve interpretability and alignment, and eventually enable recursive self-improvement, and they plan to create a benchmark to standardise the approach. 1