GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
decoder ·
MIT researchers have demonstrated that OpenAI's GPT-6 Astra can analyze and modify the internal weights of other AI models through a method called "Weight-Space Programming." This approach allows the language model to act like a mechanic, adjusting numerical settings without the need for expensive retraining. In experiments, Astra correctly identified games and languages by examining raw weight matrices, though it sometimes confused shape types. The system also generated functional game-playing models from scratch that outperformed traditionally trained baselines in certain scenarios. When tasked with improving the math skills of the Qwen3.5-0.8B model, Astra achieved significant accuracy gains by altering only a tiny fraction of its weights. However, the technique showed limitations when applied to smaller models lacking specific capabilities. While some experts view this as an evolution of older concepts, the study highlights new potential for targeted AI tuning and security audits.
decoder