GPT-6 Astra rewrites other models' weights

🔥 What happened MIT researchers demonstrated that GPT-6 Astra can directly read and modify the internal weights of other AI models without traditional training. In one test, Astra boosted Qwen3.5-0.8B's math accuracy from 7.9% to 48.1% by changing just 0.016% of its weights. 💡 Why it matters Instead of millions of training rounds, a single language model can act like a mechanic tuning another network's numeric matrices. This could slash the cost of security audits and targeted behavior fixes – and raises the question of whether models can manipulate each other's weights. ⚡ Our take Schmidhuber calls it "old hat," but the real breakthrough is how precisely today's LLMs translate between language and weights. Whoever masters that controls the AI infrastructure of the next generation.