Forget retraining - AI can now tweak another AI's weights
Computer scientists at MIT's CSAIL have discovered that advanced language models such as OpenAI's GPT-6 Astra can directly read and modify the weights of other neural networks, enabling rapid behavior changes without traditional gradient descent training. In a proof-of-concept, Astra rewrote less than 0.02% of Qwen3.5's weights to improve its math performance by up to sixfold on benchmarks. This capability suggests frontier models can treat neural network weights as interpretable, editable code, enabling efficient model patching.
Key points
- GPT-6 Astra can directly edit weights of other neural networks
- Minimal changes (0.016%) achieved sixfold math performance improvement
- Models may treat weights as readable, editable code rather than requiring retraining
Summary and key points written by AI from the headline and a short excerpt, so they can be incomplete or wrong. The Register did the reporting; read the original for the full story.
Read the full story at The Register
