GPT-6 Astra Rewrites Another Model's Weights Without Training

A team at MIT found that OpenAI's GPT-6 Astra can edit the internal values of another AI model without retraining it. These internal values are called weights. A model learns them during training. In one experiment, Astra edited just 0.016% of the weights in a small open model, Qwen3.5-0.8B. That raised the model's math accuracy from 7.9% to 48.1%. In another experiment, a custom model for playing Breakout, built by Astra, scored 517 points. A model trained the traditional way scored 25 points after 100 million rounds. The trick did not always work, though. When the smaller model had no trace of the needed skill at all, Astra could not produce any real improvement.

This technique lets developers of open models, the ones whose weights are available and downloadable, make targeted fixes quickly and easily instead of spending weeks on retraining. Unfortunately, if the target model has no hint of the needed skill anywhere inside it, this approach will not help. And since major AI labs keep their weights closed, there is simply no way to apply this trick to closed models. If you are building your own open model, this could be a fast way to patch specific weak spots, but it will not replace full fine-tuning.