'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance

MIT researchers utilized GPT-6 Astra to directly modify the internal weights of the open-source Qwen model, resulting in measurable improvements to its mathematical performance. While the team describes this technique as a form of neural network gene editing, external observers have noted that the approach relies on established machine learning engineering methods. This development highlights a growing trend in using advanced AI models to refine the capabilities of other systems without traditional retraining.