MIT research: GPT-6 Astra improved a model by rewriting 0.016% of its weights
MIT researchers published research showing that OpenAI's GPT-6 Astra can improve some AI models by directly rewriting their weights rather than retraining them, according to The Rundown AI. In the reported test, Astra changed just 0.016% of an Alibaba Qwen model's weights, reducing its repetition habit and lifting its math score from 7.9% to 48.1%. The same approach had little effect on Google's smaller Gemma 3, whose math score barely moved, and smaller tests showed Astra naming a game learned from weights alone and writing a simple Breakout player from scratch.
The reported result suggests a path toward automated model improvement and self-modifying AI research agents, but the gain was uneven across models and rests on a single secondary report of the MIT work.
This source is a digest of several news items (“MIT shows AI improving AI, no training required”). Find this story in the original.