MIT Research Reveals GPT-6 Astra's Ability to Reprogram AI Models

MIT researchers have demonstrated that OpenAI's GPT-6 Astra can read and deliberately rewrite the weights of other AI models without traditional training. In one experiment, Astra improved a Qwen model's math accuracy from 7.9 to 48.1 percent by altering just 0.016 percent of its weights. This approach could revolutionize AI optimization by allowing models to adjust their weights directly, significantly reducing the need for complex training processes. However, this method still encounters clear limitations, indicating that there are constraints on its effectiveness.