Direct intervention in weights
MIT researchers found that OpenAI's GPT-6 Astra can read the weights of other AI models and deliberately rewrite them. That means intervening without going through traditional training processes.
Weights are the numerical values that represent what an AI model has learned; they are normally updated through training data and lengthy computation. Astra's ability to change these values directly could make model tuning far faster.
Striking result in a Qwen experiment
According to the researchers, in one experiment Astra changed just 0.016% of a Qwen model's weights, raising the model's math accuracy from 7.9% to 48.1%. A very small intervention produced a clear performance gain.
Limits remain
While promising, the research also points to clear limits. It is not yet clear whether direct weight intervention will produce the same results across every model and task, or what it means for safety and oversight.
