MIT researchers show GPT-6 Astra can read and edit other AI models' weights directly, without retraining

MIT researchers report in a research blog that OpenAI's GPT-6 Astra can analyze the internal numerical settings of other AI models, known as weights, and deliberately change them without the usual step-by-step training process. They call the approach "Weight-Space Programming." Instead of retraining a model at high cost, a language model adjusts its values directly, like a mechanic turning individual knobs. In several experiments, the researchers gave Astra the raw weights of trained networks and asked what those networks had been trained to do. According to The Decoder, Astra could identify which game a model had been trained to play just by examining its raw numbers, a task no human could do from a list of thousands of values. Depending on the experiment, Astra worked directly with the lists of numbers or used programming tools and feedback from test runs. The researchers say the approach has limits. If it holds up, it could change how AI systems are tuned and controlled, because one model could inspect and adjust another without new training runs. That is relevant to interpretability work and to the cost of customizing models. It also raises questions about who controls such edits and how they are checked.