Researchers test GPT-6 Astra as a tool for reprogramming AI model weights directly
AIResearchers tested a method called Weight-Space Programming, in which the GPT-6 Astra language model edits or writes the raw weights of other AI models instead of retraining them. On Qwen3.5-0.8B, changing 0.016 percent of the weights raised MATH-500 accuracy from 7.9 to 48.1 percent, though thinking loops were not fully eliminated.
AIWhy it matters.
The source compares weight-space programming with conventional retraining, showing where targeted weight edits improved a model and where they could not add missing skills.
Source details
The Decoder · Maximilian Schreiner · the-decoder.comPublished
Original title: GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights