GPT-6 Astra edited another model's weights directly, with no training
Rikiya Takehi's post drew about 5,600 likes.
Researchers asked GPT-6 Astra to edit the weights of the small Qwen3.5-0.8B model to improve its math, according to the post.
It changed only 0.016% of the weights, improving math scores by up to 6x and instruction following by 3x, without gradients or training.
The researchers call the approach weight-space programming.
