Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
retained text
You read it right: Astra was asked to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. Next step: Models that can edit themselves.
document fetched 2026-11-03 00:30 · canonical https://reddit.com/r/singularity/comments/1xm5oy9/sparks_of_weightspace_programming_llms_can