Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
Researchers report that GPT-6 Astra inspected and edited neural-network weights, including a targeted edit that improved Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% under their test settings.
The Qwen experiment used a Python sandbox and up to 50 batches of forward passes, not just weights in a prompt. The edit primarily reduced reasoning loops; a separate attempt to improve a non-looping Gemma model did not succeed. Other experiments used small game-playing networks.
HN readers found the results striking but emphasized the distinction between fixing an existing failure mode and teaching a model new math.