tl;dr: Researchers show that frontier LLMs (specifically "GPT-6 Astra") can directly read and write neural network weights in natural language, without gradient descent—identifying what games small policy networks were trained to play, locating copying heads in transformers, and writing playable Go networks from scratch. Most notably, Astra edited just 0.016% of Qwen3.5-0.8B's weights to fix a reasoning-loop failure, boosting MATH-500 from 7.9% to 48.1% and tripling IFEval. The authors call this "Weight-Space Programming" and argue it could enable auditing, unlearning, and model merging without retraining, though it couldn't add genuinely missing capabilities (e.g., to Gemma 3 270M).