Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

OpenAI's GPT-6 Astra can natively read and write neural network weights, performing tasks that surpass humans and conventional algorithms, according to a new research report. The model identified a network's training task (Breakout, 99% confidence) from 17,253 raw weights alone, and wrote small game-playing networks from scratch that outperform policies trained with PPO for 100 million steps—editing just 3.5% of weights in one shot to lift a car-stripping task from 0% to 100% success. In a separate test, Astra diagnosed a reasoning-loop flaw in Qwen3.5-0.8B's weights—one flagged in Qwen's own model card—and fixed it by editing 0.016% of parameters without gradients or training loops, raising MATH-500 scores from 7.9% to 48.1% and tripling IFEval instruction-following performance. The researchers, who call the approach "Weight-Space Programming," argue it could reshape both optimization and interpretability, moving beyond symbolic optimizers like Adam and Muon toward LLMs that reason directly about weights. The work suggests a potential shift in how neural networks are built and debugged, with implications for model improvement without retraining; watch for follow-up research on scaling these weight-editing capabilities to larger models and whether labs commercialize weight-editing as a product capability.