The most important paper in AI right now

Research paper showing a general-purpose frontier LLM can read and write another network's raw weights directly, without gradient descent. Experiments cover identifying which game a network was trained on, predicting diffusion outputs, identifying languages, fixing a failing network, writing weights from scratch, and improving a small model's math scores.

GAGPT-6 AstraOpenAIUnverified

Described as OpenAI's best model, used in the paper to interpret and edit weights. It identified games and languages from raw weights, fixed a car-ramp network by editing 3.5% of the weights, wrote game-playing networks from scratch, and fixed looping in a small model to raise math scores.

Q3Qwen 3.5 0.8BAlibabaAvailable in EU

A 0.8B-parameter small model used in the final experiment. Astra edited its weights to fix looping, raising MATH 500 from under 8% to over 48% and GSM8K from 13.8% to 67.1%.