GPT-6 Reads Neural Nets Like Plain Text

🔥 What happened MIT researchers show GPT-6 Astra reads raw network weights and identifies what a model was trained to do. Without gradients, it boosts Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing just 0.016% of weights. 💡 Why it matters Until now, Adam and friends optimized blindly. Now an LLM reasons about weights and rewrites them directly – no training loop, no environment feedback. This could upend both optimization and interpretability. ⚡ Our take This isn't a benchmark stunt – it's a paradigm shift: the optimizer becomes a reasoner. Anyone still betting solely on gradients is watching from the sidelines.