AI Now Writes Neural Network Brains Directly

🔥 What happened Frontier models like GPT-6 Astra now edit neural network weights directly via text – no backprop. Astra pushed Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by rewriting 131,072 weights. Multiple independent replications confirm the effect, including a laptop run by Simon Willison. 💡 Why it matters Most of the gain is repair: 576 of 579 fixed answers were previously stuck in a 32K-token loop. On held-out problems, accuracy jumped from 10.6% to 60.4% – existing reasoning that can finally finish. Critics call it bug-fixing, not recursive self-improvement. ⚡ Our take Fitting weights directly against a benchmark is overfitting with extra steps. The real story: coding agents suddenly wield mechanistic-interpretability tools effectively – that's the actual paradigm shift.