AI opens another AI's weights and fixes it without training: 10 facts from MIT's demo

ccerrato147 · x · 2026-11-04

MIT CSAIL's demo had GPT-6 'Astra' open a small Qwen model's raw weights, locate and subtract a looping habit in two layers, and improve its math with no gradients or training. Ten facts from the post, model card and four independent reruns: the patch touches only 131,072 of 800M weights (0.016%); Hugging Face and Zeng Jiajun replicated within a day (4B: 59 to 92 correct on 150 problems vs 47-57 for random patches); Astra hand-wrote a Breakout network scoring 517 vs 25 for RL after 500M steps, and identified a sycophantic model from attention patterns alone 3/3 times. But the hype overreaches: '20x better than RL' holds in 1 of 3 Atari games, and '6x better at math' comes from thinking-mode looping — default mode went just 44.6% to 48.1%. Only the newest models (GPT 6, Opus 5.5) can do this. The attached image is a 'hand of god editing weights' illustration.

Original post →

AI opens another AI's weights and fixes it without training: 10 facts from MIT's demo

More from Models