LLMs · via Hacker News
MIT researchers show GPT-6 Astra can read and edit neural network weights directly
An MIT CSAIL team found GPT-6 Astra can tell what a network does just from its raw weights and fix it by hand-editing them. Editing 0.016% of Qwen3.5-0.8B's weights to stop reasoning loops lifted MATH-500 from 7.9% to 48.1%, with no training. It hit 2,390 points on Hacker News, and a rebuttal argues the results are less superhuman than they sound.