Every MIT CSAIL story collected by The AI Daily, 2 so far, newest first, refreshed hourly.
User has Claude hand-edit Gemma neurons to tune Italian translationFollowing MIT CSAIL's weight-space programming approach, a user had Claude Opus 5.5 rewrite 8 MLP neurons in Gemma 4 12B without training—about 92k of 11.9B parameters—to choose the Italian imperfect tense only in appropriate contexts.
Study probes Qwen's 'Wait' reasoning loopsA community post says MIT CSAIL fixed Qwen3.5-0.8B's reasoning loops by editing 0.016% of its weights using GPT-6 Astra, raising MATH-500 from 7.9% to 48.1%, and released the diff. The poster applied Anthropic's Jacobian lens and found the model reads 'Wrong' signals when repeating 'Wait', replaced by 'Verification' after the patch.
The A.I. Daily