MIT CSAIL fixes small models by hand-editing weights — hype or old idea rebranded?
ziv_ravid · x · 2026-11-04
Commenting on Omar Khattab's MIT CSAIL "weight-space programming" experiments: an agent (Astra) inspected and manually updated a small model's weights — no training, no gradients — fixing math, instruction-following, even writing its own game-playing network and identifying which game a network was trained on from weights alone. The author is excited ("gene editing for neural networks"), annoyed (the trick leans on interpretability tools that only work for small models, first/last-layer reads, and tasks already in training data — possibly good PR for an old idea), and jealous (his own attempts in this direction worked far worse). Bottom line: it's complicated.
Related event: MIT's Manual Weight Editing Draws Skepticism from Researchers(2 posts)→
More from Research
- Terence Tao predicts an explosion of experimental mathematics as AI maps math difficulty — haider1 · 2026-11-04
- Google DeepMind Genomics team hiring 2027 Student Researcher with RL and post-training skills — alexechu_ · 2026-11-04
- SAIL AI Challenge opens final phase: turn rare disease diagnoses into treatment reports — zakkohane · 2026-11-04
- qwenFS and qwenGPU: a filesystem and SIMT GPU compiled into Qwen3.5-2B's weights — CatAstro_Piyush · 2026-11-04
- D&A turns any VLM into a zero-shot robot reward model with SOTA on three benchmarks — KyleMorgenstein · 2026-11-04
- Berkeley's Beguš uses AI interpretability to find vowel analogs in whale communication — begusgasper · 2026-11-04