The patch is tiny: 131,072 of 800M weights, readable line by line and undoable in a second
ccerrato147 · x · 2026-11-04
What makes the fix more than a party trick is its size: 131,072 weights out of 800 million, across two layers (about 0.016%); on the 4B version just two columns in one layer. A patch that small can be read line by line, shipped as a file, and undone in a second.
Related event: MIT Shows LLMs Can Directly Edit Model Weights to Boost Performance(114 posts)→
More from Research
- NRGPT paper reimagines GPT inference as gradient descent on an explicit energy landscape — burkov · 2026-11-04
- Fireworks event slides show sliding-window bit-sharing quantization trick — sophiamyang · 2026-11-04
- UCLA Study Seeks Users of 'Griefbots' to Interview on AI and Mourning ($30 Gift Card) — Interesting-Cat-7838 · 2026-11-04
- Tiny LoRA adapter beats DNA model Carbon-X on RareBench, exposing generalization limits — danielmckinn0n · 2026-11-04
- QuickProof: formal verification for C with specs written as plain C tests — thomasahle · 2026-11-04
- RSNA 2026 Deep Learning Lab to teach medical imaging VLM dataset building from scratch — PRouzrokh · 2026-11-04