Morning Edition · Thursday, November 12, 2026Published at 1:18 AM EST · New York
Hugging Face model cards for "abliterated" Gemma 4 builds report refusal rates on harmful prompts falling from 99 in 100 to 26 in 100, achieved by editing the weights with no retraining.

A guide to running a modified Google Gemma 4 model on a personal computer, with no account, no subscription and refusals removed, has been saved by 14,000 people, the AI Post channel reported. The guide's author runs the model locally for a…
Track frontier labs, chips, export controls, model releases, regulation, and AI infrastructure.
The Global Intelligence Brief stays free.
Part of a tracked trend
Open-Weight Safeguards Are Removable
Cheap weight edits keep stripping safety tuning from open-weight models, which pushes governance away from refusal behavior and toward capability thresholds and access controls applied before release.
Start a discussion in Townsquare.
More from this edition
Comments
0No comments yet.