'One Weird Trick' for Alignment: Add Random Noise During Training
Sauers_ · x · 2026-11-15
The author jokingly presents a method to improve alignment generalization: inject random noise during training while instructing the model to keep its values stable under perturbation — effectively baking value consistency into the training objective.
Related event: Injecting Random Noise During Training Improves Alignment Generalization(7 posts)→
More from Fun
- Character design goes from 2D illustration to full 3D in one step — azed_ai · 2026-11-15
- Musk Amplifies AI-Remixed Starship Landing Video 'Made on Earth by Humans' — elonmusk · 2026-11-15
- Google Drive claims 108.27 GB used of a 15 GB quota, demands 93 GB cleanup — moonsandhues · 2026-11-15
- Swipe-comic 'Follow the little boat' shows vast mind pausing for a child-sized wonder — tallmetommy · 2026-11-15
- Musk marks 5-year leap: Starship returns from orbit, caught by Mechazilla chopsticks — elonmusk · 2026-11-15
- Schmidhuber retweets meme: flat minima date back to 1997 — SchmidhuberAI · 2026-11-15